Duplicate from kevinwang676/Bark-UI-with-Voice-Cloning

.gitattributes

@@ -0,0 +1,40 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+SE_checkpoint.pth.tar filter=lfs diff=lfs merge=lfs -text
+best_model.pth.tar filter=lfs diff=lfs merge=lfs -text
+nana_longest_vocal.wav filter=lfs diff=lfs merge=lfs -text
+test.wav filter=lfs diff=lfs merge=lfs -text
+reference.wav filter=lfs diff=lfs merge=lfs -text
+ref.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,14 @@
+---
+title: Bark with Voice Cloning
+emoji: 📊
+colorFrom: purple
+colorTo: purple
+sdk: gradio
+sdk_version: 3.27.0
+app_file: app.py
+pinned: false
+license: mit
+duplicated_from: kevinwang676/Bark-UI-with-Voice-Cloning
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

SE_checkpoint.pth.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f96efb20cbeeefd81fd8336d7f0155bf8902f82f9474e58ccb19d9e12345172
+size 44610930

app.py

@@ -0,0 +1,555 @@
+import os
+import sys
+
+os.system("git clone https://github.com/C0untFloyd/bark-gui.git")
+sys.path.append("./bark-gui/")
+
+from cProfile import label
+from distutils.command.check import check
+from doctest import Example
+import gradio as gr
+import numpy as np
+import logging
+import torch
+import pytorch_seed
+import time
+
+from xml.sax import saxutils
+from bark.api import generate_with_settings
+from bark.api import save_as_prompt
+from settings import Settings
+#import nltk
+
+from bark import SAMPLE_RATE
+from bark.clonevoice import clone_voice
+from bark.generation import SAMPLE_RATE, preload_models
+from scipy.io.wavfile import write as write_wav
+from parseinput import split_and_recombine_text, build_ssml, is_ssml, create_clips_from_ssml
+from datetime import datetime
+from tqdm.auto import tqdm
+from id3tagging import add_id3_tag
+
+import shutil
+
+import string
+import argparse
+import json
+
+from TTS.tts.utils.synthesis import synthesis
+from TTS.tts.utils.text.symbols import make_symbols, phonemes, symbols
+try:
+  from TTS.utils.audio import AudioProcessor
+except:
+  from TTS.utils.audio import AudioProcessor
+
+
+from TTS.tts.models import setup_model
+from TTS.config import load_config
+from TTS.tts.models.vits import *
+
+from TTS.tts.utils.speakers import SpeakerManager
+from pydub import AudioSegment
+
+# from google.colab import files
+import librosa
+
+from scipy.io.wavfile import write, read
+
+import subprocess
+
+
+OUTPUTFOLDER = "Outputs"
+
+
+def generate_text_to_speech(text, selected_speaker, text_temp, waveform_temp, eos_prob, quick_generation, complete_settings, seed, progress=gr.Progress(track_tqdm=True)):
+    if text == None or len(text) < 1:
+        raise gr.Error('No text entered!')
+
+    # Chunk the text into smaller pieces then combine the generated audio
+
+    # generation settings
+    if selected_speaker == 'None':
+        selected_speaker = None
+    if seed != None and seed > 2**32 - 1:
+        logger.warning(f"Seed {seed} > 2**32 - 1 (max), setting to random")
+        seed = None
+    if seed == None or seed <= 0:
+        seed = np.random.default_rng().integers(1, 2**32 - 1)
+    assert(0 < seed and seed < 2**32)
+
+    voice_name = selected_speaker
+    use_last_generation_as_history = "Use last generation as history" in complete_settings
+    save_last_generation = "Save generation as Voice" in complete_settings
+    progress(0, desc="Generating")
+
+    silenceshort = np.zeros(int((float(settings.silence_sentence) / 1000.0) * SAMPLE_RATE), dtype=np.float32)  # quarter second of silence
+    silencelong = np.zeros(int((float(settings.silence_speakers) / 1000.0) * SAMPLE_RATE), dtype=np.float32)  # half a second of silence
+    full_generation = None
+
+    all_parts = []
+    complete_text = ""
+    text = text.lstrip()
+    if is_ssml(text):
+        list_speak = create_clips_from_ssml(text)
+        prev_speaker = None
+        for i, clip in tqdm(enumerate(list_speak), total=len(list_speak)):
+            selected_speaker = clip[0]
+            # Add pause break between speakers
+            if i > 0 and selected_speaker != prev_speaker:
+                all_parts += [silencelong.copy()]
+            prev_speaker = selected_speaker
+            text = clip[1]
+            text = saxutils.unescape(text)
+            if selected_speaker == "None":
+                selected_speaker = None
+
+            print(f"\nGenerating Text ({i+1}/{len(list_speak)}) -> {selected_speaker} (Seed {seed}):`{text}`")
+            complete_text += text
+            with pytorch_seed.SavedRNG(seed):
+                audio_array = generate_with_settings(text_prompt=text, voice_name=selected_speaker, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob)
+                seed = torch.random.initial_seed()
+            if len(list_speak) > 1:
+                filename = create_filename(OUTPUTFOLDER, seed, "audioclip",".wav")
+                save_wav(audio_array, filename)
+                add_id3_tag(filename, text, selected_speaker, seed)
+
+            all_parts += [audio_array]
+    else:
+        texts = split_and_recombine_text(text, settings.input_text_desired_length, settings.input_text_max_length)
+        for i, text in tqdm(enumerate(texts), total=len(texts)):
+            print(f"\nGenerating Text ({i+1}/{len(texts)}) -> {selected_speaker} (Seed {seed}):`{text}`")
+            complete_text += text
+            if quick_generation == True:
+                with pytorch_seed.SavedRNG(seed):
+                    audio_array = generate_with_settings(text_prompt=text, voice_name=selected_speaker, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob)
+                    seed = torch.random.initial_seed()
+            else:
+                full_output = use_last_generation_as_history or save_last_generation
+                if full_output:
+                    full_generation, audio_array = generate_with_settings(text_prompt=text, voice_name=voice_name, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob, output_full=True)
+                else:
+                    audio_array = generate_with_settings(text_prompt=text, voice_name=voice_name, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob)
+
+            # Noticed this in the HF Demo - convert to 16bit int -32767/32767 - most used audio format  
+            # audio_array = (audio_array * 32767).astype(np.int16)
+
+            if len(texts) > 1:
+                filename = create_filename(OUTPUTFOLDER, seed, "audioclip",".wav")
+                save_wav(audio_array, filename)
+                add_id3_tag(filename, text, selected_speaker, seed)
+
+            if quick_generation == False and (save_last_generation == True or use_last_generation_as_history == True):
+                # save to npz
+                voice_name = create_filename(OUTPUTFOLDER, seed, "audioclip", ".npz")
+                save_as_prompt(voice_name, full_generation)
+                if use_last_generation_as_history:
+                    selected_speaker = voice_name
+
+            all_parts += [audio_array]
+            # Add short pause between sentences
+            if text[-1] in "!?.\n" and i > 1:
+                all_parts += [silenceshort.copy()]
+
+    # save & play audio
+    result = create_filename(OUTPUTFOLDER, seed, "final",".wav")
+    save_wav(np.concatenate(all_parts), result)
+    # write id3 tag with text truncated to 60 chars, as a precaution...
+    add_id3_tag(result, complete_text, selected_speaker, seed)
+    return result
+
+def create_filename(path, seed, name, extension):
+    now = datetime.now()
+    date_str =now.strftime("%m-%d-%Y")
+    outputs_folder = os.path.join(os.getcwd(), path)
+    if not os.path.exists(outputs_folder):
+        os.makedirs(outputs_folder)
+
+    sub_folder = os.path.join(outputs_folder, date_str)
+    if not os.path.exists(sub_folder):
+        os.makedirs(sub_folder)
+
+    time_str = now.strftime("%H-%M-%S")
+    file_name = f"{name}_{time_str}_s{seed}{extension}"
+    return os.path.join(sub_folder, file_name)
+
+
+def save_wav(audio_array, filename):
+    write_wav(filename, SAMPLE_RATE, audio_array)
+
+def save_voice(filename, semantic_prompt, coarse_prompt, fine_prompt):
+    np.savez_compressed(
+        filename,
+        semantic_prompt=semantic_prompt,
+        coarse_prompt=coarse_prompt,
+        fine_prompt=fine_prompt
+    )
+    
+
+def on_quick_gen_changed(checkbox):
+    if checkbox == False:
+        return gr.CheckboxGroup.update(visible=True)
+    return gr.CheckboxGroup.update(visible=False)
+
+def delete_output_files(checkbox_state):
+    if checkbox_state:
+        outputs_folder = os.path.join(os.getcwd(), OUTPUTFOLDER)
+        if os.path.exists(outputs_folder):
+            purgedir(outputs_folder)
+    return False
+
+
+# https://stackoverflow.com/a/54494779
+def purgedir(parent):
+    for root, dirs, files in os.walk(parent):                                      
+        for item in files:
+            # Delete subordinate files                                                 
+            filespec = os.path.join(root, item)
+            os.unlink(filespec)
+        for item in dirs:
+            # Recursively perform this operation for subordinate directories   
+            purgedir(os.path.join(root, item))
+
+def convert_text_to_ssml(text, selected_speaker):
+    return build_ssml(text, selected_speaker)
+
+
+def apply_settings(themes, input_server_name, input_server_port, input_server_public, input_desired_len, input_max_len, input_silence_break, input_silence_speaker):
+    settings.selected_theme = themes
+    settings.server_name = input_server_name
+    settings.server_port = input_server_port
+    settings.server_share = input_server_public
+    settings.input_text_desired_length = input_desired_len
+    settings.input_text_max_length = input_max_len
+    settings.silence_sentence = input_silence_break
+    settings.silence_speaker = input_silence_speaker
+    settings.save()
+
+def restart():
+    global restart_server
+    restart_server = True
+
+
+def create_version_html():
+    python_version = ".".join([str(x) for x in sys.version_info[0:3]])
+    versions_html = f"""
+python: <span title="{sys.version}">{python_version}</span>
+ • 
+torch: {getattr(torch, '__long_version__',torch.__version__)}
+ • 
+gradio: {gr.__version__}
+"""
+    return versions_html
+
+    
+
+logger = logging.getLogger(__name__)
+APPTITLE = "Bark UI Enhanced v0.4.6"
+
+
+autolaunch = False
+
+if len(sys.argv) > 1:
+    autolaunch = "-autolaunch" in sys.argv
+
+
+if torch.cuda.is_available() == False:
+    os.environ['BARK_FORCE_CPU'] = 'True'
+    logger.warning("No CUDA detected, fallback to CPU!")
+
+print(f'smallmodels={os.environ.get("SUNO_USE_SMALL_MODELS", False)}')
+print(f'enablemps={os.environ.get("SUNO_ENABLE_MPS", False)}')
+print(f'offloadcpu={os.environ.get("SUNO_OFFLOAD_CPU", False)}')
+print(f'forcecpu={os.environ.get("BARK_FORCE_CPU", False)}')
+print(f'autolaunch={autolaunch}\n\n')
+
+#print("Updating nltk\n")
+#nltk.download('punkt')
+
+print("Preloading Models\n")
+preload_models()
+
+settings = Settings('config.yaml')
+
+# Collect all existing speakers/voices in dir
+speakers_list = []
+
+for root, dirs, files in os.walk("./bark/assets/prompts"):
+	for file in files:
+		if(file.endswith(".npz")):
+			pathpart = root.replace("./bark/assets/prompts", "")
+			name = os.path.join(pathpart, file[:-4])
+			if name.startswith("/") or name.startswith("\\"):
+				name = name[1:]
+			speakers_list.append(name)
+
+speakers_list = sorted(speakers_list, key=lambda x: x.lower())
+speakers_list.insert(0, 'None')
+
+available_themes = ["Default", "gradio/glass", "gradio/monochrome", "gradio/seafoam", "gradio/soft", "gstaff/xkcd", "freddyaboulton/dracula_revamped", "ysharma/steampunk"]
+
+seed = -1
+server_name = settings.server_name
+if len(server_name) < 1:
+    server_name = None
+server_port = settings.server_port
+if server_port <= 0:
