Sync from GitHub repo

This Space is synced from the GitHub repo: https://github.com/SWivid/F5-TTS. Please submit contributions to the Space there

.github/workflows/sync-hf.yaml

@@ -0,0 +1,18 @@
+name: Sync to HF Space
+
+on:
+  push:
+    branches:
+      - main
+
+jobs:
+  trigger_curl:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Send cURL POST request
+      run: |
+        curl -X POST https://mrfakename-sync-f5.hf.space/gradio_api/call/refresh \
+        -s \
+        -H "Content-Type: application/json" \
+        -d "{\"data\": [\"${{ secrets.REFRESH_PASSWORD }}\"]}"

Dockerfile

@@ -0,0 +1,25 @@
+FROM pytorch/pytorch:2.4.0-cuda12.4-cudnn9-devel
+
+USER root
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+LABEL github_repo="https://github.com/SWivid/F5-TTS"
+
+RUN set -x \
+    && apt-get update \
+    && apt-get -y install wget curl man git less openssl libssl-dev unzip unar build-essential aria2 tmux vim \
+    && apt-get install -y openssh-server sox libsox-fmt-all libsox-fmt-mp3 libsndfile1-dev ffmpeg \
+    && rm -rf /var/lib/apt/lists/* \
+    && apt-get clean
+
+WORKDIR /workspace
+
+RUN git clone https://github.com/SWivid/F5-TTS.git \
+    && cd F5-TTS \
+    && pip install --no-cache-dir -r requirements.txt \
+    && pip install --no-cache-dir -r requirements_eval.txt
+
+ENV SHELL=/bin/bash
+
+WORKDIR /workspace/F5-TTS

README.md

@@ -7,7 +7,7 @@ sdk: gradio
 app_file: app.py
 pinned: true
 short_description: 'F5-TTS & E2-TTS: Zero-Shot Voice Cloning (Unofficial Demo)'
-sdk_version: 5.1.0
+sdk_version: 4.44.1
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

README_REPO.md

@@ -2,15 +2,20 @@
 
 [![python](https://img.shields.io/badge/Python-3.10-brightgreen)](https://github.com/SWivid/F5-TTS)
 [![arXiv](https://img.shields.io/badge/arXiv-2410.06885-b31b1b.svg?logo=arXiv)](https://arxiv.org/abs/2410.06885)
-[![demo](https://img.shields.io/badge/GitHub-Demo%20page-blue.svg)](https://swivid.github.io/F5-TTS/)
-[![space](https://img.shields.io/badge/🤗-Space%20demo-yellow)](https://huggingface.co/spaces/mrfakename/E2-F5-TTS)
+[![demo](https://img.shields.io/badge/GitHub-Demo%20page-orange.svg)](https://swivid.github.io/F5-TTS/)
+[![hfspace](https://img.shields.io/badge/🤗-Space%20demo-yellow)](https://huggingface.co/spaces/mrfakename/E2-F5-TTS)
+[![msspace](https://img.shields.io/badge/🤖-Space%20demo-blue)](https://modelscope.cn/studios/modelscope/E2-F5-TTS)
+[![lab](https://img.shields.io/badge/X--LANCE-Lab-grey?labelColor=lightgrey)](https://x-lance.sjtu.edu.cn/)
+<img src="https://github.com/user-attachments/assets/12d7749c-071a-427c-81bf-b87b91def670" alt="Watermark" style="width: 40px; height: auto">
 
 **F5-TTS**: Diffusion Transformer with ConvNeXt V2, faster trained and inference.
 
-**E2 TTS**: Flat-UNet Transformer, closest reproduction.
+**E2 TTS**: Flat-UNet Transformer, closest reproduction from [paper](https://arxiv.org/abs/2406.18009).
 
 **Sway Sampling**: Inference-time flow step sampling strategy, greatly improves performance
 
+### Thanks to all the contributors !
+
 ## Installation
 
 Clone the repository:
@@ -48,7 +53,7 @@ python scripts/prepare_emilia.py
 python scripts/prepare_wenetspeech4tts.py
 ```
 
-## Training
+## Training & Finetuning
 
 Once your datasets are prepared, you can start the training process.
 
@@ -60,9 +65,11 @@ accelerate launch train.py
 ```
 An initial guidance on Finetuning [#57](https://github.com/SWivid/F5-TTS/discussions/57).
 
+Gradio UI finetuning with `finetune_gradio.py` see [#143](https://github.com/SWivid/F5-TTS/discussions/143).
+
 ## Inference
 
-To run inference with pretrained models, download the checkpoints from [🤗 Hugging Face](https://huggingface.co/SWivid/F5-TTS), or automatically downloaded with `inference-cli` and `gradio_app`.
+The pretrained model checkpoints can be reached at [🤗 Hugging Face](https://huggingface.co/SWivid/F5-TTS) and [🤖 Model Scope](https://www.modelscope.cn/models/SWivid/F5-TTS_Emilia-ZH-EN), or automatically downloaded with `inference-cli` and `gradio_app`.
 