+    server_port = None
+global run_server
+global restart_server
+
+run_server = True
+
+
+
+
+'''
+from google.colab import drive
+drive.mount('/content/drive')
+src_path = os.path.join(os.path.join(os.path.join(os.path.join(os.getcwd(), 'drive'), 'MyDrive'), 'Colab Notebooks'), 'best_model_latest.pth.tar')
+dst_path = os.path.join(os.getcwd(), 'best_model.pth.tar')
+shutil.copy(src_path, dst_path)
+'''
+
+TTS_PATH = "TTS/"
+
+# add libraries into environment
+sys.path.append(TTS_PATH) # set this if TTS is not installed globally
+
+# Paths definition
+
+OUT_PATH = 'out/'
+
+# create output path
+os.makedirs(OUT_PATH, exist_ok=True)
+
+# model vars 
+MODEL_PATH = 'best_model.pth.tar'
+CONFIG_PATH = 'config.json'
+TTS_LANGUAGES = "language_ids.json"
+TTS_SPEAKERS = "speakers.json"
+USE_CUDA = torch.cuda.is_available()
+
+# load the config
+C = load_config(CONFIG_PATH)
+
+# load the audio processor
+ap = AudioProcessor(**C.audio)
+
+speaker_embedding = None
+
+C.model_args['d_vector_file'] = TTS_SPEAKERS
+C.model_args['use_speaker_encoder_as_loss'] = False
+
+model = setup_model(C)
+model.language_manager.set_language_ids_from_file(TTS_LANGUAGES)
+# print(model.language_manager.num_languages, model.embedded_language_dim)
+# print(model.emb_l)
+cp = torch.load(MODEL_PATH, map_location=torch.device('cpu'))
+# remove speaker encoder
+model_weights = cp['model'].copy()
+for key in list(model_weights.keys()):
+  if "speaker_encoder" in key:
+    del model_weights[key]
+
+model.load_state_dict(model_weights)
+
+model.eval()
+
+if USE_CUDA:
+    model = model.cuda()
+
+# synthesize voice
+use_griffin_lim = False
+
+# Paths definition
+
+CONFIG_SE_PATH = "config_se.json"
+CHECKPOINT_SE_PATH = "SE_checkpoint.pth.tar"
+
+# Load the Speaker encoder
+
+SE_speaker_manager = SpeakerManager(encoder_model_path=CHECKPOINT_SE_PATH, encoder_config_path=CONFIG_SE_PATH, use_cuda=USE_CUDA)
+
+# Define helper function
+
+def compute_spec(ref_file):
+  y, sr = librosa.load(ref_file, sr=ap.sample_rate)
+  spec = ap.spectrogram(y)
+  spec = torch.FloatTensor(spec).unsqueeze(0)
+  return spec
+
+
+def voice_conversion(ta, ra, da):
+
+  target_audio = 'target.wav'
+  reference_audio = 'reference.wav'
+  driving_audio = 'driving.wav'
+
+  write(target_audio, ta[0], ta[1])
+  write(reference_audio, ra[0], ra[1])
+  write(driving_audio, da[0], da[1])          
+
+  # !ffmpeg-normalize $target_audio -nt rms -t=-27 -o $target_audio -ar 16000 -f
+  # !ffmpeg-normalize $reference_audio -nt rms -t=-27 -o $reference_audio -ar 16000 -f
+  # !ffmpeg-normalize $driving_audio -nt rms -t=-27 -o $driving_audio -ar 16000 -f
+
+  files = [target_audio, reference_audio, driving_audio]
+
+  for file in files:
+      subprocess.run(["ffmpeg-normalize", file, "-nt", "rms", "-t=-27", "-o", file, "-ar", "16000", "-f"])
+
+  # ta_ = read(target_audio)
+
+  target_emb = SE_speaker_manager.compute_d_vector_from_clip([target_audio])
+  target_emb = torch.FloatTensor(target_emb).unsqueeze(0)
+
+  driving_emb = SE_speaker_manager.compute_d_vector_from_clip([reference_audio])
+  driving_emb = torch.FloatTensor(driving_emb).unsqueeze(0)
+
+  # Convert the voice
+
+  driving_spec = compute_spec(driving_audio)
+  y_lengths = torch.tensor([driving_spec.size(-1)])
+  if USE_CUDA:
+      ref_wav_voc, _, _ = model.voice_conversion(driving_spec.cuda(), y_lengths.cuda(), driving_emb.cuda(), target_emb.cuda())
+      ref_wav_voc = ref_wav_voc.squeeze().cpu().detach().numpy()
+  else:
+      ref_wav_voc, _, _ = model.voice_conversion(driving_spec, y_lengths, driving_emb, target_emb)
+      ref_wav_voc = ref_wav_voc.squeeze().detach().numpy()
+
+  # print("Reference Audio after decoder:")
+  # IPython.display.display(Audio(ref_wav_voc, rate=ap.sample_rate))
+
+  return (ap.sample_rate, ref_wav_voc)
+
+
+while run_server:
+    print(f'Launching {APPTITLE} Server')
+
+    # Create Gradio Blocks
+
+    with gr.Blocks(title=f"{APPTITLE}", mode=f"{APPTITLE}", theme=settings.selected_theme) as barkgui:
+        with gr.Row():
+            with gr.Column():
+                gr.Markdown(f"### [{APPTITLE}](https://github.com/C0untFloyd/bark-gui)")
+            with gr.Column():
+                gr.HTML(create_version_html(), elem_id="versions")
+
+        with gr.Tab("TTS"):
+            with gr.Row():
+                with gr.Column():
+                    placeholder = "Enter text here."
+                    input_text = gr.Textbox(label="Input Text", lines=4, placeholder=placeholder)
+                with gr.Column():
+                    seedcomponent = gr.Number(label="Seed (default -1 = Random)", precision=0, value=-1)
+                    convert_to_ssml_button = gr.Button("Convert Text to SSML")
+            with gr.Row():
+                with gr.Column():
+                    examples = [
+                        "Special meanings: [laughter] [laughs] [sighs] [music] [gasps] [clears throat] MAN: WOMAN:",
+                       "♪ Never gonna make you cry, never gonna say goodbye, never gonna tell a lie and hurt you ♪",
+                       "And now — a picture of a larch [laughter]",
+                       """
+                            WOMAN: I would like an oatmilk latte please.
+                            MAN: Wow, that's expensive!
+                       """,
+                       """<?xml version="1.0"?>
+    <speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis"
+             xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+             xsi:schemaLocation="http://www.w3.org/2001/10/synthesis
+                       http://www.w3.org/TR/speech-synthesis/synthesis.xsd"
+             xml:lang="en-US">
+    <voice name="en_speaker_9">Look at that drunk guy!</voice>
+    <voice name="en_speaker_3">Who is he?</voice>
+    <voice name="en_speaker_9">WOMAN: [clears throat] 10 years ago, he proposed me and I rejected him.</voice>
+    <voice name="en_speaker_3">Oh my God [laughs] he is still celebrating</voice>
+    </speak>"""
+                       ]
+                    examples = gr.Examples(examples=examples, inputs=input_text)
+
+            with gr.Row():
+                with gr.Column():
+                    gr.Markdown("[Voice Prompt Library](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c)")
+                    speaker = gr.Dropdown(speakers_list, value=speakers_list[0], label="Voice")
+                with gr.Column():
+                    text_temp = gr.Slider(0.1, 1.0, value=0.6, label="Generation Temperature", info="1.0 more diverse, 0.1 more conservative")
+                    waveform_temp = gr.Slider(0.1, 1.0, value=0.7, label="Waveform temperature", info="1.0 more diverse, 0.1 more conservative")
+
+            with gr.Row():
+                with gr.Column():
+                    quick_gen_checkbox = gr.Checkbox(label="Quick Generation", value=True)
+                    settings_checkboxes = ["Use last generation as history", "Save generation as Voice"]
+                    complete_settings = gr.CheckboxGroup(choices=settings_checkboxes, value=settings_checkboxes, label="Detailed Generation Settings", type="value", interactive=True, visible=False)
+                with gr.Column():
+                    eos_prob = gr.Slider(0.0, 0.5, value=0.05, label="End of sentence probability")
+
+            with gr.Row():
+                with gr.Column():
+                    tts_create_button = gr.Button("Generate")
+                with gr.Column():
+                    hidden_checkbox = gr.Checkbox(visible=False)
+                    button_stop_generation = gr.Button("Stop generation")
+            with gr.Row():
+                output_audio = gr.Audio(label="Generated Audio")
+
+            with gr.Row():
+                inp1 = gr.Audio(label='Target Speaker - Reference Clip')
+                inp2 = output_audio
+                inp3 = output_audio
+                btn = gr.Button("Generate")
+                out1 = gr.Audio(label='Target Speaker - Converted Clip')
+            btn.click(voice_conversion, [inp1, inp2, inp3], [out1])
+                
+
+
+        with gr.Tab("Clone Voice"):
+            input_audio_filename = gr.Audio(label="Input audio.wav", source="upload", type="filepath")
+            transcription_text = gr.Textbox(label="Transcription Text", lines=1, placeholder="Enter Text of your Audio Sample here...")
+            initialname = "./bark/assets/prompts/custom/MeMyselfAndI"
+            output_voice = gr.Textbox(label="Filename of trained Voice", lines=1, placeholder=initialname, value=initialname)
+            clone_voice_button = gr.Button("Create Voice")
+            dummy = gr.Text(label="Progress")
+
+        with gr.Tab("Settings"):
+            with gr.Row():
+                themes = gr.Dropdown(available_themes, label="Theme", info="Change needs complete restart", value=settings.selected_theme)
+            with gr.Row():
+                input_server_name = gr.Textbox(label="Server Name", lines=1, info="Leave blank to run locally", value=settings.server_name)
+                input_server_port = gr.Number(label="Server Port", precision=0, info="Leave at 0 to use default", value=settings.server_port)
+                share_checkbox = gr.Checkbox(label="Public Server", value=settings.server_share)
+            with gr.Row():
+                input_desired_len = gr.Slider(100, 150, value=settings.input_text_desired_length, label="Desired Input Text Length", info="Ideal length to split input sentences")
+                input_max_len = gr.Slider(150, 256, value=settings.input_text_max_length, label="Max Input Text Length", info="Maximum Input Text Length")
+            with gr.Row():
+                input_silence_break = gr.Slider(1, 1000, value=settings.silence_sentence, label="Sentence Pause Time (ms)", info="Silence between sentences in milliseconds")
+                input_silence_speakers = gr.Slider(1, 5000, value=settings.silence_speakers, label="Speaker Pause Time (ms)", info="Silence between different speakers in milliseconds")
+
+            with gr.Row():
+                button_apply_settings = gr.Button("Apply Settings")
+                button_apply_restart = gr.Button("Restart Server")
+                button_delete_files = gr.Button("Clear output folder")
+
+        quick_gen_checkbox.change(fn=on_quick_gen_changed, inputs=quick_gen_checkbox, outputs=complete_settings)
+        convert_to_ssml_button.click(convert_text_to_ssml, inputs=[input_text, speaker],outputs=input_text)
+        gen_click = tts_create_button.click(generate_text_to_speech, inputs=[input_text, speaker, text_temp, waveform_temp, eos_prob, quick_gen_checkbox, complete_settings, seedcomponent],outputs=output_audio)
+        button_stop_generation.click(fn=None, inputs=None, outputs=None, cancels=[gen_click])
+        # Javascript hack to display modal confirmation dialog
+        js = "(x) => confirm('Are you sure? This will remove all files from output folder')"
+        button_delete_files.click(None, None, hidden_checkbox, _js=js)
+        hidden_checkbox.change(delete_output_files, [hidden_checkbox], [hidden_checkbox])
+        clone_voice_button.click(clone_voice, inputs=[input_audio_filename, transcription_text, output_voice], outputs=dummy)
+        button_apply_settings.click(apply_settings, inputs=[themes, input_server_name, input_server_port, share_checkbox, input_desired_len, input_max_len, input_silence_break, input_silence_speakers])
+        button_apply_restart.click(restart)
+        restart_server = False
+        try:
+            barkgui.queue().launch(show_error=True)
+        except:
+            restart_server = True
+            run_server = False
+        try:
+            while restart_server == False:
+                time.sleep(1.0)
+        except (KeyboardInterrupt, OSError):
+            print("Keyboard interruption in main thread... closing server.")
+            run_server = False
+        barkgui.close()