 Currently support 30s for a single generation, which is the **TOTAL** length of prompt audio and the generated. Batch inference with chunks is supported by `inference-cli` and `gradio_app`. 
 - To avoid possible inference failures, make sure you have seen through the following instructions.
@@ -86,6 +93,9 @@ python inference-cli.py \
 --ref_audio "tests/ref_audio/test_zh_1_ref_short.wav" \
 --ref_text "对，这就是我，万人敬仰的太乙真人。" \
 --gen_text "突然，身边一阵笑声。我看着他们，意气风发地挺直了胸膛，甩了甩那稍显肉感的双臂，轻笑道，我身上的肉，是为了掩饰我爆棚的魅力，否则，岂不吓坏了你们呢？"
+
+# Multi voice
+python inference-cli.py -c samples/story.toml
 ```
 
 ### Gradio App
@@ -148,6 +158,12 @@ bash scripts/eval_infer_batch.sh
 
 ### Objective Evaluation
 
+Install packages for evaluation:
+
+```bash
+pip install -r requirements_eval.txt
+```
+
 **Some Notes**
 
 For faster-whisper with CUDA 11:
@@ -178,11 +194,13 @@ python scripts/eval_librispeech_test_clean.py
 - [lucidrains](https://github.com/lucidrains) initial CFM structure with also [bfs18](https://github.com/bfs18) for discussion
 - [SD3](https://arxiv.org/abs/2403.03206) & [Hugging Face diffusers](https://github.com/huggingface/diffusers) DiT and MMDiT code structure
 - [torchdiffeq](https://github.com/rtqichen/torchdiffeq) as ODE solver, [Vocos](https://huggingface.co/charactr/vocos-mel-24khz) as vocoder
-- [mrfakename](https://x.com/realmrfakename) huggingface space demo ~
 - [FunASR](https://github.com/modelscope/FunASR), [faster-whisper](https://github.com/SYSTRAN/faster-whisper), [UniSpeech](https://github.com/microsoft/UniSpeech) for evaluation tools
 - [ctc-forced-aligner](https://github.com/MahmoudAshraf97/ctc-forced-aligner) for speech edit test
+- [mrfakename](https://x.com/realmrfakename) huggingface space demo ~
+- [f5-tts-mlx](https://github.com/lucasnewman/f5-tts-mlx/tree/main) Implementation of F5-TTS, with the MLX framework.
 
 ## Citation
+If our work and codebase is useful for you, please cite as:
 ```
 @article{chen-etal-2024-f5tts,
       title={F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching}, 
@@ -193,4 +211,4 @@ python scripts/eval_librispeech_test_clean.py
 ```
 ## License
 
-Our code is released under MIT License.
+Our code is released under MIT License. The pre-trained models are licensed under the CC-BY-NC license due to the training data Emilia, which is an in-the-wild dataset. Sorry for any inconvenience this may cause.

app.py

@@ -1,4 +1,3 @@
-import os
 import re
 import torch
 import torchaudio
@@ -17,7 +16,6 @@ from model.utils import (
     save_spectrogram,
 )
 from transformers import pipeline
-import librosa
 import click
 import soundfile as sf
 
@@ -33,19 +31,6 @@ def gpu_decorator(func):
     else:
         return func
 
-
-
-SPLIT_WORDS = [
-    "but", "however", "nevertheless", "yet", "still",
-    "therefore", "thus", "hence", "consequently",
-    "moreover", "furthermore", "additionally",
-    "meanwhile", "alternatively", "otherwise",
-    "namely", "specifically", "for example", "such as",
-    "in fact", "indeed", "notably",
-    "in contrast", "on the other hand", "conversely",
-    "in conclusion", "to summarize", "finally"
-]
-
 device = (
     "cuda"
     if torch.cuda.is_available()
@@ -73,7 +58,6 @@ cfg_strength = 2.0
 ode_method = "euler"
 sway_sampling_coef = -1.0
 speed = 1.0
-# fix_duration = 27  # None or float (duration in seconds)
 fix_duration = None
 
 
@@ -114,104 +98,37 @@ E2TTS_ema_model = load_model(
     "E2-TTS", "E2TTS_Base", UNetT, E2TTS_model_cfg, 1200000
 )
 
-def split_text_into_batches(text, max_chars=200, split_words=SPLIT_WORDS):
-    if len(text.encode('utf-8')) <= max_chars:
-        return [text]
-    if text[-1] not in ['。', '.', '!', '！', '?', '？']:
-        text += '.'
-        
-    sentences = re.split('([。.!?！？])', text)
-    sentences = [''.join(i) for i in zip(sentences[0::2], sentences[1::2])]
-    
-    batches = []
-    current_batch = ""
-    
-    def split_by_words(text):
-        words = text.split()
-        current_word_part = ""
-        word_batches = []
-        for word in words:
-            if len(current_word_part.encode('utf-8')) + len(word.encode('utf-8')) + 1 <= max_chars:
-                current_word_part += word + ' '
-            else:
-                if current_word_part:
-                    # Try to find a suitable split word
-                    for split_word in split_words:
-                        split_index = current_word_part.rfind(' ' + split_word + ' ')
-                        if split_index != -1:
-                            word_batches.append(current_word_part[:split_index].strip())
-                            current_word_part = current_word_part[split_index:].strip() + ' '
-                            break
-                    else:
-                        # If no suitable split word found, just append the current part
-                        word_batches.append(current_word_part.strip())
-                        current_word_part = ""
-                current_word_part += word + ' '
-        if current_word_part:
-            word_batches.append(current_word_part.strip())
-        return word_batches
+def chunk_text(text, max_chars=135):
+    """
+    Splits the input text into chunks, each with a maximum number of characters.
+
+    Args:
+        text (str): The text to be split.
+        max_chars (int): The maximum number of characters per chunk.
+
+    Returns:
+        List[str]: A list of text chunks.
+    """
+    chunks = []
+    current_chunk = ""
+    # Split the text into sentences based on punctuation followed by whitespace
+    sentences = re.split(r'(?<=[;:,.!?])\s+|(?<=[；：，。！？])', text)
 
     for sentence in sentences:
-        if len(current_batch.encode('utf-8')) + len(sentence.encode('utf-8')) <= max_chars:
-            current_batch += sentence
+        if len(current_chunk.encode('utf-8')) + len(sentence.encode('utf-8')) <= max_chars:
+            current_chunk += sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
         else:
-            # If adding this sentence would exceed the limit
-            if current_batch:
-                batches.append(current_batch)
-                current_batch = ""
-            
-            # If the sentence itself is longer than max_chars, split it
-            if len(sentence.encode('utf-8')) > max_chars:
-                # First, try to split by colon
-                colon_parts = sentence.split(':')
-                if len(colon_parts) > 1:
-                    for part in colon_parts:
-                        if len(part.encode('utf-8')) <= max_chars:
-                            batches.append(part)
-                        else:
-                            # If colon part is still too long, split by comma
-                            comma_parts = re.split('[,，]', part)
-                            if len(comma_parts) > 1:
-                                current_comma_part = ""
-                                for comma_part in comma_parts:
-                                    if len(current_comma_part.encode('utf-8')) + len(comma_part.encode('utf-8')) <= max_chars:
-                                        current_comma_part += comma_part + ','
-                                    else:
-                                        if current_comma_part:
-                                            batches.append(current_comma_part.rstrip(','))
-                                        current_comma_part = comma_part + ','
-                                if current_comma_part:
-                                    batches.append(current_comma_part.rstrip(','))
-                            else:
-                                # If no comma, split by words
-                                batches.extend(split_by_words(part))
-                else:
-                    # If no colon, split by comma
-                    comma_parts = re.split('[,，]', sentence)
-                    if len(comma_parts) > 1:
-                        current_comma_part = ""
-                        for comma_part in comma_parts:
-                            if len(current_comma_part.encode('utf-8')) + len(comma_part.encode('utf-8')) <= max_chars:
-                                current_comma_part += comma_part + ','
-                            else:
-                                if current_comma_part:
-                                    batches.append(current_comma_part.rstrip(','))
-                                current_comma_part = comma_part + ','
-                        if current_comma_part:
-                            batches.append(current_comma_part.rstrip(','))
-                    else:
-                        # If no comma, split by words
-                        batches.extend(split_by_words(sentence))
-            else:
-                current_batch = sentence
-    
-    if current_batch:
-        batches.append(current_batch)
-    
-    return batches
+            if current_chunk:
+                chunks.append(current_chunk.strip())
+            current_chunk = sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
 
-@spaces.GPU
-def infer_batch(ref_audio, ref_text, gen_text_batches, exp_name, remove_silence, progress=gr.Progress()):
+    if current_chunk:
+        chunks.append(current_chunk.strip())
+
+    return chunks
+
+@gpu_decorator