bark/__init__.py

@@ -0,0 +1,2 @@
+from .api import generate_audio, text_to_semantic, semantic_to_waveform, save_as_prompt
+from .generation import SAMPLE_RATE, preload_models

bark/api.py

@@ -0,0 +1,158 @@
+from typing import Dict, Optional, Union
+
+import numpy as np
+
+from .generation import codec_decode, generate_coarse, generate_fine, generate_text_semantic
+
+
+def generate_with_settings(text_prompt, semantic_temp=0.6, eos_p=0.2, coarse_temp=0.7, fine_temp=0.5, voice_name=None, output_full=False):
+
+    # generation with more control
+    x_semantic = generate_text_semantic(
+        text_prompt,
+        history_prompt=voice_name,
+        temp=semantic_temp,
+        min_eos_p = eos_p,
+        use_kv_caching=True
+    )
+
+    x_coarse_gen = generate_coarse(
+        x_semantic,
+        history_prompt=voice_name,
+        temp=coarse_temp,
+        use_kv_caching=True
+    )
+    x_fine_gen = generate_fine(
+        x_coarse_gen,
+        history_prompt=voice_name,
+        temp=fine_temp,
+    )
+
+    if output_full:
+        full_generation = {
+            'semantic_prompt': x_semantic,
+            'coarse_prompt': x_coarse_gen,
+            'fine_prompt': x_fine_gen
+        }
+        return full_generation, codec_decode(x_fine_gen)
+    return codec_decode(x_fine_gen)
+
+
+def text_to_semantic(
+    text: str,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    temp: float = 0.7,
+    silent: bool = False,
+):
+    """Generate semantic array from text.
+
+    Args:
+        text: text to be turned into audio
+        history_prompt: history choice for audio cloning
+        temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+
+    Returns:
+        numpy semantic array to be fed into `semantic_to_waveform`
+    """
+    x_semantic = generate_text_semantic(
+        text,
+        history_prompt=history_prompt,
+        temp=temp,
+        silent=silent,
+        use_kv_caching=True
+    )
+    return x_semantic
+
+
+def semantic_to_waveform(
+    semantic_tokens: np.ndarray,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    temp: float = 0.7,
+    silent: bool = False,
+    output_full: bool = False,
+):
+    """Generate audio array from semantic input.
+
+    Args:
+        semantic_tokens: semantic token output from `text_to_semantic`
+        history_prompt: history choice for audio cloning
+        temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+        output_full: return full generation to be used as a history prompt
+
+    Returns:
+        numpy audio array at sample frequency 24khz
+    """
+    coarse_tokens = generate_coarse(
+        semantic_tokens,
+        history_prompt=history_prompt,
+        temp=temp,
+        silent=silent,
+        use_kv_caching=True
+    )
+    fine_tokens = generate_fine(
+        coarse_tokens,
+        history_prompt=history_prompt,
+        temp=0.5,
+    )
+    audio_arr = codec_decode(fine_tokens)
+    if output_full:
+        full_generation = {
+            "semantic_prompt": semantic_tokens,
+            "coarse_prompt": coarse_tokens,
+            "fine_prompt": fine_tokens,
+        }
+        return full_generation, audio_arr
+    return audio_arr
+
+
+def save_as_prompt(filepath, full_generation):
+    assert(filepath.endswith(".npz"))
+    assert(isinstance(full_generation, dict))
+    assert("semantic_prompt" in full_generation)
+    assert("coarse_prompt" in full_generation)
+    assert("fine_prompt" in full_generation)
+    np.savez(filepath, **full_generation)
+
+
+def generate_audio(
+    text: str,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    text_temp: float = 0.7,
+    waveform_temp: float = 0.7,
+    silent: bool = False,
+    output_full: bool = False,
+):
+    """Generate audio array from input text.
+
+    Args:
+        text: text to be turned into audio
+        history_prompt: history choice for audio cloning
+        text_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        waveform_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+        output_full: return full generation to be used as a history prompt
+
+    Returns:
+        numpy audio array at sample frequency 24khz
+    """
+    semantic_tokens = text_to_semantic(
+        text,
+        history_prompt=history_prompt,
+        temp=text_temp,
+        silent=silent,
+    )
+    out = semantic_to_waveform(
+        semantic_tokens,
+        history_prompt=history_prompt,
+        temp=waveform_temp,
+        silent=silent,
+        output_full=output_full,
+    )
+    if output_full:
+        full_generation, audio_arr = out
+        return full_generation, audio_arr
+    else:
+        audio_arr = out
+    return audio_arr

bark/assets/prompts/announcer.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26f2d1a9e3b6fe453cf5fc8191de26cbfae6276c5b0f7c376c6a0f3c35867f83
+size 16794

bark/assets/prompts/en_speaker_0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:932f40d879ba8659f1ca26319ba64ea3b0647b2050fe24313bf42b0dff1fe241
+size 28100

bark/assets/prompts/en_speaker_1.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e7f18015e1ab9b6302ded1e28a971af5306a72f193bb6c411f1948a083c8578
+size 25220

bark/assets/prompts/en_speaker_2.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d218990680ece5f2d4fc18ea4783b016b3ae353ec413eaee2058f2d57263c9b3
+size 26236

bark/assets/prompts/en_speaker_3.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92c2e2a29145c83738e9b63f082fd1c873d9422468a155463cb27f814aeaea66
+size 34980

bark/assets/prompts/en_speaker_4.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:992f91991a9a5359d72f00b09a11a550e71bb8ebfc0cfd877e39d7d41f98b714
+size 23780

bark/assets/prompts/en_speaker_5.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18831c3f6014e4a2ff60ad5169b1fae06e28ed07f43f8a3616aafb84515091bf
+size 24740

bark/assets/prompts/en_speaker_6.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fab38dc6b6bc9226bcc414f4c5a9524bc1b2441865a586153fb620127a8faa4e
+size 25540

bark/assets/prompts/en_speaker_7.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f4c4eb33f5994be8de5cfd1744ebce13da1618a6da3a7d244514178c61ef7db
+size 22716

bark/assets/prompts/en_speaker_8.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fc9f11b539588f51bbf78150a73e0365c49b2306bd72e5a22b28ef09c4fb15d
+size 23300

bark/assets/prompts/en_speaker_9.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78b3ba32eb9aeb9ed34556856c40633ecc8332d1c3ae3c81e6f5015ac3eefbd5
+size 30180

bark/assets/prompts/zh_speaker_0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd7ac118a3e944b3f20c89f2446056a00850a630ee16318922acc6572ce80929
+size 20636

bark/assets/prompts/zh_speaker_1.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0eacf5c862dfd3c5ac825f2ebb26f323e64309cb712e7e264cbd31c5bca3f038
+size 19836

bark/assets/prompts/zh_speaker_2.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e324b47f8250e5798c314f395d4e049575e7ca369d0b6074e91c7bba70e9f26d
+size 21060

bark/assets/prompts/zh_speaker_3.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98c476abc7bf634ffb2d71d363284e7bd8c8abd5e33ec5ca21d4aa5b15730d18
+size 31300

bark/assets/prompts/zh_speaker_4.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fa8673a9895ad3302d13ac94193b5ad5da481f1cc276e6181fa895acaae133b
+size 29964

bark/assets/prompts/zh_speaker_5.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:226edfe5fabc72eeb83a13e350599bc8babe5adc2264b3cdb661fd1258dc4044
+size 17436

bark/assets/prompts/zh_speaker_6.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:285d51fbe81cc263636b5b487fbb6633e6f3cf92c53ca9ab8e6b7f55d4b4a31d
+size 16900

bark/assets/prompts/zh_speaker_7.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0967cdb14ffa79895747b0d52df9f15bdad80d6c55b7630894345c9a7ec87c91
+size 21060

bark/assets/prompts/zh_speaker_8.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c028f78530013f29ab8c0c1cf4fe2138106fbe5252951f5f36e0168056779549
+size 19300

bark/assets/prompts/zh_speaker_9.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6265bb827008d7af8a45a8e057fe3e91efb347d56208180a9ed990ad54e4d75e
+size 16156

bark/clonevoice.py

@@ -0,0 +1,41 @@
+from bark.generation import load_codec_model, generate_text_semantic, grab_best_device
+from encodec.utils import convert_audio
+import torchaudio
+import torch
+import os
+import gradio
+
+
+def clone_voice(audio_filepath, text, dest_filename, progress=gradio.Progress(track_tqdm=True)):
+    if len(text) < 1:
+        raise gradio.Error('No transcription text entered!')
+
+    use_gpu = not os.environ.get("BARK_FORCE_CPU", False)
+    progress(0, desc="Loading Codec")
+    model = load_codec_model(use_gpu=use_gpu)
+    progress(0.25, desc="Converting WAV")
+
+    # Load and pre-process the audio waveform
+    device = grab_best_device(use_gpu)
+    wav, sr = torchaudio.load(audio_filepath)
+    wav = convert_audio(wav, sr, model.sample_rate, model.channels)
+    wav = wav.unsqueeze(0).to(device)
+    progress(0.5, desc="Extracting codes")
+
+    # Extract discrete codes from EnCodec
+    with torch.no_grad():
+        encoded_frames = model.encode(wav)
+    codes = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1).squeeze()  # [n_q, T]
+
+    # get seconds of audio
+    seconds = wav.shape[-1] / model.sample_rate
+    # generate semantic tokens
+    semantic_tokens = generate_text_semantic(text, max_gen_duration_s=seconds, top_k=50, top_p=.95, temp=0.7)
+
+    # move codes to cpu
+    codes = codes.cpu().numpy()
+
+    import numpy as np
+    output_path = dest_filename + '.npz'
+    np.savez(output_path, fine_prompt=codes, coarse_prompt=codes[:2, :], semantic_prompt=semantic_tokens)
+    return "Finished"

bark/generation.py

@@ -0,0 +1,865 @@
+import contextlib
+import gc
+import os
+import re
+import requests
+import gc
+import sys
+
+from encodec import EncodecModel
+import funcy
+import logging
+import numpy as np
+from scipy.special import softmax
+import torch
+import torch.nn.functional as F
+import tqdm
+from transformers import BertTokenizer
+from huggingface_hub import hf_hub_download, hf_hub_url
+
+from .model import GPTConfig, GPT
+from .model_fine import FineGPT, FineGPTConfig
+from .settings import initenv
+
+initenv(sys.argv)
+global_force_cpu = os.environ.get("BARK_FORCE_CPU", False)
+if (
+    global_force_cpu != True and
+    torch.cuda.is_available() and
+    hasattr(torch.cuda, "amp") and
+    hasattr(torch.cuda.amp, "autocast") and
+    hasattr(torch.cuda, "is_bf16_supported") and
+    torch.cuda.is_bf16_supported()
+):
+    autocast = funcy.partial(torch.cuda.amp.autocast, dtype=torch.bfloat16)
+else:
+    @contextlib.contextmanager
+    def autocast():
+        yield
+
+
+# hold models in global scope to lazy load
+global models
+models = {}
+
+global models_devices
+models_devices = {}
+
+
+CONTEXT_WINDOW_SIZE = 1024
+
+SEMANTIC_RATE_HZ = 49.9
+SEMANTIC_VOCAB_SIZE = 10_000
+
+CODEBOOK_SIZE = 1024
+N_COARSE_CODEBOOKS = 2
+N_FINE_CODEBOOKS = 8
+COARSE_RATE_HZ = 75
+
+SAMPLE_RATE = 24_000
+
+
+SUPPORTED_LANGS = [
+    ("English", "en"),
+    ("German", "de"),
+    ("Spanish", "es"),
+    ("French", "fr"),
+    ("Hindi", "hi"),
+    ("Italian", "it"),
+    ("Japanese", "ja"),
+    ("Korean", "ko"),
+    ("Polish", "pl"),
+    ("Portuguese", "pt"),
+    ("Russian", "ru"),
+    ("Turkish", "tr"),
+    ("Chinese", "zh"),
+]
+
+ALLOWED_PROMPTS = {"announcer"}
+for _, lang in SUPPORTED_LANGS:
+    for prefix in ("", f"v2{os.path.sep}"):
+        for n in range(10):
+            ALLOWED_PROMPTS.add(f"{prefix}{lang}_speaker_{n}")
+
+
+logger = logging.getLogger(__name__)
+
+
+CUR_PATH = os.path.dirname(os.path.abspath(__file__))
+
+
+#default_cache_dir = os.path.join(os.path.expanduser("~"), ".cache")
+#CACHE_DIR = os.path.join(os.getenv("XDG_CACHE_HOME", default_cache_dir), "suno", "bark_v0")
+#CACHE_DIR = os.path.join(os.getcwd(), "models"
+CACHE_DIR = "./models"
+
+
+def _cast_bool_env_var(s):
+    return s.lower() in ('true', '1', 't')
+
+USE_SMALL_MODELS = _cast_bool_env_var(os.environ.get("SUNO_USE_SMALL_MODELS", "False"))
+GLOBAL_ENABLE_MPS = _cast_bool_env_var(os.environ.get("SUNO_ENABLE_MPS", "False"))
+OFFLOAD_CPU = _cast_bool_env_var(os.environ.get("SUNO_OFFLOAD_CPU", "False"))
+
+REMOTE_MODEL_PATHS = {
+    "text_small": {
+        "repo_id": "suno/bark",
+        "file_name": "text.pt",
+    },
+    "coarse_small": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse.pt",
+    },
+    "fine_small": {
+        "repo_id": "suno/bark",
+        "file_name": "fine.pt",
+    },
+    "text": {
+        "repo_id": "suno/bark",
+        "file_name": "text_2.pt",
+    },
+    "coarse": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse_2.pt",
+    },
+    "fine": {
+        "repo_id": "suno/bark",
+        "file_name": "fine_2.pt",
+    },
+}
+
+
+if not hasattr(torch.nn.functional, 'scaled_dot_product_attention') and torch.cuda.is_available():
+    logger.warning(
+        "torch version does not support flash attention. You will get faster" +
+        " inference speed by upgrade torch to newest nightly version."
+    )
+
+
+def grab_best_device(use_gpu=True):
+    if torch.cuda.device_count() > 0 and use_gpu:
+        device = "cuda"
+    elif torch.backends.mps.is_available() and use_gpu and GLOBAL_ENABLE_MPS:
+        device = "mps"
+    else:
+        device = "cpu"
+    return device
+
+
+def _get_ckpt_path(model_type, use_small=False):
+    key = model_type
+    if use_small or USE_SMALL_MODELS:
+        key += "_small"
+    return os.path.join(CACHE_DIR, REMOTE_MODEL_PATHS[key]["file_name"])
+
+"""
+def _download(from_hf_path, file_name, destfilename):
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    hf_hub_download(repo_id=from_hf_path, filename=file_name, local_dir=CACHE_DIR, local_dir_use_symlinks=False)
+    # Bug in original repo? Downloaded name differs from expected...
+    if not os.path.exists(destfilename):
+        localname = os.path.join(CACHE_DIR, file_name)
+        os.rename(localname, destfilename)
+"""
+def _download(from_hf_path, file_name):
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    hf_hub_download(repo_id=from_hf_path, filename=file_name, local_dir=CACHE_DIR)
+
+
+class InferenceContext:
+    def __init__(self, benchmark=False):
+        # we can't expect inputs to be the same length, so disable benchmarking by default
+        self._chosen_cudnn_benchmark = benchmark
+        self._cudnn_benchmark = None
+
+    def __enter__(self):
+        self._cudnn_benchmark = torch.backends.cudnn.benchmark
+        torch.backends.cudnn.benchmark = self._chosen_cudnn_benchmark
+
+    def __exit__(self, exc_type, exc_value, exc_traceback):
+        torch.backends.cudnn.benchmark = self._cudnn_benchmark
+
+
+if torch.cuda.is_available():
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+
+
+@contextlib.contextmanager
+def _inference_mode():
+    with InferenceContext(), torch.inference_mode(), torch.no_grad(), autocast():
+        yield
+
+
+def _clear_cuda_cache():
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+
+def clean_models(model_key=None):
+    global models
+    model_keys = [model_key] if model_key is not None else models.keys()
+    for k in model_keys:
+        if k in models:
+            del models[k]
+    _clear_cuda_cache()
+    gc.collect()
+
+
+def _load_model(ckpt_path, device, use_small=False, model_type="text"):
+    if model_type == "text":
+        ConfigClass = GPTConfig
+        ModelClass = GPT
+    elif model_type == "coarse":
+        ConfigClass = GPTConfig
+        ModelClass = GPT
+    elif model_type == "fine":
+        ConfigClass = FineGPTConfig
+        ModelClass = FineGPT
+    else:
+        raise NotImplementedError()
+
+    # Force-remove Models to allow running on >12Gb GPU
+    # CF: Probably not needed anymore
+    #global models
+    #models.clear()
+    #gc.collect()
+    #torch.cuda.empty_cache()
+    # to here...
+
+    model_key = f"{model_type}_small" if use_small or USE_SMALL_MODELS else model_type
+    model_info = REMOTE_MODEL_PATHS[model_key]
+    if not os.path.exists(ckpt_path):
+        logger.info(f"{model_type} model not found, downloading into `{CACHE_DIR}`.")
+        ## added next two lines to make it super clear which model is being downloaded
+        remote_filename = hf_hub_url(model_info["repo_id"], model_info["file_name"])
+        print(f"Downloading {model_key} {model_info['repo_id']} remote model file {remote_filename} {model_info['file_name']} to {CACHE_DIR}")
+        _download(model_info["repo_id"], model_info["file_name"])
+        # add next line to make it super clear which model is being loaded
+    print(f"Loading {model_key} model from {ckpt_path} to {device}") # added
+    checkpoint = torch.load(ckpt_path, map_location=device)
+    # this is a hack
+    model_args = checkpoint["model_args"]
+    if "input_vocab_size" not in model_args:
+        model_args["input_vocab_size"] = model_args["vocab_size"]
+        model_args["output_vocab_size"] = model_args["vocab_size"]
+        del model_args["vocab_size"]
+    gptconf = ConfigClass(**checkpoint["model_args"])
+    model = ModelClass(gptconf)
+    state_dict = checkpoint["model"]
+    # fixup checkpoint
+    unwanted_prefix = "_orig_mod."
+    for k, v in list(state_dict.items()):
+        if k.startswith(unwanted_prefix):
+            state_dict[k[len(unwanted_prefix) :]] = state_dict.pop(k)
+    extra_keys = set(state_dict.keys()) - set(model.state_dict().keys())
+    extra_keys = set([k for k in extra_keys if not k.endswith(".attn.bias")])
+    missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
+    missing_keys = set([k for k in missing_keys if not k.endswith(".attn.bias")])
+    if len(extra_keys) != 0:
+        raise ValueError(f"extra keys found: {extra_keys}")
+    if len(missing_keys) != 0:
+        raise ValueError(f"missing keys: {missing_keys}")
+    model.load_state_dict(state_dict, strict=False)
+    n_params = model.get_num_params()
+    val_loss = checkpoint["best_val_loss"].item()
+    logger.info(f"model loaded: {round(n_params/1e6,1)}M params, {round(val_loss,3)} loss")
+    model.eval()
+    model.to(device)
+    del checkpoint, state_dict
+    _clear_cuda_cache()
+    if model_type == "text":
+        tokenizer = BertTokenizer.from_pretrained("bert-base-multilingual-cased")
+        return {
+            "model": model,
+            "tokenizer": tokenizer,
+        }
+    return model
+
+
+def _load_codec_model(device):
+    model = EncodecModel.encodec_model_24khz()
+    model.set_target_bandwidth(6.0)
+    model.eval()
+    model.to(device)
+    _clear_cuda_cache()
+    return model
+
+
+def load_model(use_gpu=True, use_small=False, force_reload=False, model_type="text"):
+    _load_model_f = funcy.partial(_load_model, model_type=model_type, use_small=use_small)
+    if model_type not in ("text", "coarse", "fine"):
+        raise NotImplementedError()
+    global models
+    global models_devices
+    device = grab_best_device(use_gpu=use_gpu)
+    model_key = f"{model_type}"
+    if OFFLOAD_CPU:
+        models_devices[model_key] = device
+        device = "cpu"
+    if model_key not in models or force_reload:
+        ckpt_path = _get_ckpt_path(model_type, use_small=use_small)
+        clean_models(model_key=model_key)
+        model = _load_model_f(ckpt_path, device)
+        models[model_key] = model
+    if model_type == "text":
+        models[model_key]["model"].to(device)
+    else:
+        models[model_key].to(device)
+    return models[model_key]
+
+
+def load_codec_model(use_gpu=True, force_reload=False):
+    global models
+    global models_devices
+    device = grab_best_device(use_gpu=use_gpu)
+    if device == "mps":
+        # encodec doesn't support mps
+        device = "cpu"
+    model_key = "codec"
+    if OFFLOAD_CPU:
+        models_devices[model_key] = device
+        device = "cpu"
+    if model_key not in models or force_reload:
+        clean_models(model_key=model_key)
+        model = _load_codec_model(device)
+        models[model_key] = model
+    models[model_key].to(device)
+    return models[model_key]
+
+
+def preload_models(
+    text_use_gpu=True,
+    text_use_small=False,
+    coarse_use_gpu=True,
+    coarse_use_small=False,
+    fine_use_gpu=True,
+    fine_use_small=False,
+    codec_use_gpu=True,
+    force_reload=False
+):
+    """Load all the necessary models for the pipeline."""
+    if grab_best_device() == "cpu" and (
+        text_use_gpu or coarse_use_gpu or fine_use_gpu or codec_use_gpu
+    ):
+        logger.warning("No GPU being used. Careful, inference might be very slow!")
+    _ = load_model(
+        model_type="text", use_gpu=text_use_gpu, use_small=text_use_small, force_reload=force_reload
+    )
+    _ = load_model(
+        model_type="coarse",
+        use_gpu=coarse_use_gpu,
+        use_small=coarse_use_small,
+        force_reload=force_reload,
+    )
+    _ = load_model(
+        model_type="fine", use_gpu=fine_use_gpu, use_small=fine_use_small, force_reload=force_reload
+    )
+    _ = load_codec_model(use_gpu=codec_use_gpu, force_reload=force_reload)
+
+
+####
+# Generation Functionality
+####
+
+
+def _tokenize(tokenizer, text):
+    return tokenizer.encode(text, add_special_tokens=False)
+
+
+def _detokenize(tokenizer, enc_text):
+    return tokenizer.decode(enc_text)
+
+
+def _normalize_whitespace(text):
+    return re.sub(r"\s+", " ", text).strip()
+
+
+TEXT_ENCODING_OFFSET = 10_048
+SEMANTIC_PAD_TOKEN = 10_000
+TEXT_PAD_TOKEN = 129_595
+SEMANTIC_INFER_TOKEN = 129_599
+
+
+def _load_history_prompt(history_prompt_input):
+    if isinstance(history_prompt_input, str) and history_prompt_input.endswith(".npz"):
+        history_prompt = np.load(history_prompt_input)
+    elif isinstance(history_prompt_input, str):
+        # make sure this works on non-ubuntu
+        history_prompt_input = os.path.join(*history_prompt_input.split("/"))
+#        if history_prompt_input not in ALLOWED_PROMPTS:
+#            raise ValueError("history prompt not found")
+        history_prompt = np.load(
+            os.path.join(CUR_PATH, "assets", "prompts", f"{history_prompt_input}.npz")
+        )
+    elif isinstance(history_prompt_input, dict):
+        assert("semantic_prompt" in history_prompt_input)
+        assert("coarse_prompt" in history_prompt_input)
+        assert("fine_prompt" in history_prompt_input)
+        history_prompt = history_prompt_input
+    else:
+        raise ValueError("history prompt format unrecognized")
+    return history_prompt
+
+
+def generate_text_semantic(
+    text,
+    history_prompt=None,
+    temp=0.7,
+    top_k=None,
+    top_p=None,
+    silent=False,
+    min_eos_p=0.2,
+    max_gen_duration_s=None,
+    allow_early_stop=True,
+    use_kv_caching=False,
+):
+    """Generate semantic tokens from text."""
+    assert isinstance(text, str)
+    text = _normalize_whitespace(text)
+    assert len(text.strip()) > 0
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        semantic_history = history_prompt["semantic_prompt"]
+        assert (
+            isinstance(semantic_history, np.ndarray)
+            and len(semantic_history.shape) == 1
+            and len(semantic_history) > 0
+            and semantic_history.min() >= 0
+            and semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
+        )
+    else:
+        semantic_history = None
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "text" not in models:
+        preload_models()
+    model_container = models["text"]
+    model = model_container["model"]
+    tokenizer = model_container["tokenizer"]
+    encoded_text = np.array(_tokenize(tokenizer, text)) + TEXT_ENCODING_OFFSET
+    if OFFLOAD_CPU:
+        model.to(models_devices["text"])
+    device = next(model.parameters()).device
+    if len(encoded_text) > 256:
+        p = round((len(encoded_text) - 256) / len(encoded_text) * 100, 1)
+        logger.warning(f"warning, text too long, lopping of last {p}%")
+        encoded_text = encoded_text[:256]
+    encoded_text = np.pad(
+        encoded_text,
+        (0, 256 - len(encoded_text)),
+        constant_values=TEXT_PAD_TOKEN,
+        mode="constant",
+    )
+    if semantic_history is not None:
+        semantic_history = semantic_history.astype(np.int64)
+        # lop off if history is too long, pad if needed
+        semantic_history = semantic_history[-256:]
+        semantic_history = np.pad(
+            semantic_history,
+            (0, 256 - len(semantic_history)),
+            constant_values=SEMANTIC_PAD_TOKEN,
+            mode="constant",
+        )
+    else:
+        semantic_history = np.array([SEMANTIC_PAD_TOKEN] * 256)
+    x = torch.from_numpy(
+        np.hstack([
+            encoded_text, semantic_history, np.array([SEMANTIC_INFER_TOKEN])
+        ]).astype(np.int64)
+    )[None]
+    assert x.shape[1] == 256 + 256 + 1
+    with _inference_mode():
+        x = x.to(device)
+        n_tot_steps = 768
+        # custom tqdm updates since we don't know when eos will occur
+        pbar = tqdm.tqdm(disable=silent, total=100)
+        pbar_state = 0
+        tot_generated_duration_s = 0
+        kv_cache = None
+        for n in range(n_tot_steps):
+            if use_kv_caching and kv_cache is not None:
+                x_input = x[:, [-1]]
+            else:
+                x_input = x
+            logits, kv_cache = model(
+                x_input, merge_context=True, use_cache=use_kv_caching, past_kv=kv_cache
+            )
+            relevant_logits = logits[0, 0, :SEMANTIC_VOCAB_SIZE]
+            if allow_early_stop:
+                relevant_logits = torch.hstack(
+                    (relevant_logits, logits[0, 0, [SEMANTIC_PAD_TOKEN]])  # eos