+def infer_batch(ref_audio, ref_text, gen_text_batches, exp_name, remove_silence, cross_fade_duration=0.15, progress=gr.Progress()):
     if exp_name == "F5-TTS":
         ema_model = F5TTS_ema_model
     elif exp_name == "E2-TTS":
@@ -269,8 +186,44 @@ def infer_batch(ref_audio, ref_text, gen_text_batches, exp_name, remove_silence,
         generated_waves.append(generated_wave)
         spectrograms.append(generated_mel_spec[0].cpu().numpy())
 
-    # Combine all generated waves
-    final_wave = np.concatenate(generated_waves)
+    # Combine all generated waves with cross-fading
+    if cross_fade_duration <= 0:
+        # Simply concatenate
+        final_wave = np.concatenate(generated_waves)
+    else:
+        final_wave = generated_waves[0]
+        for i in range(1, len(generated_waves)):
+            prev_wave = final_wave
+            next_wave = generated_waves[i]
+
+            # Calculate cross-fade samples, ensuring it does not exceed wave lengths
+            cross_fade_samples = int(cross_fade_duration * target_sample_rate)
+            cross_fade_samples = min(cross_fade_samples, len(prev_wave), len(next_wave))
+
+            if cross_fade_samples <= 0:
+                # No overlap possible, concatenate
+                final_wave = np.concatenate([prev_wave, next_wave])
+                continue
+
+            # Overlapping parts
+            prev_overlap = prev_wave[-cross_fade_samples:]
+            next_overlap = next_wave[:cross_fade_samples]
+
+            # Fade out and fade in
+            fade_out = np.linspace(1, 0, cross_fade_samples)
+            fade_in = np.linspace(0, 1, cross_fade_samples)
+
+            # Cross-faded overlap
+            cross_faded_overlap = prev_overlap * fade_out + next_overlap * fade_in
+
+            # Combine
+            new_wave = np.concatenate([
+                prev_wave[:-cross_fade_samples],
+                cross_faded_overlap,
+                next_wave[cross_fade_samples:]
+            ])
+
+            final_wave = new_wave
 
     # Remove silence
     if remove_silence:
@@ -295,12 +248,8 @@ def infer_batch(ref_audio, ref_text, gen_text_batches, exp_name, remove_silence,
 
     return (target_sample_rate, final_wave), spectrogram_path
 
-@spaces.GPU
-def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, custom_split_words=''):
-    if not custom_split_words.strip():
-        custom_words = [word.strip() for word in custom_split_words.split(',')]
-        global SPLIT_WORDS
-        SPLIT_WORDS = custom_words
+@gpu_decorator
+def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, cross_fade_duration=0.15):
 
     print(gen_text)
 
@@ -308,7 +257,9 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, custom_s
     with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
         aseg = AudioSegment.from_file(ref_audio_orig)
 
-        non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
+        non_silent_segs = silence.split_on_silence(
+            aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=1000
+        )
         non_silent_wave = AudioSegment.silent(duration=0)
         for non_silent_seg in non_silent_segs:
             non_silent_wave += non_silent_seg
@@ -334,18 +285,27 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, custom_s
     else:
         gr.Info("Using custom reference text...")
 
-    # Split the input text into batches
+    # Add the functionality to ensure it ends with ". "
+    if not ref_text.endswith(". "):
+        if ref_text.endswith("."):
+            ref_text += " "
+        else:
+            ref_text += ". "
+
     audio, sr = torchaudio.load(ref_audio)
-    max_chars = int(len(ref_text.encode('utf-8')) / (audio.shape[-1] / sr) * (30 - audio.shape[-1] / sr))
-    gen_text_batches = split_text_into_batches(gen_text, max_chars=max_chars)
+
+    # Use the new chunk_text function to split gen_text
+    max_chars = int(len(ref_text.encode('utf-8')) / (audio.shape[-1] / sr) * (25 - audio.shape[-1] / sr))
+    gen_text_batches = chunk_text(gen_text, max_chars=max_chars)
     print('ref_text', ref_text)
-    for i, gen_text in enumerate(gen_text_batches):
-        print(f'gen_text {i}', gen_text)
+    for i, batch_text in enumerate(gen_text_batches):
+        print(f'gen_text {i}', batch_text)
     
     gr.Info(f"Generating audio using {exp_name} in {len(gen_text_batches)} batches")
-    return infer_batch((audio, sr), ref_text, gen_text_batches, exp_name, remove_silence)
+    return infer_batch((audio, sr), ref_text, gen_text_batches, exp_name, remove_silence, cross_fade_duration)
+
 
-@spaces.GPU
+@gpu_decorator
 def generate_podcast(script, speaker1_name, ref_audio1, ref_text1, speaker2_name, ref_audio2, ref_text2, exp_name, remove_silence):
     # Split the script into speaker blocks
     speaker_pattern = re.compile(f"^({re.escape(speaker1_name)}|{re.escape(speaker2_name)}):", re.MULTILINE)
@@ -429,6 +389,7 @@ with gr.Blocks() as app_credits:
 
 * [mrfakename](https://github.com/fakerybakery) for the original [online demo](https://huggingface.co/spaces/mrfakename/E2-F5-TTS)
 * [RootingInLoad](https://github.com/RootingInLoad) for the podcast generation
+* [jpgallegoar](https://github.com/jpgallegoar) for multiple speech-type generation
 """)
 with gr.Blocks() as app_tts:
     gr.Markdown("# Batched TTS")
@@ -447,12 +408,7 @@ with gr.Blocks() as app_tts:
         remove_silence = gr.Checkbox(
             label="Remove Silences",
             info="The model tends to produce silences, especially on longer audio. We can manually remove silences if needed. Note that this is an experimental feature and may produce strange results. This will also increase generation time.",
-            value=True,
-        )
-        split_words_input = gr.Textbox(
-            label="Custom Split Words",
-            info="Enter custom words to split on, separated by commas. Leave blank to use default list.",
-            lines=2,
+            value=False,
         )
         speed_slider = gr.Slider(
             label="Speed",
@@ -462,6 +418,14 @@ with gr.Blocks() as app_tts:
             step=0.1,
             info="Adjust the speed of the audio.",
         )
+        cross_fade_duration_slider = gr.Slider(
+            label="Cross-Fade Duration (s)",
+            minimum=0.0,
+            maximum=1.0,
+            value=0.15,
+            step=0.01,
+            info="Set the duration of the cross-fade between audio clips.",
+        )
     speed_slider.change(update_speed, inputs=speed_slider)
 
     audio_output = gr.Audio(label="Synthesized Audio")
@@ -475,7 +439,7 @@ with gr.Blocks() as app_tts:
             gen_text_input,
             model_choice,
             remove_silence,
-            split_words_input,
+            cross_fade_duration_slider,
         ],
         outputs=[audio_output, spectrogram_output],
     )
@@ -568,8 +532,8 @@ with gr.Blocks() as app_emotional:
         regular_audio = gr.Audio(label='Regular Reference Audio', type='filepath')
         regular_ref_text = gr.Textbox(label='Reference Text (Regular)', lines=2)
 
-    # Additional speech types (up to 9 more)
-    max_speech_types = 10
+    # Additional speech types (up to 99 more)
+    max_speech_types = 100
     speech_type_names = []
     speech_type_audios = []
     speech_type_ref_texts = []
@@ -681,8 +645,7 @@ with gr.Blocks() as app_emotional:
 
     # Output audio
     audio_output_emotional = gr.Audio(label="Synthesized Audio")
-    
-    @spaces.GPU
+    @gpu_decorator
     def generate_emotional_speech(
         regular_audio,
         regular_ref_text,
@@ -724,7 +687,7 @@ with gr.Blocks() as app_emotional:
             ref_text = speech_types[current_emotion].get('ref_text', '')
 
             # Generate speech for this segment
-            audio, _ = infer(ref_audio, ref_text, text, model_choice, remove_silence, "")
+            audio, _ = infer(ref_audio, ref_text, text, model_choice, remove_silence, 0)
             sr, audio_data = audio
 
             generated_audio_segments.append(audio_data)
@@ -805,4 +768,27 @@ If you're having issues, try converting your reference audio to WAV or MP3, clip
     )
     gr.TabbedInterface([app_tts, app_podcast, app_emotional, app_credits], ["TTS", "Podcast", "Multi-Style", "Credits"])
 
-app.queue().launch()
+@click.command()
+@click.option("--port", "-p", default=None, type=int, help="Port to run the app on")
+@click.option("--host", "-H", default=None, help="Host to run the app on")
+@click.option(
+    "--share",
+    "-s",
+    default=False,
+    is_flag=True,
+    help="Share the app via Gradio share link",
+)
+@click.option("--api", "-a", default=True, is_flag=True, help="Allow API access")
+def main(port, host, share, api):
+    global app
+    print(f"Starting app...")
+    app.queue(api_open=api).launch(
+        server_name=host, server_port=port, share=share, show_api=api
+    )
+
+
+if __name__ == "__main__":
+    if not USING_SPACES:
+        main()
+    else:
+        app.queue().launch()

finetune-cli.py

@@ -0,0 +1,108 @@
+import argparse
+from model import CFM, UNetT, DiT, MMDiT, Trainer
+from model.utils import get_tokenizer
+from model.dataset import load_dataset
+from cached_path import cached_path
+import shutil,os
+# -------------------------- Dataset Settings --------------------------- #
+target_sample_rate = 24000
+n_mel_channels = 100
+hop_length = 256
+
+tokenizer = "pinyin"  # 'pinyin', 'char', or 'custom'
+tokenizer_path = None  # if tokenizer = 'custom', define the path to the tokenizer you want to use (should be vocab.txt)
+
+# -------------------------- Argument Parsing --------------------------- #
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train CFM Model')
+
+    parser.add_argument('--exp_name', type=str, default="F5TTS_Base", choices=["F5TTS_Base", "E2TTS_Base"],help='Experiment name')
+    parser.add_argument('--dataset_name', type=str, default="Emilia_ZH_EN", help='Name of the dataset to use')
+    parser.add_argument('--learning_rate', type=float, default=1e-4, help='Learning rate for training')
+    parser.add_argument('--batch_size_per_gpu', type=int, default=256, help='Batch size per GPU')
+    parser.add_argument('--batch_size_type', type=str, default="frame", choices=["frame", "sample"],help='Batch size type')
+    parser.add_argument('--max_samples', type=int, default=16, help='Max sequences per batch')
+    parser.add_argument('--grad_accumulation_steps', type=int, default=1,help='Gradient accumulation steps')
+    parser.add_argument('--max_grad_norm', type=float, default=1.0, help='Max gradient norm for clipping')
+    parser.add_argument('--epochs', type=int, default=10, help='Number of training epochs')
+    parser.add_argument('--num_warmup_updates', type=int, default=5, help='Warmup steps')
+    parser.add_argument('--save_per_updates', type=int, default=10, help='Save checkpoint every X steps')
+    parser.add_argument('--last_per_steps', type=int, default=10, help='Save last checkpoint every X steps')
+    parser.add_argument('--finetune', type=bool, default=True, help='Use Finetune')
+    
+    return parser.parse_args()
+
+# -------------------------- Training Settings -------------------------- #
+
+def main():
+    args = parse_args()
+    
+
+    # Model parameters based on experiment name
+    if args.exp_name == "F5TTS_Base":
+        wandb_resume_id = None
+        model_cls = DiT
+        model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
+        if args.finetune:
+           ckpt_path = str(cached_path(f"hf://SWivid/F5-TTS/F5TTS_Base/model_1200000.pt"))
+    elif args.exp_name == "E2TTS_Base":
+        wandb_resume_id = None
+        model_cls = UNetT
+        model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
+        if args.finetune:
+           ckpt_path = str(cached_path(f"hf://SWivid/E2-TTS/E2TTS_Base/model_1200000.pt"))
+           
+    if args.finetune:
+       path_ckpt = os.path.join("ckpts",args.dataset_name)
+       if os.path.isdir(path_ckpt)==False:
+          os.makedirs(path_ckpt,exist_ok=True)
+          shutil.copy2(ckpt_path,os.path.join(path_ckpt,os.path.basename(ckpt_path)))
+    
+    checkpoint_path=os.path.join("ckpts",args.dataset_name)
+
+    # Use the dataset_name provided in the command line
+    tokenizer_path = args.dataset_name if tokenizer != "custom" else tokenizer_path
+    vocab_char_map, vocab_size = get_tokenizer(tokenizer_path, tokenizer)
+
+    mel_spec_kwargs = dict(
+        target_sample_rate=target_sample_rate,
+        n_mel_channels=n_mel_channels,
+        hop_length=hop_length,
+    )
+
+    e2tts = CFM(
+        transformer=model_cls(
+            **model_cfg,
+            text_num_embeds=vocab_size,
+            mel_dim=n_mel_channels
+        ),
+        mel_spec_kwargs=mel_spec_kwargs,
+        vocab_char_map=vocab_char_map,
+    )
+
+    trainer = Trainer(
+        e2tts,
+        args.epochs,
+        args.learning_rate,
+        num_warmup_updates=args.num_warmup_updates,
+        save_per_updates=args.save_per_updates,
+        checkpoint_path=checkpoint_path,
+        batch_size=args.batch_size_per_gpu,
+        batch_size_type=args.batch_size_type,
+        max_samples=args.max_samples,
+        grad_accumulation_steps=args.grad_accumulation_steps,
+        max_grad_norm=args.max_grad_norm,
+        wandb_project="CFM-TTS",
+        wandb_run_name=args.exp_name,
+        wandb_resume_id=wandb_resume_id,
+        last_per_steps=args.last_per_steps,
+    )
+
+    train_dataset = load_dataset(args.dataset_name, tokenizer, mel_spec_kwargs=mel_spec_kwargs)
+    trainer.train(train_dataset,
+                  resumable_with_seed=666  # seed for shuffling dataset
+                  )
+
+
+if __name__ == '__main__':
+    main()

finetune_gradio.py

@@ -0,0 +1,730 @@
+import os,sys
+
+from transformers import pipeline
+import gradio as gr
+import torch
+import click
+import torchaudio
+from glob import glob
+import librosa
+import numpy as np
+from scipy.io import wavfile
+import shutil
+import time
+
+import json
+from model.utils import convert_char_to_pinyin
+import signal
+import psutil
+import platform
+import subprocess
+from datasets.arrow_writer import ArrowWriter
+
+import json
+
+training_process = None    
+system = platform.system()
+python_executable = sys.executable or "python"
+
+path_data="data"
+
+device = (
+"cuda"
+    if torch.cuda.is_available()
+    else "mps" if torch.backends.mps.is_available() else "cpu"
+)
+
+pipe = None
+
+# Load metadata
+def get_audio_duration(audio_path):
+    """Calculate the duration of an audio file."""
+    audio, sample_rate = torchaudio.load(audio_path)
+    num_channels = audio.shape[0]  
+    return audio.shape[1] / (sample_rate * num_channels)
+
+def clear_text(text):
+    """Clean and prepare text by lowering the case and stripping whitespace."""
+    return text.lower().strip()
+
+def get_rms(y,frame_length=2048,hop_length=512,pad_mode="constant",): # https://github.com/RVC-Boss/GPT-SoVITS/blob/main/tools/slicer2.py
+    padding = (int(frame_length // 2), int(frame_length // 2))
+    y = np.pad(y, padding, mode=pad_mode)
+
+    axis = -1
+    # put our new within-frame axis at the end for now
+    out_strides = y.strides + tuple([y.strides[axis]])
+    # Reduce the shape on the framing axis
+    x_shape_trimmed = list(y.shape)
+    x_shape_trimmed[axis] -= frame_length - 1
+    out_shape = tuple(x_shape_trimmed) + tuple([frame_length])
+    xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)
+    if axis < 0:
+        target_axis = axis - 1
+    else:
+        target_axis = axis + 1
+    xw = np.moveaxis(xw, -1, target_axis)
+    # Downsample along the target axis
+    slices = [slice(None)] * xw.ndim
+    slices[axis] = slice(0, None, hop_length)
+    x = xw[tuple(slices)]
+
+    # Calculate power
+    power = np.mean(np.abs(x) ** 2, axis=-2, keepdims=True)
+
+    return np.sqrt(power)
+
+class Slicer: # https://github.com/RVC-Boss/GPT-SoVITS/blob/main/tools/slicer2.py
+    def __init__(
+        self,
+        sr: int,
+        threshold: float = -40.0,
+        min_length: int = 2000,
+        min_interval: int = 300,
+        hop_size: int = 20,
+        max_sil_kept: int = 2000,
+    ):
+        if not min_length >= min_interval >= hop_size:
+            raise ValueError(
+                "The following condition must be satisfied: min_length >= min_interval >= hop_size"
+            )
+        if not max_sil_kept >= hop_size:
+            raise ValueError(
+                "The following condition must be satisfied: max_sil_kept >= hop_size"
+            )
+        min_interval = sr * min_interval / 1000
+        self.threshold = 10 ** (threshold / 20.0)
+        self.hop_size = round(sr * hop_size / 1000)
+        self.win_size = min(round(min_interval), 4 * self.hop_size)
+        self.min_length = round(sr * min_length / 1000 / self.hop_size)
+        self.min_interval = round(min_interval / self.hop_size)
+        self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+
+    def _apply_slice(self, waveform, begin, end):
+        if len(waveform.shape) > 1:
+            return waveform[
+                :, begin * self.hop_size : min(waveform.shape[1], end * self.hop_size)
+            ]
+        else:
+            return waveform[
+                begin * self.hop_size : min(waveform.shape[0], end * self.hop_size)
+            ]
+
+    # @timeit
+    def slice(self, waveform):
+        if len(waveform.shape) > 1:
+            samples = waveform.mean(axis=0)
+        else:
+            samples = waveform
+        if samples.shape[0] <= self.min_length:
+            return [waveform]
+        rms_list = get_rms(
+            y=samples, frame_length=self.win_size, hop_length=self.hop_size
+        ).squeeze(0)
+        sil_tags = []
+        silence_start = None
+        clip_start = 0
+        for i, rms in enumerate(rms_list):
+            # Keep looping while frame is silent.
+            if rms < self.threshold:
+                # Record start of silent frames.
+                if silence_start is None:
+                    silence_start = i
+                continue
+            # Keep looping while frame is not silent and silence start has not been recorded.
+            if silence_start is None:
+                continue
+            # Clear recorded silence start if interval is not enough or clip is too short
+            is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+            need_slice_middle = (
+                i - silence_start >= self.min_interval
+                and i - clip_start >= self.min_length
+            )