+                )
+            if top_p is not None:
+                # faster to convert to numpy
+                logits_device = relevant_logits.device
+                logits_dtype = relevant_logits.type()
+                relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
+                sorted_indices = np.argsort(relevant_logits)[::-1]
+                sorted_logits = relevant_logits[sorted_indices]
+                cumulative_probs = np.cumsum(softmax(sorted_logits))
+                sorted_indices_to_remove = cumulative_probs > top_p
+                sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
+                sorted_indices_to_remove[0] = False
+                relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
+                relevant_logits = torch.from_numpy(relevant_logits)
+                relevant_logits = relevant_logits.to(logits_device).type(logits_dtype)
+            if top_k is not None:
+                v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
+                relevant_logits[relevant_logits < v[-1]] = -float("Inf")
+            probs = F.softmax(relevant_logits / temp, dim=-1)
+            # multinomial bugged on mps: shuttle to cpu if necessary
+            inf_device = probs.device
+            if probs.device.type == "mps":
+                probs = probs.to("cpu")
+            item_next = torch.multinomial(probs, num_samples=1)
+            probs = probs.to(inf_device)
+            item_next = item_next.to(inf_device)
+            if allow_early_stop and (
+                item_next == SEMANTIC_VOCAB_SIZE
+                or (min_eos_p is not None and probs[-1] >= min_eos_p)
+            ):
+                # eos found, so break
+                pbar.update(100 - pbar_state)
+                break
+            x = torch.cat((x, item_next[None]), dim=1)
+            tot_generated_duration_s += 1 / SEMANTIC_RATE_HZ
+            if max_gen_duration_s is not None and tot_generated_duration_s > max_gen_duration_s:
+                pbar.update(100 - pbar_state)
+                break
+            if n == n_tot_steps - 1:
+                pbar.update(100 - pbar_state)
+                break
+            del logits, relevant_logits, probs, item_next
+            req_pbar_state = np.min([100, int(round(100 * n / n_tot_steps))])
+            if req_pbar_state > pbar_state:
+                pbar.update(req_pbar_state - pbar_state)
+            pbar_state = req_pbar_state
+        pbar.close()
+        out = x.detach().cpu().numpy().squeeze()[256 + 256 + 1 :]
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    assert all(0 <= out) and all(out < SEMANTIC_VOCAB_SIZE)
+    _clear_cuda_cache()
+    return out
+
+
+def _flatten_codebooks(arr, offset_size=CODEBOOK_SIZE):
+    assert len(arr.shape) == 2
+    arr = arr.copy()
+    if offset_size is not None:
+        for n in range(1, arr.shape[0]):
+            arr[n, :] += offset_size * n
+    flat_arr = arr.ravel("F")
+    return flat_arr
+
+
+COARSE_SEMANTIC_PAD_TOKEN = 12_048
+COARSE_INFER_TOKEN = 12_050
+
+
+def generate_coarse(
+    x_semantic,
+    history_prompt=None,
+    temp=0.7,
+    top_k=None,
+    top_p=None,
+    silent=False,
+    max_coarse_history=630,  # min 60 (faster), max 630 (more context)
+    sliding_window_len=60,
+    use_kv_caching=False,
+):
+    """Generate coarse audio codes from semantic tokens."""
+    assert (
+        isinstance(x_semantic, np.ndarray)
+        and len(x_semantic.shape) == 1
+        and len(x_semantic) > 0
+        and x_semantic.min() >= 0
+        and x_semantic.max() <= SEMANTIC_VOCAB_SIZE - 1
+    )
+    assert 60 <= max_coarse_history <= 630
+    assert max_coarse_history + sliding_window_len <= 1024 - 256
+    semantic_to_coarse_ratio = COARSE_RATE_HZ / SEMANTIC_RATE_HZ * N_COARSE_CODEBOOKS
+    max_semantic_history = int(np.floor(max_coarse_history / semantic_to_coarse_ratio))
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        x_semantic_history = history_prompt["semantic_prompt"]
+        x_coarse_history = history_prompt["coarse_prompt"]
+        assert (
+            isinstance(x_semantic_history, np.ndarray)
+            and len(x_semantic_history.shape) == 1
+            and len(x_semantic_history) > 0
+            and x_semantic_history.min() >= 0
+            and x_semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
+            and isinstance(x_coarse_history, np.ndarray)
+            and len(x_coarse_history.shape) == 2
+            and x_coarse_history.shape[0] == N_COARSE_CODEBOOKS
+            and x_coarse_history.shape[-1] >= 0
+            and x_coarse_history.min() >= 0
+            and x_coarse_history.max() <= CODEBOOK_SIZE - 1
+            and (
+                round(x_coarse_history.shape[-1] / len(x_semantic_history), 1)
+                == round(semantic_to_coarse_ratio / N_COARSE_CODEBOOKS, 1)
+            )
+        )
+        x_coarse_history = _flatten_codebooks(x_coarse_history) + SEMANTIC_VOCAB_SIZE
+        # trim histories correctly
+        n_semantic_hist_provided = np.min(
+            [
+                max_semantic_history,
+                len(x_semantic_history) - len(x_semantic_history) % 2,
+                int(np.floor(len(x_coarse_history) / semantic_to_coarse_ratio)),
+            ]
+        )
+        n_coarse_hist_provided = int(round(n_semantic_hist_provided * semantic_to_coarse_ratio))
+        x_semantic_history = x_semantic_history[-n_semantic_hist_provided:].astype(np.int32)
+        x_coarse_history = x_coarse_history[-n_coarse_hist_provided:].astype(np.int32)
+        # TODO: bit of a hack for time alignment (sounds better)
+        x_coarse_history = x_coarse_history[:-2]
+    else:
+        x_semantic_history = np.array([], dtype=np.int32)
+        x_coarse_history = np.array([], dtype=np.int32)
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "coarse" not in models:
+        preload_models()
+    model = models["coarse"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["coarse"])
+    device = next(model.parameters()).device
+    # start loop
+    n_steps = int(
+        round(
+            np.floor(len(x_semantic) * semantic_to_coarse_ratio / N_COARSE_CODEBOOKS)
+            * N_COARSE_CODEBOOKS
+        )
+    )
+    assert n_steps > 0 and n_steps % N_COARSE_CODEBOOKS == 0
+    x_semantic = np.hstack([x_semantic_history, x_semantic]).astype(np.int32)
+    x_coarse = x_coarse_history.astype(np.int32)
+    base_semantic_idx = len(x_semantic_history)
+    with _inference_mode():
+        x_semantic_in = torch.from_numpy(x_semantic)[None].to(device)
+        x_coarse_in = torch.from_numpy(x_coarse)[None].to(device)
+        n_window_steps = int(np.ceil(n_steps / sliding_window_len))
+        n_step = 0
+        for _ in tqdm.tqdm(range(n_window_steps), total=n_window_steps, disable=silent):
+            semantic_idx = base_semantic_idx + int(round(n_step / semantic_to_coarse_ratio))
+            # pad from right side
+            x_in = x_semantic_in[:, np.max([0, semantic_idx - max_semantic_history]) :]
+            x_in = x_in[:, :256]
+            x_in = F.pad(
+                x_in,
+                (0, 256 - x_in.shape[-1]),
+                "constant",
+                COARSE_SEMANTIC_PAD_TOKEN,
+            )
+            x_in = torch.hstack(
+                [
+                    x_in,
+                    torch.tensor([COARSE_INFER_TOKEN])[None].to(device),
+                    x_coarse_in[:, -max_coarse_history:],
+                ]
+            )
+            kv_cache = None
+            for _ in range(sliding_window_len):
+                if n_step >= n_steps:
+                    continue
+                is_major_step = n_step % N_COARSE_CODEBOOKS == 0
+
+                if use_kv_caching and kv_cache is not None:
+                    x_input = x_in[:, [-1]]
+                else:
+                    x_input = x_in
+
+                logits, kv_cache = model(x_input, use_cache=use_kv_caching, past_kv=kv_cache)
+                logit_start_idx = (
+                    SEMANTIC_VOCAB_SIZE + (1 - int(is_major_step)) * CODEBOOK_SIZE
+                )
+                logit_end_idx = (
+                    SEMANTIC_VOCAB_SIZE + (2 - int(is_major_step)) * CODEBOOK_SIZE
+                )
+                relevant_logits = logits[0, 0, logit_start_idx:logit_end_idx]
+                if top_p is not None:
+                    # faster to convert to numpy
+                    logits_device = relevant_logits.device
+                    logits_dtype = relevant_logits.type()
+                    relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
+                    sorted_indices = np.argsort(relevant_logits)[::-1]
+                    sorted_logits = relevant_logits[sorted_indices]
+                    cumulative_probs = np.cumsum(softmax(sorted_logits))
+                    sorted_indices_to_remove = cumulative_probs > top_p
+                    sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
+                    sorted_indices_to_remove[0] = False
+                    relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
+                    relevant_logits = torch.from_numpy(relevant_logits)
+                    relevant_logits = relevant_logits.to(logits_device).type(logits_dtype)
+                if top_k is not None:
+                    v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
+                    relevant_logits[relevant_logits < v[-1]] = -float("Inf")
+                probs = F.softmax(relevant_logits / temp, dim=-1)
+                # multinomial bugged on mps: shuttle to cpu if necessary
+                inf_device = probs.device
+                if probs.device.type == "mps":
+                    probs = probs.to("cpu")
+                item_next = torch.multinomial(probs, num_samples=1)
+                probs = probs.to(inf_device)
+                item_next = item_next.to(inf_device)
+                item_next += logit_start_idx
+                x_coarse_in = torch.cat((x_coarse_in, item_next[None]), dim=1)
+                x_in = torch.cat((x_in, item_next[None]), dim=1)
+                del logits, relevant_logits, probs, item_next
+                n_step += 1
+            del x_in
+        del x_semantic_in
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    gen_coarse_arr = x_coarse_in.detach().cpu().numpy().squeeze()[len(x_coarse_history) :]
+    del x_coarse_in
+    assert len(gen_coarse_arr) == n_steps
+    gen_coarse_audio_arr = gen_coarse_arr.reshape(-1, N_COARSE_CODEBOOKS).T - SEMANTIC_VOCAB_SIZE
+    for n in range(1, N_COARSE_CODEBOOKS):
+        gen_coarse_audio_arr[n, :] -= n * CODEBOOK_SIZE
+    _clear_cuda_cache()
+    return gen_coarse_audio_arr
+
+
+def generate_fine(
+    x_coarse_gen,
+    history_prompt=None,
+    temp=0.5,
+    silent=True,
+):
+    """Generate full audio codes from coarse audio codes."""
+    assert (
+        isinstance(x_coarse_gen, np.ndarray)
+        and len(x_coarse_gen.shape) == 2
+        and 1 <= x_coarse_gen.shape[0] <= N_FINE_CODEBOOKS - 1
+        and x_coarse_gen.shape[1] > 0
+        and x_coarse_gen.min() >= 0
+        and x_coarse_gen.max() <= CODEBOOK_SIZE - 1
+    )
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        x_fine_history = history_prompt["fine_prompt"]
+        assert (
+            isinstance(x_fine_history, np.ndarray)
+            and len(x_fine_history.shape) == 2
+            and x_fine_history.shape[0] == N_FINE_CODEBOOKS
+            and x_fine_history.shape[1] >= 0
+            and x_fine_history.min() >= 0
+            and x_fine_history.max() <= CODEBOOK_SIZE - 1
+        )
+    else:
+        x_fine_history = None
+    n_coarse = x_coarse_gen.shape[0]
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "fine" not in models:
+        preload_models()
+    model = models["fine"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["fine"])
+    device = next(model.parameters()).device
+    # make input arr
+    in_arr = np.vstack(
+        [
+            x_coarse_gen,
+            np.zeros((N_FINE_CODEBOOKS - n_coarse, x_coarse_gen.shape[1]))
+            + CODEBOOK_SIZE,  # padding
+        ]
+    ).astype(np.int32)
+    # prepend history if available (max 512)
+    if x_fine_history is not None:
+        x_fine_history = x_fine_history.astype(np.int32)
+        in_arr = np.hstack(
+            [
+                x_fine_history[:, -512:].astype(np.int32),
+                in_arr,
+            ]
+        )
+        n_history = x_fine_history[:, -512:].shape[1]
+    else:
+        n_history = 0
+    n_remove_from_end = 0
+    # need to pad if too short (since non-causal model)
+    if in_arr.shape[1] < 1024:
+        n_remove_from_end = 1024 - in_arr.shape[1]
+        in_arr = np.hstack(
+            [
+                in_arr,
+                np.zeros((N_FINE_CODEBOOKS, n_remove_from_end), dtype=np.int32) + CODEBOOK_SIZE,
+            ]
+        )
+    # we can be lazy about fractional loop and just keep overwriting codebooks
+    n_loops = np.max([0, int(np.ceil((x_coarse_gen.shape[1] - (1024 - n_history)) / 512))]) + 1
+    with _inference_mode():
+        in_arr = torch.tensor(in_arr.T).to(device)
+        for n in tqdm.tqdm(range(n_loops), disable=silent):
+            start_idx = np.min([n * 512, in_arr.shape[0] - 1024])
+            start_fill_idx = np.min([n_history + n * 512, in_arr.shape[0] - 512])
+            rel_start_fill_idx = start_fill_idx - start_idx
+            in_buffer = in_arr[start_idx : start_idx + 1024, :][None]
+            for nn in range(n_coarse, N_FINE_CODEBOOKS):
+                logits = model(nn, in_buffer)
+                if temp is None:
+                    relevant_logits = logits[0, rel_start_fill_idx:, :CODEBOOK_SIZE]
+                    codebook_preds = torch.argmax(relevant_logits, -1)
+                else:
+                    relevant_logits = logits[0, :, :CODEBOOK_SIZE] / temp
+                    probs = F.softmax(relevant_logits, dim=-1)
+                    # multinomial bugged on mps: shuttle to cpu if necessary
+                    inf_device = probs.device
+                    if probs.device.type == "mps":
+                        probs = probs.to("cpu")
+                    codebook_preds = torch.hstack(
+                        [
+                            torch.multinomial(probs[nnn], num_samples=1).to(inf_device)
+                            for nnn in range(rel_start_fill_idx, 1024)
+                        ]
+                    )
+                in_buffer[0, rel_start_fill_idx:, nn] = codebook_preds
+                del logits, codebook_preds
+            # transfer over info into model_in and convert to numpy
+            for nn in range(n_coarse, N_FINE_CODEBOOKS):
+                in_arr[
+                    start_fill_idx : start_fill_idx + (1024 - rel_start_fill_idx), nn
+                ] = in_buffer[0, rel_start_fill_idx:, nn]
+            del in_buffer
+        gen_fine_arr = in_arr.detach().cpu().numpy().squeeze().T
+        del in_arr
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    gen_fine_arr = gen_fine_arr[:, n_history:]
+    if n_remove_from_end > 0:
+        gen_fine_arr = gen_fine_arr[:, :-n_remove_from_end]
+    assert gen_fine_arr.shape[-1] == x_coarse_gen.shape[-1]
+    _clear_cuda_cache()
+    return gen_fine_arr
+
+
+def codec_decode(fine_tokens):
+    """Turn quantized audio codes into audio array using encodec."""
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "codec" not in models:
+        preload_models()
+    model = models["codec"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["codec"])
+    device = next(model.parameters()).device
+    arr = torch.from_numpy(fine_tokens)[None]
+    arr = arr.to(device)
+    arr = arr.transpose(0, 1)
+    emb = model.quantizer.decode(arr)
+    out = model.decoder(emb)
+    audio_arr = out.detach().cpu().numpy().squeeze()
+    del arr, emb, out
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    return audio_arr