+            if not is_leading_silence and not need_slice_middle:
+                silence_start = None
+                continue
+            # Need slicing. Record the range of silent frames to be removed.
+            if i - silence_start <= self.max_sil_kept:
+                pos = rms_list[silence_start : i + 1].argmin() + silence_start
+                if silence_start == 0:
+                    sil_tags.append((0, pos))
+                else:
+                    sil_tags.append((pos, pos))
+                clip_start = pos
+            elif i - silence_start <= self.max_sil_kept * 2:
+                pos = rms_list[
+                    i - self.max_sil_kept : silence_start + self.max_sil_kept + 1
+                ].argmin()
+                pos += i - self.max_sil_kept
+                pos_l = (
+                    rms_list[
+                        silence_start : silence_start + self.max_sil_kept + 1
+                    ].argmin()
+                    + silence_start
+                )
+                pos_r = (
+                    rms_list[i - self.max_sil_kept : i + 1].argmin()
+                    + i
+                    - self.max_sil_kept
+                )
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                    clip_start = pos_r
+                else:
+                    sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+                    clip_start = max(pos_r, pos)
+            else:
+                pos_l = (
+                    rms_list[
+                        silence_start : silence_start + self.max_sil_kept + 1
+                    ].argmin()
+                    + silence_start
+                )
+                pos_r = (
+                    rms_list[i - self.max_sil_kept : i + 1].argmin()
+                    + i
+                    - self.max_sil_kept
+                )
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                else:
+                    sil_tags.append((pos_l, pos_r))
+                clip_start = pos_r
+            silence_start = None
+        # Deal with trailing silence.
+        total_frames = rms_list.shape[0]
+        if (
+            silence_start is not None
+            and total_frames - silence_start >= self.min_interval
+        ):
+            silence_end = min(total_frames, silence_start + self.max_sil_kept)
+            pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start
+            sil_tags.append((pos, total_frames + 1))
+        # Apply and return slices.
+        ####音频+起始时间+终止时间
+        if len(sil_tags) == 0:
+            return [[waveform,0,int(total_frames*self.hop_size)]]
+        else:
+            chunks = []
+            if sil_tags[0][0] > 0:
+                chunks.append([self._apply_slice(waveform, 0, sil_tags[0][0]),0,int(sil_tags[0][0]*self.hop_size)])
+            for i in range(len(sil_tags) - 1):
+                chunks.append(
+                    [self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0]),int(sil_tags[i][1]*self.hop_size),int(sil_tags[i + 1][0]*self.hop_size)]
+                )
+            if sil_tags[-1][1] < total_frames:
+                chunks.append(
+                    [self._apply_slice(waveform, sil_tags[-1][1], total_frames),int(sil_tags[-1][1]*self.hop_size),int(total_frames*self.hop_size)]
+                )
+            return chunks
+
+#terminal
+def terminate_process_tree(pid, including_parent=True):  
+    try:
+        parent = psutil.Process(pid)
+    except psutil.NoSuchProcess:
+        # Process already terminated
+        return
+
+    children = parent.children(recursive=True)
+    for child in children:
+        try:
+            os.kill(child.pid, signal.SIGTERM)  # or signal.SIGKILL
+        except OSError:
+            pass
+    if including_parent:
+        try:
+            os.kill(parent.pid, signal.SIGTERM)  # or signal.SIGKILL
+        except OSError:
+            pass
+
+def terminate_process(pid):
+    if system == "Windows":
+        cmd = f"taskkill /t /f /pid {pid}"
+        os.system(cmd)
+    else:
+        terminate_process_tree(pid)
+
+def start_training(dataset_name="",     
+    exp_name="F5TTS_Base",            
+    learning_rate=1e-4,               
+    batch_size_per_gpu=400,           
+    batch_size_type="frame",          
+    max_samples=64,                   
+    grad_accumulation_steps=1,        
+    max_grad_norm=1.0,                
+    epochs=11,                        
+    num_warmup_updates=200,           
+    save_per_updates=400,             
+    last_per_steps=800,               
+    finetune=True,                    
+    ):
+
+
+    global training_process
+
+    path_project = os.path.join(path_data, dataset_name + "_pinyin")
+
+    if os.path.isdir(path_project)==False:
+        yield f"There is not project with name {dataset_name}",gr.update(interactive=True),gr.update(interactive=False)
+        return
+
+    file_raw = os.path.join(path_project,"raw.arrow")
+    if os.path.isfile(file_raw)==False:
+       yield f"There is no file {file_raw}",gr.update(interactive=True),gr.update(interactive=False)
+       return    
+
+    # Check if a training process is already running
+    if training_process is not None:
+        return "Train run already!",gr.update(interactive=False),gr.update(interactive=True)
+
+    yield "start train",gr.update(interactive=False),gr.update(interactive=False)
+
+    # Command to run the training script with the specified arguments
+    cmd = f"accelerate launch finetune-cli.py --exp_name {exp_name} " \
+          f"--learning_rate {learning_rate} " \
+          f"--batch_size_per_gpu {batch_size_per_gpu} " \
+          f"--batch_size_type {batch_size_type} " \
+          f"--max_samples {max_samples} " \
+          f"--grad_accumulation_steps {grad_accumulation_steps} " \
+          f"--max_grad_norm {max_grad_norm} " \
+          f"--epochs {epochs} " \
+          f"--num_warmup_updates {num_warmup_updates} " \
+          f"--save_per_updates {save_per_updates} " \
+          f"--last_per_steps {last_per_steps} " \
+          f"--dataset_name {dataset_name}"  
+    if finetune:cmd += f" --finetune {finetune}"   
+    print(cmd)      
+    try:
+      # Start the training process
+      training_process = subprocess.Popen(cmd, shell=True)
+
+      time.sleep(5) 
+      yield "check terminal for wandb",gr.update(interactive=False),gr.update(interactive=True) 
+      
+      # Wait for the training process to finish
+      training_process.wait()
+      time.sleep(1)
+      
+      if training_process is None: 
+         text_info = 'train stop'
+      else:     
+         text_info = "train complete !"
+
+    except Exception as e:  # Catch all exceptions
+        # Ensure that we reset the training process variable in case of an error
+        text_info=f"An error occurred: {str(e)}"
+    
+    training_process=None
+
+    yield text_info,gr.update(interactive=True),gr.update(interactive=False)
+
+def stop_training():
+    global training_process
+    if training_process is None:return f"Train not run !",gr.update(interactive=True),gr.update(interactive=False)
+    terminate_process_tree(training_process.pid)
+    training_process = None
+    return 'train stop',gr.update(interactive=True),gr.update(interactive=False)
+
+def create_data_project(name):
+    name+="_pinyin"
+    os.makedirs(os.path.join(path_data,name),exist_ok=True)
+    os.makedirs(os.path.join(path_data,name,"dataset"),exist_ok=True)
+    
+def transcribe(file_audio,language="english"):
+    global pipe
+
+    if pipe is None:
+       pipe = pipeline("automatic-speech-recognition",model="openai/whisper-large-v3-turbo", torch_dtype=torch.float16,device=device)
+
+    text_transcribe = pipe(
+        file_audio,
+        chunk_length_s=30,
+        batch_size=128,
+        generate_kwargs={"task": "transcribe","language": language},
+        return_timestamps=False,
+    )["text"].strip()
+    return text_transcribe
+
+def transcribe_all(name_project,audio_files,language,user=False,progress=gr.Progress()):
+    name_project+="_pinyin"
+    path_project= os.path.join(path_data,name_project)
+    path_dataset = os.path.join(path_project,"dataset")
+    path_project_wavs = os.path.join(path_project,"wavs")
+    file_metadata = os.path.join(path_project,"metadata.csv")
+
+    if audio_files is None:return "You need to load an audio file."
+
+    if os.path.isdir(path_project_wavs):
+       shutil.rmtree(path_project_wavs)
+
+    if os.path.isfile(file_metadata):
+       os.remove(file_metadata)
+
+    os.makedirs(path_project_wavs,exist_ok=True)
+    
+    if user:
+       file_audios = [file for format in ('*.wav', '*.ogg', '*.opus', '*.mp3', '*.flac') for file in glob(os.path.join(path_dataset, format))]
+       if file_audios==[]:return "No audio file was found in the dataset."
+    else:
+       file_audios = audio_files
+       
+
+    alpha = 0.5
+    _max = 1.0
+    slicer = Slicer(24000)
+
+    num = 0
+    error_num = 0
+    data=""
+    for file_audio in progress.tqdm(file_audios, desc="transcribe files",total=len((file_audios))):
+        
+        audio, _ = librosa.load(file_audio, sr=24000, mono=True) 
+
+        list_slicer=slicer.slice(audio)
+        for chunk, start, end in progress.tqdm(list_slicer,total=len(list_slicer), desc="slicer files"): 