bark/model.py

@@ -0,0 +1,218 @@
+"""
+Much of this code is adapted from Andrej Karpathy's NanoGPT
+(https://github.com/karpathy/nanoGPT)
+"""
+import math
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+class LayerNorm(nn.Module):
+    """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """
+
+    def __init__(self, ndim, bias):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(ndim))
+        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None
+
+    def forward(self, input):
+        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
+        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
+        if not self.flash:
+            # print("WARNING: using slow attention. Flash Attention atm needs PyTorch nightly and dropout=0.0")
+            # causal mask to ensure that attention is only applied to the left in the input sequence
+            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
+                                        .view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x, past_kv=None, use_cache=False):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        q, k ,v  = self.c_attn(x).split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        if past_kv is not None:
+            past_key = past_kv[0]
+            past_value = past_kv[1]
+            k = torch.cat((past_key, k), dim=-2)
+            v = torch.cat((past_value, v), dim=-2)
+
+        FULL_T = k.shape[-2]
+
+        if use_cache is True:
+            present = (k, v)
+        else:
+            present = None
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        if self.flash:
+            # efficient attention using Flash Attention CUDA kernels
+            if past_kv is not None:
+                # When `past_kv` is provided, we're doing incremental decoding and `q.shape[2] == 1`: q only contains
+                # the query for the last token. scaled_dot_product_attention interprets this as the first token in the
+                # sequence, so if is_causal=True it will mask out all attention from it. This is not what we want, so 
+                # to work around this we set is_causal=False.
+                is_causal = False
+            else:
+                is_causal = True
+
+            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout, is_causal=is_causal)
+        else:
+            # manual implementation of attention
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = att.masked_fill(self.bias[:,:,FULL_T-T:FULL_T,:FULL_T] == 0, float('-inf'))
+            att = F.softmax(att, dim=-1)
+            att = self.attn_dropout(att)
+            y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_dropout(self.c_proj(y))
+        return (y, present)
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
+        self.dropout = nn.Dropout(config.dropout)
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        x = self.dropout(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
+        self.mlp = MLP(config)
+        self.layer_idx = layer_idx
+
+    def forward(self, x, past_kv=None, use_cache=False):
+        attn_output, prev_kvs = self.attn(self.ln_1(x), past_kv=past_kv, use_cache=use_cache)
+        x = x + attn_output
+        x = x + self.mlp(self.ln_2(x))
+        return (x, prev_kvs)
+
+@dataclass
+class GPTConfig:
+    block_size: int = 1024
+    input_vocab_size: int = 10_048
+    output_vocab_size: int = 10_048
+    n_layer: int = 12
+    n_head: int = 12
+    n_embd: int = 768
+    dropout: float = 0.0
+    bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.input_vocab_size is not None
+        assert config.output_vocab_size is not None
+        assert config.block_size is not None
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.input_vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            drop = nn.Dropout(config.dropout),
+            h = nn.ModuleList([Block(config, idx) for idx in range(config.n_layer)]),
+            ln_f = LayerNorm(config.n_embd, bias=config.bias),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            n_params -= self.transformer.wte.weight.numel()
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+    def forward(self, idx, merge_context=False, past_kv=None, position_ids=None, use_cache=False):
+        device = idx.device
+        b, t = idx.size()
+        if past_kv is not None:
+            assert t == 1
+            tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
+        else:
+            if merge_context:
+                assert(idx.shape[1] >= 256+256+1)
+                t = idx.shape[1] - 256
+            else:
+                assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
+
+            # forward the GPT model itself
+            if merge_context:
+                tok_emb = torch.cat([
+                    self.transformer.wte(idx[:,:256]) + self.transformer.wte(idx[:,256:256+256]),
+                    self.transformer.wte(idx[:,256+256:])
+                ], dim=1)
+            else:
+                tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
+
+        if past_kv is None:
+            past_length = 0
+            past_kv = tuple([None] * len(self.transformer.h))
+        else:
+            past_length = past_kv[0][0].size(-2)
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, t + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0) # shape (1, t)
+            assert position_ids.shape == (1, t)
+
+        pos_emb = self.transformer.wpe(position_ids) # position embeddings of shape (1, t, n_embd)
+
+        x = self.transformer.drop(tok_emb + pos_emb)
+
+        new_kv = () if use_cache else None
+
+        for i, (block, past_layer_kv) in enumerate(zip(self.transformer.h, past_kv)):
+            x, kv = block(x, past_kv=past_layer_kv, use_cache=use_cache)
+
+            if use_cache:
+                new_kv = new_kv + (kv,)
+
+        x = self.transformer.ln_f(x)
+
+        # inference-time mini-optimization: only forward the lm_head on the very last position
+        logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
+
+        return (logits, new_kv)

bark/model_fine.py

@@ -0,0 +1,149 @@
+"""
+Much of this code is adapted from Andrej Karpathy's NanoGPT
+(https://github.com/karpathy/nanoGPT)
+"""
+from dataclasses import dataclass
+import math
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from .model import GPT, GPTConfig, MLP
+
+
+class NonCausalSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
+        self.flash = (
+            hasattr(torch.nn.functional, "scaled_dot_product_attention") and self.dropout == 0.0
+        )
+
+    def forward(self, x):
+        B, T, C = x.size()  # batch size, sequence length, embedding dimensionality (n_embd)
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        if self.flash:
+            # efficient attention using Flash Attention CUDA kernels
+            y = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=False
+            )
+        else:
+            # manual implementation of attention
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = F.softmax(att, dim=-1)
+            att = self.attn_dropout(att)
+            y = att @ v  # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = (
+            y.transpose(1, 2).contiguous().view(B, T, C)
+        )  # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_dropout(self.c_proj(y))
+        return y
+
+
+class FineBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = NonCausalSelfAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class FineGPT(GPT):
+    def __init__(self, config):
+        super().__init__(config)
+        del self.lm_head
+        self.config = config
+        self.n_codes_total = config.n_codes_total
+        self.transformer = nn.ModuleDict(
+            dict(
+                wtes=nn.ModuleList(
+                    [
+                        nn.Embedding(config.input_vocab_size, config.n_embd)
+                        for _ in range(config.n_codes_total)
+                    ]
+                ),
+                wpe=nn.Embedding(config.block_size, config.n_embd),
+                drop=nn.Dropout(config.dropout),
+                h=nn.ModuleList([FineBlock(config) for _ in range(config.n_layer)]),
+                ln_f=nn.LayerNorm(config.n_embd),
+            )
+        )
+        self.lm_heads = nn.ModuleList(
+            [
+                nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
+                for _ in range(config.n_codes_given, self.n_codes_total)
+            ]
+        )
+        for i in range(self.n_codes_total - config.n_codes_given):
+            self.transformer.wtes[i + 1].weight = self.lm_heads[i].weight
+
+    def forward(self, pred_idx, idx):
+        device = idx.device
+        b, t, codes = idx.size()
+        assert (
+            t <= self.config.block_size
+        ), f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
+        assert pred_idx > 0, "cannot predict 0th codebook"
+        assert codes == self.n_codes_total, (b, t, codes)
+        pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0)  # shape (1, t)
+
+        # forward the GPT model itself
+        tok_embs = [
+            wte(idx[:, :, i]).unsqueeze(-1) for i, wte in enumerate(self.transformer.wtes)
+        ]  # token embeddings of shape (b, t, n_embd)
+        tok_emb = torch.cat(tok_embs, dim=-1)
+        pos_emb = self.transformer.wpe(pos)  # position embeddings of shape (1, t, n_embd)
+        x = tok_emb[:, :, :, : pred_idx + 1].sum(dim=-1)
+        x = self.transformer.drop(x + pos_emb)
+        for block in self.transformer.h:
+            x = block(x)
+        x = self.transformer.ln_f(x)
+        logits = self.lm_heads[pred_idx - self.config.n_codes_given](x)
+        return logits
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            for wte in self.transformer.wtes:
+                n_params -= wte.weight.numel()
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+
+@dataclass
+class FineGPTConfig(GPTConfig):
+    n_codes_total: int = 8
+    n_codes_given: int = 1

bark/settings.py

@@ -0,0 +1,7 @@
+import os
+
+def initenv(args):
+    os.environ['SUNO_USE_SMALL_MODELS'] = str("-smallmodels" in args)
+    os.environ['BARK_FORCE_CPU'] = str("-forcecpu" in args)
+    os.environ['SUNO_ENABLE_MPS'] = str("-enablemps" in args)
+    os.environ['SUNO_OFFLOAD_CPU'] = str("-offloadcpu" in args)

best_model.pth.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:017bfd8907c80bb5857d65d0223f0e4e4b9d699ef52e2a853d9cc7eb7e308cf0
+size 379957289