+            
+            name_segment = os.path.join(f"segment_{num}")
+            file_segment = os.path.join(path_project_wavs, f"{name_segment}.wav")    
+            
+            tmp_max = np.abs(chunk).max()
+            if(tmp_max>1):chunk/=tmp_max
+            chunk = (chunk / tmp_max * (_max * alpha)) + (1 - alpha) * chunk
+            wavfile.write(file_segment,24000, (chunk * 32767).astype(np.int16))
+            
+            try:
+               text=transcribe(file_segment,language)
+               text = text.lower().strip().replace('"',"")
+
+               data+= f"{name_segment}|{text}\n"
+
+               num+=1
+            except:
+               error_num +=1
+
+    with open(file_metadata,"w",encoding="utf-8") as f:
+        f.write(data)
+    
+    if error_num!=[]:
+       error_text=f"\nerror files : {error_num}"    
+    else:
+       error_text=""
+       
+    return f"transcribe complete samples : {num}\npath : {path_project_wavs}{error_text}"
+
+def format_seconds_to_hms(seconds):
+    hours = int(seconds / 3600)
+    minutes = int((seconds % 3600) / 60)
+    seconds = seconds % 60
+    return "{:02d}:{:02d}:{:02d}".format(hours, minutes, int(seconds))
+
+def create_metadata(name_project,progress=gr.Progress()):
+    name_project+="_pinyin"
+    path_project= os.path.join(path_data,name_project)
+    path_project_wavs = os.path.join(path_project,"wavs")
+    file_metadata = os.path.join(path_project,"metadata.csv")
+    file_raw = os.path.join(path_project,"raw.arrow")
+    file_duration = os.path.join(path_project,"duration.json")
+    file_vocab = os.path.join(path_project,"vocab.txt")
+
+    if os.path.isfile(file_metadata)==False: return "The file was not found in " + file_metadata 
+    
+    with open(file_metadata,"r",encoding="utf-8") as f:
+        data=f.read()
+    
+    audio_path_list=[]
+    text_list=[]
+    duration_list=[]
+    
+    count=data.split("\n")
+    lenght=0
+    result=[]
+    error_files=[]
+    for line in progress.tqdm(data.split("\n"),total=count):
+        sp_line=line.split("|")
+        if len(sp_line)!=2:continue
+        name_audio,text = sp_line[:2]
+
+        file_audio = os.path.join(path_project_wavs, name_audio + ".wav")
+
+        if os.path.isfile(file_audio)==False:
+            error_files.append(file_audio)
+            continue
+
+        duraction = get_audio_duration(file_audio)
+        if duraction<2 and duraction>15:continue
+        if len(text)<4:continue
+
+        text = clear_text(text)
+        text = convert_char_to_pinyin([text], polyphone = True)[0]
+
+        audio_path_list.append(file_audio)
+        duration_list.append(duraction)
+        text_list.append(text)
+      
+        result.append({"audio_path": file_audio, "text": text, "duration": duraction})
+
+        lenght+=duraction
+
+    if duration_list==[]:
+        error_files_text="\n".join(error_files)
+        return f"Error: No audio files found in the specified path : \n{error_files_text}"
+    
+    min_second = round(min(duration_list),2)   
+    max_second = round(max(duration_list),2)
+
+    with ArrowWriter(path=file_raw, writer_batch_size=1) as writer:
+        for line in progress.tqdm(result,total=len(result), desc=f"prepare data"):
+            writer.write(line)
+
+    with open(file_duration, 'w', encoding='utf-8') as f:
+        json.dump({"duration": duration_list}, f, ensure_ascii=False)
+ 
+    file_vocab_finetune = "data/Emilia_ZH_EN_pinyin/vocab.txt"    
+    if os.path.isfile(file_vocab_finetune==False):return "Error: Vocabulary file 'Emilia_ZH_EN_pinyin' not found!"
+    shutil.copy2(file_vocab_finetune, file_vocab)
+    
+    if error_files!=[]:
+       error_text="error files\n" + "\n".join(error_files)
+    else:
+       error_text=""
+        
+    return f"prepare complete \nsamples : {len(text_list)}\ntime data : {format_seconds_to_hms(lenght)}\nmin sec : {min_second}\nmax sec : {max_second}\nfile_arrow : {file_raw}\n{error_text}"
+
+def check_user(value):
+    return gr.update(visible=not value),gr.update(visible=value)
+
+def calculate_train(name_project,batch_size_type,max_samples,learning_rate,num_warmup_updates,save_per_updates,last_per_steps,finetune):
+    name_project+="_pinyin"
+    path_project= os.path.join(path_data,name_project)
+    file_duraction = os.path.join(path_project,"duration.json")
+
+    with open(file_duraction, 'r') as file:
+        data = json.load(file)  
+    
+    duration_list = data['duration']
+
+    samples = len(duration_list)
+
+    if torch.cuda.is_available():
+        gpu_properties = torch.cuda.get_device_properties(0)
+        total_memory = gpu_properties.total_memory / (1024 ** 3)  
+    elif torch.backends.mps.is_available():
+        total_memory = psutil.virtual_memory().available / (1024 ** 3)
+    
+    if batch_size_type=="frame":
+       batch = int(total_memory * 0.5)
+       batch = (lambda num: num + 1 if num % 2 != 0 else num)(batch)
+       batch_size_per_gpu = int(38400 / batch )
+    else:
+       batch_size_per_gpu = int(total_memory / 8)   
+       batch_size_per_gpu = (lambda num: num + 1 if num % 2 != 0 else num)(batch_size_per_gpu)
+       batch = batch_size_per_gpu
+
+    if batch_size_per_gpu<=0:batch_size_per_gpu=1
+
+    if samples<64:
+       max_samples = int(samples * 0.25) 
+    
+    num_warmup_updates = int(samples * 0.10) 
+    save_per_updates = int(samples * 0.25) 
+    last_per_steps =int(save_per_updates * 5)
+    
+    max_samples = (lambda num: num + 1 if num % 2 != 0 else num)(max_samples)
+    num_warmup_updates = (lambda num: num + 1 if num % 2 != 0 else num)(num_warmup_updates)
+    save_per_updates = (lambda num: num + 1 if num % 2 != 0 else num)(save_per_updates)
+    last_per_steps = (lambda num: num + 1 if num % 2 != 0 else num)(last_per_steps)
+
+    if finetune:learning_rate=1e-4
+    else:learning_rate=7.5e-5
+
+    return batch_size_per_gpu,max_samples,num_warmup_updates,save_per_updates,last_per_steps,samples,learning_rate
+
+def extract_and_save_ema_model(checkpoint_path: str, new_checkpoint_path: str) -> None:
+    try:
+        checkpoint = torch.load(checkpoint_path)
+        print("Original Checkpoint Keys:", checkpoint.keys())
+        
+        ema_model_state_dict = checkpoint.get('ema_model_state_dict', None)
+
+        if ema_model_state_dict is not None:
+            new_checkpoint = {'ema_model_state_dict': ema_model_state_dict}
+            torch.save(new_checkpoint, new_checkpoint_path)
+            return f"New checkpoint saved at: {new_checkpoint_path}"
+        else:
+            return "No 'ema_model_state_dict' found in the checkpoint."
+
+    except Exception as e:
+        return f"An error occurred: {e}"
+
+def vocab_check(project_name):
+    name_project = project_name + "_pinyin"
+    path_project = os.path.join(path_data, name_project)
+
+    file_metadata = os.path.join(path_project, "metadata.csv")
+   
+    file_vocab="data/Emilia_ZH_EN_pinyin/vocab.txt"
+    if os.path.isfile(file_vocab)==False:
+        return f"the file {file_vocab} not found !"
+    
+    with open(file_vocab,"r",encoding="utf-8") as f:
+         data=f.read()
+
+    vocab = data.split("\n")
+
+    if os.path.isfile(file_metadata)==False:
+        return f"the file {file_metadata} not found !"
+
+    with open(file_metadata,"r",encoding="utf-8") as f:
+         data=f.read()
+
+    miss_symbols=[]
+    miss_symbols_keep={}
+    for item in data.split("\n"):
+         sp=item.split("|")
+         if len(sp)!=2:continue
+         text=sp[1].lower().strip()
+
+         for t in text:
+             if (t in vocab)==False and (t in miss_symbols_keep)==False:
+                miss_symbols.append(t) 
+                miss_symbols_keep[t]=t
+
+            
+    if miss_symbols==[]:info ="You can train using your language !"     
+    else:info = f"The following symbols are missing in your language : {len(miss_symbols)}\n\n" + "\n".join(miss_symbols)   
+
+    return info
+
+
+
+with gr.Blocks() as app:
+
+    with gr.Row():
+         project_name=gr.Textbox(label="project name",value="my_speak")
+         bt_create=gr.Button("create new project")
+    
+    bt_create.click(fn=create_data_project,inputs=[project_name])
+
+    with gr.Tabs():
+         
+
+         with gr.TabItem("transcribe Data"):
+
+
+              ch_manual = gr.Checkbox(label="user",value=False)
+
+              mark_info_transcribe=gr.Markdown(
+                         """```plaintext    
+     Place your 'wavs' folder and 'metadata.csv' file in the {your_project_name}' directory. 
+                 
+     my_speak/
+     │
+     └── dataset/
+         ├── audio1.wav
+         └── audio2.wav
+         ...
+     ```""",visible=False)
+
+              audio_speaker = gr.File(label="voice",type="filepath",file_count="multiple")
+              txt_lang = gr.Text(label="Language",value="english")
+              bt_transcribe=bt_create=gr.Button("transcribe")
+              txt_info_transcribe=gr.Text(label="info",value="")
+              bt_transcribe.click(fn=transcribe_all,inputs=[project_name,audio_speaker,txt_lang,ch_manual],outputs=[txt_info_transcribe])
+              ch_manual.change(fn=check_user,inputs=[ch_manual],outputs=[audio_speaker,mark_info_transcribe])
+         
+         with gr.TabItem("prepare Data"):