config.json

@@ -0,0 +1,373 @@
+{
+    "model": "vits",
+    "run_name": "vits_tts-portuguese",
+    "run_description": "",
+    "epochs": 1000,
+    "batch_size": 52,
+    "eval_batch_size": 52,
+    "mixed_precision": false,
+    "scheduler_after_epoch": true,
+    "run_eval": true,
+    "test_delay_epochs": -1,
+    "print_eval": true,
+    "dashboard_logger": "tensorboard",
+    "print_step": 25,
+    "plot_step": 100,
+    "model_param_stats": false,
+    "project_name": null,
+    "log_model_step": 10000,
+    "wandb_entity": null,
+    "save_step": 10000,
+    "checkpoint": true,
+    "keep_all_best": false,
+    "keep_after": 10000,
+    "num_loader_workers": 4,
+    "num_eval_loader_workers": 4,
+    "use_noise_augment": false,
+    "use_language_weighted_sampler": true,
+    "output_path": "../checkpoints/VITS-multilingual/VITS_fixes/new/new-SE/use_noise_aument_false/xlarge-ZS-PT-VCTK/pt-en+LibriTTS-fr/speaker_encoder_as_loss_9_alpha/mixed-p-false-bug-SDP-fixed/",
+    "distributed_backend": "nccl",
+    "distributed_url": "tcp://localhost:54321",
+    "audio": {
+        "fft_size": 1024,
+        "win_length": 1024,
+        "hop_length": 256,
+        "frame_shift_ms": null,
+        "frame_length_ms": null,
+        "stft_pad_mode": "reflect",
+        "sample_rate": 16000,
+        "resample": false,
+        "preemphasis": 0.0,
+        "ref_level_db": 20,
+        "do_sound_norm": false,
+        "log_func": "np.log",
+        "do_trim_silence": true,
+        "trim_db": 45,
+        "power": 1.5,
+        "griffin_lim_iters": 60,
+        "num_mels": 80,
+        "mel_fmin": 0.0,
+        "mel_fmax": null,
+        "spec_gain": 1,
+        "do_amp_to_db_linear": false,
+        "do_amp_to_db_mel": true,
+        "signal_norm": false,
+        "min_level_db": -100,
+        "symmetric_norm": true,
+        "max_norm": 4.0,
+        "clip_norm": true,
+        "stats_path": null
+    },
+    "use_phonemes": false,
+    "use_espeak_phonemes": false,
+    "phoneme_language": "pt-br",
+    "compute_input_seq_cache": false,
+    "text_cleaner": "multilingual_cleaners",
+    "enable_eos_bos_chars": false,
+    "test_sentences_file": "",
+    "phoneme_cache_path": null,
+    "characters": {
+        "pad": "_",
+        "eos": "&",
+        "bos": "*",
+        "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz\u00af\u00b7\u00df\u00e0\u00e1\u00e2\u00e3\u00e4\u00e6\u00e7\u00e8\u00e9\u00ea\u00eb\u00ec\u00ed\u00ee\u00ef\u00f1\u00f2\u00f3\u00f4\u00f5\u00f6\u00f9\u00fa\u00fb\u00fc\u00ff\u0101\u0105\u0107\u0113\u0119\u011b\u012b\u0131\u0142\u0144\u014d\u0151\u0153\u015b\u016b\u0171\u017a\u017c\u01ce\u01d0\u01d2\u01d4\u0430\u0431\u0432\u0433\u0434\u0435\u0436\u0437\u0438\u0439\u043a\u043b\u043c\u043d\u043e\u043f\u0440\u0441\u0442\u0443\u0444\u0445\u0446\u0447\u0448\u0449\u044a\u044b\u044c\u044d\u044e\u044f\u0451\u0454\u0456\u0457\u0491\u2013!'(),-.:;? ",
+        "punctuations": "!'(),-.:;? ",
+        "phonemes": "iy\u0268\u0289\u026fu\u026a\u028f\u028ae\u00f8\u0258\u0259\u0275\u0264o\u025b\u0153\u025c\u025e\u028c\u0254\u00e6\u0250a\u0276\u0251\u0252\u1d7b\u0298\u0253\u01c0\u0257\u01c3\u0284\u01c2\u0260\u01c1\u029bpbtd\u0288\u0256c\u025fk\u0261q\u0262\u0294\u0274\u014b\u0272\u0273n\u0271m\u0299r\u0280\u2c71\u027e\u027d\u0278\u03b2fv\u03b8\u00f0sz\u0283\u0292\u0282\u0290\u00e7\u029dx\u0263\u03c7\u0281\u0127\u0295h\u0266\u026c\u026e\u028b\u0279\u027bj\u0270l\u026d\u028e\u029f\u02c8\u02cc\u02d0\u02d1\u028dw\u0265\u029c\u02a2\u02a1\u0255\u0291\u027a\u0267\u025a\u02de\u026b'\u0303' ",
+        "unique": true
+    },
+    "batch_group_size": 0,
+    "loss_masking": null,
+    "min_seq_len": 90,
+    "max_seq_len": 270,
+    "compute_f0": false,
+    "compute_linear_spec": true,
+    "add_blank": true,
+    "datasets": [
+        {
+            "name": "vctk",
+            "path": "../../datasets/VCTK-Corpus-removed-silence_16Khz/",
+            "meta_file_train": null,
+            "ununsed_speakers": [
+                "p225",
+                "p234",
+                "p238",
+                "p245",
+                "p248",
+                "p261",
+                "p294",
+                "p302",
+                "p326",
+                "p335",
+                "p347"
+            ],
+            "language": "en",
+            "meta_file_val": null,
+            "meta_file_attn_mask": ""
+        },
+        {
+            "name": "libri_tts",
+            "path": "../../datasets/LibriTTS/LibriTTS/dataset-preprocessed-clean-100-and-360/dataset-22k/",
+            "meta_file_train": "metadata_all.csv",
+            "ununsed_speakers": null,
+            "language": "en",
+            "meta_file_val": "dev-clean_500.csv",
+            "meta_file_attn_mask": ""
+        },
+        {
+            "name": "brspeech",
+            "path": "../../datasets/TTS-Portuguese-Corpus_16khz/",
+            "meta_file_train": "train_TTS-Portuguese_Corpus_metadata.csv",
+            "ununsed_speakers": null,
+            "language": "pt-br",
+            "meta_file_val": "eval_TTS-Portuguese_Corpus_metadata.csv",
+            "meta_file_attn_mask": ""
+        },
+        {
+            "name": "mailabs",
+            "path": "../../datasets/M-AILABS/fr_FR",
+            "meta_file_train": "",
+            "ununsed_speakers": null,
+            "language": "fr-fr",
+            "meta_file_val": null,
+            "meta_file_attn_mask": null
+        }
+    ],
+    "optimizer": "AdamW",
+    "optimizer_params": {
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-09,
+        "weight_decay": 0.01
+    },
+    "lr_scheduler": "",
+    "lr_scheduler_params": null,
+    "test_sentences": [
+        [
+            "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
+            "VCTK_p225",
+            null,
+            "en"
+        ],
+        [
+            "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
+            "ED",
+            null,
+            "en"
+        ],
+        [
+            "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
+            "bernard",
+            null,
+            "en"
+        ],
+        [
+            "This cake is great. It's so delicious and moist.",
+            "VCTK_p234",
+            null,
+            "en"
+        ],
+        [
+            "This cake is great. It's so delicious and moist.",
+            "ED",
+            null,
+            "en"
+        ],
+        [
+            "This cake is great. It's so delicious and moist.",
+            "ezwa",
+            null,
+            "en"
+        ],
+        [
+            "Hoje \u00e9 fundamental encontrar a raz\u00e3o da exist\u00eancia humana.",
+            "ED",
+            null,
+            "pt-br"
+        ],
+        [
+            "Hoje \u00e9 fundamental encontrar a raz\u00e3o da exist\u00eancia humana.",
+            "VCTK_p238",
+            null,
+            "pt-br"
+        ],
+        [
+            "Hoje \u00e9 fundamental encontrar a raz\u00e3o da exist\u00eancia humana.",
+            "gilles_g_le_blanc",
+            null,
+            "pt-br"
+        ],
+        [
+            "Em muitas cidades a popula\u00e7\u00e3o est\u00e1 diminuindo.",
+            "ED",
+            null,
+            "pt-br"
+        ],
+        [
+            "Em muitas cidades a popula\u00e7\u00e3o est\u00e1 diminuindo.",
+            "VCTK_p245",
+            null,
+            "pt-br"
+        ],
+        [
+            "Em muitas cidades a popula\u00e7\u00e3o est\u00e1 diminuindo.",
+            "nadine_eckert_boulet",
+            null,
+            "pt-br"
+        ],
+        [
+            "Il m'a fallu beaucoup de temps pour d\u00e9velopper une voix, et maintenant que je l'ai, je ne vais pas me taire.",
+            "VCTK_p245",
+            null,
+            "fr-fr"
+        ],
+        [
+            "Il m'a fallu beaucoup de temps pour d\u00e9velopper une voix, et maintenant que je l'ai, je ne vais pas me taire.",
+            "ED",
+            null,
+            "fr-fr"
+        ],
+        [
+            "Il m'a fallu beaucoup de temps pour d\u00e9velopper une voix, et maintenant que je l'ai, je ne vais pas me taire.",
+            "ezwa",
+            null,
+            "fr-fr"
+        ],
+        [
+            "Il m'a fallu beaucoup de temps pour d\u00e9velopper une voix, et maintenant que je l'ai, je ne vais pas me taire.",
+            "bernard",
+            null,
+            "fr-fr"
+        ],
+        [
+            "Il m'a fallu beaucoup de temps pour d\u00e9velopper une voix, et maintenant que je l'ai, je ne vais pas me taire.",
+            "gilles_g_le_blanc",
+            null,
+            "fr-fr"
+        ],
+        [
+            "Il m'a fallu beaucoup de temps pour d\u00e9velopper une voix, et maintenant que je l'ai, je ne vais pas me taire.",
+            "nadine_eckert_boulet",
+            null,
+            "fr-fr"
+        ],
+        [
+            "Il m'a fallu beaucoup de temps pour d\u00e9velopper une voix, et maintenant que je l'ai, je ne vais pas me taire.",
+            "zeckou",
+            null,
+            "fr-fr"
+        ]
+    ],
+    "use_speaker_embedding": true,
+    "use_d_vector_file": true,
+    "d_vector_dim": 512,
+    "model_args": {
+        "num_chars": 165,
+        "out_channels": 513,
+        "spec_segment_size": 62,
+        "hidden_channels": 192,
+        "hidden_channels_ffn_text_encoder": 768,
+        "num_heads_text_encoder": 2,
+        "num_layers_text_encoder": 10,
+        "kernel_size_text_encoder": 3,
+        "dropout_p_text_encoder": 0.1,
+        "dropout_p_duration_predictor": 0.5,
+        "kernel_size_posterior_encoder": 5,
+        "dilation_rate_posterior_encoder": 1,
+        "num_layers_posterior_encoder": 16,
+        "kernel_size_flow": 5,
+        "dilation_rate_flow": 1,
+        "num_layers_flow": 4,
+        "resblock_type_decoder": 1,
+        "resblock_kernel_sizes_decoder": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes_decoder": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates_decoder": [
+            8,
+            8,
+            2,
+            2
+        ],
+        "upsample_initial_channel_decoder": 512,
+        "upsample_kernel_sizes_decoder": [
+            16,
+            16,
+            4,
+            4
+        ],
+        "use_sdp": true,
+        "noise_scale": 1.0,
+        "inference_noise_scale": 0.667,
+        "length_scale": 1,
+        "noise_scale_dp": 1.0,
+        "inference_noise_scale_dp": 0.8,
+        "max_inference_len": null,
+        "init_discriminator": true,
+        "use_spectral_norm_disriminator": false,
+        "use_speaker_embedding": true,
+        "num_speakers": 1244,
+        "speakers_file": null,
+        "d_vector_file": "../speaker_embeddings/new-SE/VCTK-LibriTTS+TTS-PT+MAILABS-FR/speakers.json",
+        "speaker_embedding_channels": 512,
+        "use_d_vector_file": true,
+        "d_vector_dim": 512,
+        "detach_dp_input": true,
+        "use_language_embedding": true,
+        "embedded_language_dim": 4,
+        "num_languages": 3,
+        "use_speaker_encoder_as_loss": true,
+        "speaker_encoder_config_path": "../checkpoints/Speaker_Encoder/Resnet-original-paper/config.json",
+        "speaker_encoder_model_path": "../checkpoints/Speaker_Encoder/Resnet-original-paper/converted_checkpoint.pth.tar",
+        "fine_tuning_mode": 0,
+        "freeze_encoder": false,
+        "freeze_DP": false,
+        "freeze_PE": false,
+        "freeze_flow_decoder": false,
+        "freeze_waveform_decoder": false
+    },
+    "grad_clip": [
+        5.0,
+        5.0
+    ],
+    "lr_gen": 0.0002,
+    "lr_disc": 0.0002,
+    "lr_scheduler_gen": "ExponentialLR",
+    "lr_scheduler_gen_params": {
+        "gamma": 0.999875,
+        "last_epoch": -1
+    },
+    "lr_scheduler_disc": "ExponentialLR",
+    "lr_scheduler_disc_params": {
+        "gamma": 0.999875,
+        "last_epoch": -1
+    },
+    "kl_loss_alpha": 1.0,
+    "disc_loss_alpha": 1.0,
+    "gen_loss_alpha": 1.0,
+    "feat_loss_alpha": 1.0,
+    "mel_loss_alpha": 45.0,
+    "dur_loss_alpha": 1.0,
+    "speaker_encoder_loss_alpha": 9.0,
+    "return_wav": true,
+    "r": 1
+}

config.yaml

@@ -0,0 +1,8 @@
+input_text_desired_length: 110
+input_text_max_length: 170
+selected_theme: gradio/soft
+server_name: ''
+server_port: 0
+server_share: false
+silence_between_sentences: 250
+silence_between_speakers: 500

config_se.json

@@ -0,0 +1,119 @@
+{
+    "model": "speaker_encoder",
+    "run_name": "speaker_encoder",
+    "run_description": "resnet speaker encoder trained with commonvoice all languages dev and train, Voxceleb 1 dev and Voxceleb 2 dev",
+    "epochs": 100000,
+    "batch_size": null,
+    "eval_batch_size": null,
+    "mixed_precision": false,
+    "run_eval": true,
+    "test_delay_epochs": 0,
+    "print_eval": false,
+    "print_step": 50,
+    "tb_plot_step": 100,
+    "tb_model_param_stats": false,
+    "save_step": 1000,
+    "checkpoint": true,
+    "keep_all_best": false,
+    "keep_after": 10000,
+    "num_loader_workers": 8,
+    "num_val_loader_workers": 0,
+    "use_noise_augment": false,
+    "output_path": "../checkpoints/speaker_encoder/language_balanced/normalized/angleproto-4-samples-by-speakers/",
+    "distributed_backend": "nccl",
+    "distributed_url": "tcp://localhost:54321",
+    "audio": {
+        "fft_size": 512,
+        "win_length": 400,
+        "hop_length": 160,
+        "frame_shift_ms": null,
+        "frame_length_ms": null,
+        "stft_pad_mode": "reflect",
+        "sample_rate": 16000,
+        "resample": false,
+        "preemphasis": 0.97,
+        "ref_level_db": 20,
+        "do_sound_norm": false,
+        "do_trim_silence": false,
+        "trim_db": 60,
+        "power": 1.5,
+        "griffin_lim_iters": 60,
+        "num_mels": 64,
+        "mel_fmin": 0.0,
+        "mel_fmax": 8000.0,
+        "spec_gain": 20,
+        "signal_norm": false,
+        "min_level_db": -100,
+        "symmetric_norm": false,
+        "max_norm": 4.0,
+        "clip_norm": false,
+        "stats_path": null
+    },
+    "datasets": [
+        {
+            "name": "voxceleb2",
+            "path": "/workspace/scratch/ecasanova/datasets/VoxCeleb/vox2_dev_aac/",
+            "meta_file_train": null,
+            "ununsed_speakers": null,
+            "meta_file_val": null,
+            "meta_file_attn_mask": "",
+            "language": "voxceleb"
+        }
+    ],
+    "model_params": {
+        "model_name": "resnet",
+        "input_dim": 64,
+        "use_torch_spec": true,
+        "log_input": true,
+        "proj_dim": 512
+    },
+    "audio_augmentation": {
+        "p": 0.5,
+        "rir": {
+            "rir_path": "/workspace/store/ecasanova/ComParE/RIRS_NOISES/simulated_rirs/",
+            "conv_mode": "full"
+        },
+        "additive": {
+            "sounds_path": "/workspace/store/ecasanova/ComParE/musan/",
+            "speech": {
+                "min_snr_in_db": 13,
+                "max_snr_in_db": 20,
+                "min_num_noises": 1,
+                "max_num_noises": 1
+            },
+            "noise": {
+                "min_snr_in_db": 0,
+                "max_snr_in_db": 15,
+                "min_num_noises": 1,
+                "max_num_noises": 1
+            },
+            "music": {
+                "min_snr_in_db": 5,
+                "max_snr_in_db": 15,
+                "min_num_noises": 1,
+                "max_num_noises": 1
+            }
+        },
+        "gaussian": {
+            "p": 0.0,
+            "min_amplitude": 0.0,
+            "max_amplitude": 1e-05
+        }
+    },
+    "storage": {
+        "sample_from_storage_p": 0.5,
+        "storage_size": 40
+    },
+    "max_train_step": 1000000,
+    "loss": "angleproto",
+    "grad_clip": 3.0,
+    "lr": 0.0001,
+    "lr_decay": false,
+    "warmup_steps": 4000,
+    "wd": 1e-06,
+    "steps_plot_stats": 100,
+    "num_speakers_in_batch": 100,
+    "num_utters_per_speaker": 4,
+    "skip_speakers": true,
+    "voice_len": 2.0
+}

id3tagging.py

@@ -0,0 +1,14 @@
+from mutagen.wave import WAVE
+from mutagen.id3._frames import *
+
+def add_id3_tag(filename, text, speakername, seed):
+    audio = WAVE(filename)
+    if speakername == None:
+        speakername = "Unconditional"
+
+    # write id3 tag with text truncated to 60 chars, as a precaution...
+    audio["TIT2"] = TIT2(encoding=3, text=text[:60])
+    audio["TPE1"] = TPE1(encoding=3, text=f"Voice {speakername} using Seed={seed}")
+    audio["TPUB"] = TPUB(encoding=3, text="Bark by Suno AI")
+    audio["COMMENT"] = COMM(encoding=3, text="Generated with Bark GUI - Text-Prompted Generative Audio Model. Visit https://github.com/C0untFloyd/bark-gui")
+    audio.save()

language_ids.json

@@ -0,0 +1,5 @@
+{
+    "en": 0,
+    "fr-fr": 1,
+    "pt-br": 2
+}

parseinput.py

@@ -0,0 +1,129 @@
+import re
+import xml.etree.ElementTree as ET
+from xml.sax import saxutils
+#import nltk
+
+# Chunked generation originally from https://github.com/serp-ai/bark-with-voice-clone
+def split_and_recombine_text(text, desired_length=100, max_length=150):
+    # return nltk.sent_tokenize(text)
+
+    # from https://github.com/neonbjb/tortoise-tts
+    """Split text it into chunks of a desired length trying to keep sentences intact."""
+    # normalize text, remove redundant whitespace and convert non-ascii quotes to ascii
+    text = re.sub(r"\n\n+", "\n", text)
+    text = re.sub(r"\s+", " ", text)
+    text = re.sub(r"[“”]", '"', text)
+
+    rv = []
+    in_quote = False
+    current = ""
+    split_pos = []
+    pos = -1
+    end_pos = len(text) - 1
+
+    def seek(delta):
+        nonlocal pos, in_quote, current
+        is_neg = delta < 0
+        for _ in range(abs(delta)):
+            if is_neg:
+                pos -= 1
+                current = current[:-1]
+            else:
+                pos += 1
+                current += text[pos]
+            if text[pos] == '"':
+                in_quote = not in_quote
+        return text[pos]
+
+    def peek(delta):
+        p = pos + delta
+        return text[p] if p < end_pos and p >= 0 else ""
+
+    def commit():
+        nonlocal rv, current, split_pos
+        rv.append(current)
+        current = ""
+        split_pos = []
+
+    while pos < end_pos:
+        c = seek(1)
+        # do we need to force a split?
+        if len(current) >= max_length:
+            if len(split_pos) > 0 and len(current) > (desired_length / 2):
+                # we have at least one sentence and we are over half the desired length, seek back to the last split
+                d = pos - split_pos[-1]
+                seek(-d)
+            else:
+                # no full sentences, seek back until we are not in the middle of a word and split there
+                while c not in "!?.,\n " and pos > 0 and len(current) > desired_length:
+                    c = seek(-1)
+            commit()
+        # check for sentence boundaries
+        elif not in_quote and (c in "!?]\n" or (c == "." and peek(1) in "\n ")):
+            # seek forward if we have consecutive boundary markers but still within the max length
+            while (
+                pos < len(text) - 1 and len(current) < max_length and peek(1) in "!?.]"
+            ):
+                c = seek(1)
+            split_pos.append(pos)
+            if len(current) >= desired_length:
+                commit()
+        # treat end of quote as a boundary if its followed by a space or newline
+        elif in_quote and peek(1) == '"' and peek(2) in "\n ":
+            seek(2)
+            split_pos.append(pos)
+    rv.append(current)
+
+    # clean up, remove lines with only whitespace or punctuation
+    rv = [s.strip() for s in rv]
+    rv = [s for s in rv if len(s) > 0 and not re.match(r"^[\s\.,;:!?]*$", s)]
+
+    return rv
+
+def is_ssml(value):
+    try:
+        ET.fromstring(value)
+    except ET.ParseError:
+        return False
+    return True
+
+def build_ssml(rawtext, selected_voice):
+    texts = rawtext.split("\n")
+    joinedparts = ""
+    for textpart in texts:
+        textpart = textpart.strip()
+        if len(textpart) < 1:
+            continue
+        joinedparts = joinedparts + f"\n<voice name=\"{selected_voice}\">{saxutils.escape(textpart)}</voice>"
+    ssml = f"""<?xml version="1.0"?>
+<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis"
+         xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+         xsi:schemaLocation="http://www.w3.org/2001/10/synthesis
+                   http://www.w3.org/TR/speech-synthesis/synthesis.xsd"
+         xml:lang="en-US">
+         {joinedparts}
+</speak>
+    """
+    return ssml
+
+def create_clips_from_ssml(ssmlinput):
+    # Parse the XML
+    tree = ET.ElementTree(ET.fromstring(ssmlinput))
+    root = tree.getroot()
+
+    # Create an empty list
+    voice_list = []
+
+    # Loop through all voice tags
+    for voice in root.iter('{http://www.w3.org/2001/10/synthesis}voice'):
+        # Extract the voice name attribute and the content text
+        voice_name = voice.attrib['name']
+        voice_content = voice.text.strip() if voice.text else ''
+        if(len(voice_content) > 0):
+            parts = split_and_recombine_text(voice_content)
+            for p in parts:
+                if(len(p) > 1):
+                    # add to tuple list
+                    voice_list.append((voice_name, p))
+    return voice_list
+

pyproject.toml

@@ -0,0 +1,63 @@
+[build-system]
+requires = ["setuptools"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "bark-ui-enhanced"
+version = "0.4.7"
+description = "Bark text to audio model with addition features and a Web UI"
+readme = "README.md"
+requires-python = ">=3.8"
+authors =  [
+    {name = "Suno Inc (original Bark)", email = "hello@suno.ai"},
+    {name = "Count Floyd"},
+]
+# MIT License
+license = {file = "LICENSE"}
+
+dependencies = [
+    "boto3",
+    "encodec",
+    "funcy",
+    "mutagen",
+    "numpy",
+    "pytorch_seed",
+    "scipy",
+    "tokenizers",
+    "torch",
+    "tqdm",
+    "transformers",
+    "pyyaml"
+]
+
+[project.urls]
+source = "https://github.com/C0untFloyd/bark-gui"
+
+[project.optional-dependencies]
+dev = [
+    "bandit",
+    "black",
+    "codecov",
+    "flake8",
+    "huggingface-hub>=0.14.1",
+    "hypothesis>=6.14,<7",
+    "isort>=5.0.0,<6",
+    "jupyter",
+    "mypy",
+    "nbconvert",
+    "nbformat",
+    "pydocstyle",
+    "pylint",
+    "pytest",
+    "pytest-cov",
+]
+
+[tool.setuptools]
+packages = ["bark"]
+
+[tool.setuptools.package-data]
+bark = ["assets/prompts/*.npz", "assets/prompts/v2/*.npz"]
+
+
+[tool.black]
+line-length = 100

requirements.txt

@@ -0,0 +1,10 @@
+pytorch_seed
+encodec
+funcy
+transformers
+scipy
+mutagen
+git+https://github.com/Edresson/Coqui-TTS@multilingual-torchaudio-SE
+torchaudio
+pydub 
+ffmpeg-normalize==1.21.0

settings.py

@@ -0,0 +1,39 @@
+import yaml
+
+class Settings:
+    def __init__(self, config_file):
+        self.config_file = config_file
+        self.load()
+
+    def load(self):
+        try:
+            with open(self.config_file, 'r') as f:
+                data = yaml.load(f, Loader=yaml.FullLoader)
+            self.selected_theme = data.get('selected_theme', "gstaff/xkcd")
+            self.server_name = data.get('server_name', "")
+            self.server_port = data.get('server_port', 0)
+            self.server_share = data.get('server_share', False)
+            self.input_text_desired_length = data.get('input_text_desired_length', 110)
+            self.input_text_max_length = data.get('input_text_max_length', 170)
+            self.silence_sentence = data.get('silence_between_sentences', 250)
+            self.silence_speakers = data.get('silence_between_speakers', 500)
+
+        except:
+            self.selected_theme = "gstaff/xkcd"
+
+    def save(self):
+        data = {
+            'selected_theme': self.selected_theme,
+            'server_name': self.server_name,
+            'server_port': self.server_port,
+            'server_share': self.server_share,
+            'input_text_desired_length' : self.input_text_desired_length,
+            'input_text_max_length' : self.input_text_max_length, 
+            'silence_between_sentences': self.silence_sentence,
+            'silence_between_speakers': self.silence_speakers,
+        }
+        with open(self.config_file, 'w') as f:
+            yaml.dump(data, f)
+
+
+

setup.py

@@ -0,0 +1,3 @@
+from setuptools import setup
+
+setup()