+              gr.Markdown(
+                         """```plaintext    
+     place all your wavs folder and your metadata.csv file in {your name project}                                 
+     my_speak/
+     │
+     ├── wavs/
+     │   ├── audio1.wav
+     │   └── audio2.wav
+     |   ...
+     │
+     └── metadata.csv
+      
+     file format metadata.csv
+
+     audio1|text1
+     audio2|text1
+     ...
+
+     ```""")
+
+              bt_prepare=bt_create=gr.Button("prepare")
+              txt_info_prepare=gr.Text(label="info",value="")
+              bt_prepare.click(fn=create_metadata,inputs=[project_name],outputs=[txt_info_prepare])
+         
+         with gr.TabItem("train Data"):
+              
+              with gr.Row():    
+                   bt_calculate=bt_create=gr.Button("Auto Settings") 
+                   ch_finetune=bt_create=gr.Checkbox(label="finetune",value=True)   
+                   lb_samples = gr.Label(label="samples")  
+                   batch_size_type = gr.Radio(label="Batch Size Type", choices=["frame", "sample"], value="frame")  
+                
+              with gr.Row():     
+                   exp_name = gr.Radio(label="Model", choices=["F5TTS_Base", "E2TTS_Base"], value="F5TTS_Base")
+                   learning_rate = gr.Number(label="Learning Rate", value=1e-4, step=1e-4)
+                   
+              with gr.Row():
+                   batch_size_per_gpu = gr.Number(label="Batch Size per GPU", value=1000)
+                   max_samples = gr.Number(label="Max Samples", value=16)
+                   
+              with gr.Row():
+                   grad_accumulation_steps = gr.Number(label="Gradient Accumulation Steps", value=1)
+                   max_grad_norm = gr.Number(label="Max Gradient Norm", value=1.0)
+                   
+              with gr.Row():
+                   epochs = gr.Number(label="Epochs", value=10)
+                   num_warmup_updates = gr.Number(label="Warmup Updates", value=5)
+
+              with gr.Row():     
+                   save_per_updates = gr.Number(label="Save per Updates", value=10)
+                   last_per_steps = gr.Number(label="Last per Steps", value=50)
+       
+              with gr.Row():
+                   start_button = gr.Button("Start Training")
+                   stop_button = gr.Button("Stop Training",interactive=False)
+         
+              txt_info_train=gr.Text(label="info",value="")
+              start_button.click(fn=start_training,inputs=[project_name,exp_name,learning_rate,batch_size_per_gpu,batch_size_type,max_samples,grad_accumulation_steps,max_grad_norm,epochs,num_warmup_updates,save_per_updates,last_per_steps,ch_finetune],outputs=[txt_info_train,start_button,stop_button])
+              stop_button.click(fn=stop_training,outputs=[txt_info_train,start_button,stop_button])
+              bt_calculate.click(fn=calculate_train,inputs=[project_name,batch_size_type,max_samples,learning_rate,num_warmup_updates,save_per_updates,last_per_steps,ch_finetune],outputs=[batch_size_per_gpu,max_samples,num_warmup_updates,save_per_updates,last_per_steps,lb_samples,learning_rate]) 
+         
+         with gr.TabItem("reduse checkpoint"):
+              txt_path_checkpoint = gr.Text(label="path checkpoint :") 
+              txt_path_checkpoint_small = gr.Text(label="path output :") 
+              txt_info_reduse = gr.Text(label="info",value="")
+              reduse_button = gr.Button("reduse") 
+              reduse_button.click(fn=extract_and_save_ema_model,inputs=[txt_path_checkpoint,txt_path_checkpoint_small],outputs=[txt_info_reduse])
+
+         with gr.TabItem("vocab check experiment"):
+              check_button = gr.Button("check vocab") 
+              txt_info_check=gr.Text(label="info",value="")
+              check_button.click(fn=vocab_check,inputs=[project_name],outputs=[txt_info_check])
+       
+
+@click.command()
+@click.option("--port", "-p", default=None, type=int, help="Port to run the app on")
+@click.option("--host", "-H", default=None, help="Host to run the app on")
+@click.option(
+    "--share",
+    "-s",
+    default=False,
+    is_flag=True,
+    help="Share the app via Gradio share link",
+)
+@click.option("--api", "-a", default=True, is_flag=True, help="Allow API access")
+def main(port, host, share, api):
+    global app
+    print(f"Starting app...")
+    app.queue(api_open=api).launch(
+        server_name=host, server_port=port, share=share, show_api=api
+    )
+
+if __name__ == "__main__":
+    main()

inference-cli.py

@@ -1,25 +1,24 @@
+import argparse
+import codecs
 import re
+import tempfile
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+import tomli
 import torch
 import torchaudio
-import numpy as np
-import tempfile
+import tqdm
+from cached_path import cached_path
 from einops import rearrange
-from vocos import Vocos
 from pydub import AudioSegment, silence
-from model import CFM, UNetT, DiT, MMDiT
-from cached_path import cached_path
-from model.utils import (
-    load_checkpoint,
-    get_tokenizer,
-    convert_char_to_pinyin,
-    save_spectrogram,
-)
 from transformers import pipeline
-import soundfile as sf
-import tomli
-import argparse
-import tqdm
-from pathlib import Path
+from vocos import Vocos
+
+from model import CFM, DiT, MMDiT, UNetT
+from model.utils import (convert_char_to_pinyin, get_tokenizer,
+                         load_checkpoint, save_spectrogram)
 
 parser = argparse.ArgumentParser(
     prog="python3 inference-cli.py",
@@ -56,6 +55,12 @@ parser.add_argument(
     type=str,
     help="Text to generate.",
 )
+parser.add_argument(
+    "-f",
+    "--gen_file",
+    type=str,
+    help="File with text to generate. Ignores --text",
+)
 parser.add_argument(
     "-o",
     "--output_dir",
@@ -66,6 +71,11 @@ parser.add_argument(
     "--remove_silence",
     help="Remove silence.",
 )
+parser.add_argument(
+    "--load_vocoder_from_local",
+    action="store_true",
+    help="load vocoder from local. Default: ../checkpoints/charactr/vocos-mel-24khz",
+)
 args = parser.parse_args()
 
 config = tomli.load(open(args.config, "rb"))
@@ -73,29 +83,31 @@ config = tomli.load(open(args.config, "rb"))
 ref_audio = args.ref_audio if args.ref_audio else config["ref_audio"]
 ref_text = args.ref_text if args.ref_text != "666" else config["ref_text"]
 gen_text = args.gen_text if args.gen_text else config["gen_text"]
+gen_file = args.gen_file if args.gen_file else config["gen_file"]
+if gen_file:
+    gen_text = codecs.open(gen_file, "r", "utf-8").read()
 output_dir = args.output_dir if args.output_dir else config["output_dir"]
 model = args.model if args.model else config["model"]
 remove_silence = args.remove_silence if args.remove_silence else config["remove_silence"]
 wave_path = Path(output_dir)/"out.wav"
 spectrogram_path = Path(output_dir)/"out.png"
-
-SPLIT_WORDS = [
-    "but", "however", "nevertheless", "yet", "still",
-    "therefore", "thus", "hence", "consequently",
-    "moreover", "furthermore", "additionally",
-    "meanwhile", "alternatively", "otherwise",
-    "namely", "specifically", "for example", "such as",
-    "in fact", "indeed", "notably",
-    "in contrast", "on the other hand", "conversely",
-    "in conclusion", "to summarize", "finally"
-]
+vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
 
 device = (
     "cuda"
     if torch.cuda.is_available()
     else "mps" if torch.backends.mps.is_available() else "cpu"
 )
-vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
+
+if args.load_vocoder_from_local:
+    print(f"Load vocos from local path {vocos_local_path}")
+    vocos = Vocos.from_hparams(f"{vocos_local_path}/config.yaml")
+    state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", map_location=device)
+    vocos.load_state_dict(state_dict)
+    vocos.eval()
+else:
+    print("Donwload Vocos from huggingface charactr/vocos-mel-24khz")
+    vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
 
 print(f"Using {device} device")
 
@@ -114,8 +126,9 @@ speed = 1.0
 fix_duration = None
 
 def load_model(repo_name, exp_name, model_cls, model_cfg, ckpt_step):
-    ckpt_path = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.safetensors"))
-    # ckpt_path = f"ckpts/{exp_name}/model_{ckpt_step}.pt"  # .pt | .safetensors
+    ckpt_path = f"ckpts/{exp_name}/model_{ckpt_step}.pt" # .pt | .safetensors
+    if not Path(ckpt_path).exists():
+        ckpt_path = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.safetensors"))
     vocab_char_map, vocab_size = get_tokenizer("Emilia_ZH_EN", "pinyin")
     model = CFM(
         transformer=model_cls(
@@ -143,103 +156,36 @@ F5TTS_model_cfg = dict(
 )
 E2TTS_model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
 
-def split_text_into_batches(text, max_chars=200, split_words=SPLIT_WORDS):
-    if len(text.encode('utf-8')) <= max_chars:
-        return [text]
-    if text[-1] not in ['。', '.', '!', '！', '?', '？']:
-        text += '.'
-        
-    sentences = re.split('([。.!?！？])', text)
-    sentences = [''.join(i) for i in zip(sentences[0::2], sentences[1::2])]
-    
-    batches = []
-    current_batch = ""
-    
-    def split_by_words(text):
-        words = text.split()
-        current_word_part = ""
-        word_batches = []
-        for word in words:
-            if len(current_word_part.encode('utf-8')) + len(word.encode('utf-8')) + 1 <= max_chars:
-                current_word_part += word + ' '
-            else:
-                if current_word_part:
-                    # Try to find a suitable split word
-                    for split_word in split_words:
-                        split_index = current_word_part.rfind(' ' + split_word + ' ')
-                        if split_index != -1:
-                            word_batches.append(current_word_part[:split_index].strip())
-                            current_word_part = current_word_part[split_index:].strip() + ' '
-                            break
-                    else:
-                        # If no suitable split word found, just append the current part
-                        word_batches.append(current_word_part.strip())
-                        current_word_part = ""
-                current_word_part += word + ' '
-        if current_word_part:
-            word_batches.append(current_word_part.strip())
-        return word_batches
+
+def chunk_text(text, max_chars=135):
+    """
+    Splits the input text into chunks, each with a maximum number of characters.
+    Args:
+        text (str): The text to be split.
+        max_chars (int): The maximum number of characters per chunk.
+    Returns:
+        List[str]: A list of text chunks.
+    """
+    chunks = []
+    current_chunk = ""
+    # Split the text into sentences based on punctuation followed by whitespace
+    sentences = re.split(r'(?<=[;:,.!?])\s+|(?<=[；：，。！？])', text)
 
     for sentence in sentences:
-        if len(current_batch.encode('utf-8')) + len(sentence.encode('utf-8')) <= max_chars:
-            current_batch += sentence
+        if len(current_chunk.encode('utf-8')) + len(sentence.encode('utf-8')) <= max_chars:
+            current_chunk += sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
         else:
-            # If adding this sentence would exceed the limit
-            if current_batch:
-                batches.append(current_batch)
-                current_batch = ""
-            
-            # If the sentence itself is longer than max_chars, split it
-            if len(sentence.encode('utf-8')) > max_chars:
-                # First, try to split by colon
-                colon_parts = sentence.split(':')
-                if len(colon_parts) > 1:
-                    for part in colon_parts:
-                        if len(part.encode('utf-8')) <= max_chars:
-                            batches.append(part)
-                        else:
-                            # If colon part is still too long, split by comma
-                            comma_parts = re.split('[,，]', part)
-                            if len(comma_parts) > 1:
-                                current_comma_part = ""
-                                for comma_part in comma_parts:
-                                    if len(current_comma_part.encode('utf-8')) + len(comma_part.encode('utf-8')) <= max_chars:
-                                        current_comma_part += comma_part + ','
-                                    else:
-                                        if current_comma_part:
-                                            batches.append(current_comma_part.rstrip(','))
-                                        current_comma_part = comma_part + ','
-                                if current_comma_part:
-                                    batches.append(current_comma_part.rstrip(','))
-                            else:
-                                # If no comma, split by words
-                                batches.extend(split_by_words(part))
-                else:
-                    # If no colon, split by comma
-                    comma_parts = re.split('[,，]', sentence)
-                    if len(comma_parts) > 1:
-                        current_comma_part = ""
-                        for comma_part in comma_parts:
-                            if len(current_comma_part.encode('utf-8')) + len(comma_part.encode('utf-8')) <= max_chars:
-                                current_comma_part += comma_part + ','
-                            else:
-                                if current_comma_part:
-                                    batches.append(current_comma_part.rstrip(','))
-                                current_comma_part = comma_part + ','
-                        if current_comma_part:
-                            batches.append(current_comma_part.rstrip(','))
-                    else:
-                        # If no comma, split by words
-                        batches.extend(split_by_words(sentence))
-            else:
-                current_batch = sentence
-    
-    if current_batch:
-        batches.append(current_batch)
-    
-    return batches
+            if current_chunk:
+                chunks.append(current_chunk.strip())
+            current_chunk = sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
 
-def infer_batch(ref_audio, ref_text, gen_text_batches, model, remove_silence):
+    if current_chunk:
+        chunks.append(current_chunk.strip())
+
+    return chunks
+
+
+def infer_batch(ref_audio, ref_text, gen_text_batches, model, remove_silence, cross_fade_duration=0.15):
     if model == "F5-TTS":
         ema_model = load_model(model, "F5TTS_Base", DiT, F5TTS_model_cfg, 1200000)
     elif model == "E2-TTS":
@@ -297,41 +243,56 @@ def infer_batch(ref_audio, ref_text, gen_text_batches, model, remove_silence):
         generated_waves.append(generated_wave)
         spectrograms.append(generated_mel_spec[0].cpu().numpy())
 
-    # Combine all generated waves
-    final_wave = np.concatenate(generated_waves)
-
-    with open(wave_path, "wb") as f:
-        sf.write(f.name, final_wave, target_sample_rate)
-        # Remove silence
-        if remove_silence:
-            aseg = AudioSegment.from_file(f.name)
-            non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
-            non_silent_wave = AudioSegment.silent(duration=0)
-            for non_silent_seg in non_silent_segs:
-                non_silent_wave += non_silent_seg
-            aseg = non_silent_wave
-            aseg.export(f.name, format="wav")
-        print(f.name)
+    # Combine all generated waves with cross-fading
+    if cross_fade_duration <= 0:
+        # Simply concatenate
+        final_wave = np.concatenate(generated_waves)
+    else:
+        final_wave = generated_waves[0]
+        for i in range(1, len(generated_waves)):
+            prev_wave = final_wave
+            next_wave = generated_waves[i]
 
-    # Create a combined spectrogram
-    combined_spectrogram = np.concatenate(spectrograms, axis=1)
-    save_spectrogram(combined_spectrogram, spectrogram_path)
-    print(spectrogram_path)
+            # Calculate cross-fade samples, ensuring it does not exceed wave lengths
+            cross_fade_samples = int(cross_fade_duration * target_sample_rate)
+            cross_fade_samples = min(cross_fade_samples, len(prev_wave), len(next_wave))
 
+            if cross_fade_samples <= 0:
+                # No overlap possible, concatenate
+                final_wave = np.concatenate([prev_wave, next_wave])
+                continue
 
-def infer(ref_audio_orig, ref_text, gen_text, model, remove_silence, custom_split_words):
-    if not custom_split_words.strip():
-        custom_words = [word.strip() for word in custom_split_words.split(',')]
-        global SPLIT_WORDS
-        SPLIT_WORDS = custom_words
+            # Overlapping parts
+            prev_overlap = prev_wave[-cross_fade_samples:]
+            next_overlap = next_wave[:cross_fade_samples]
 
-    print(gen_text)
+            # Fade out and fade in
+            fade_out = np.linspace(1, 0, cross_fade_samples)
+            fade_in = np.linspace(0, 1, cross_fade_samples)
+
+            # Cross-faded overlap
+            cross_faded_overlap = prev_overlap * fade_out + next_overlap * fade_in
+
+            # Combine
+            new_wave = np.concatenate([
+                prev_wave[:-cross_fade_samples],
+                cross_faded_overlap,
+                next_wave[cross_fade_samples:]
+            ])
+
+            final_wave = new_wave
+
+    # Create a combined spectrogram
+    combined_spectrogram = np.concatenate(spectrograms, axis=1)
 
+    return final_wave, combined_spectrogram
+
+def process_voice(ref_audio_orig, ref_text):
     print("Converting audio...")
     with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
         aseg = AudioSegment.from_file(ref_audio_orig)
 
-        non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
+        non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=1000)
         non_silent_wave = AudioSegment.silent(duration=0)
         for non_silent_seg in non_silent_segs:
             non_silent_wave += non_silent_seg
@@ -362,17 +323,70 @@ def infer(ref_audio_orig, ref_text, gen_text, model, remove_silence, custom_spli
         print("Finished transcription")
     else:
         print("Using custom reference text...")
+    return ref_audio, ref_text    
+
+def infer(ref_audio, ref_text, gen_text, model, remove_silence, cross_fade_duration=0.15):
+    print(gen_text)
+    # Add the functionality to ensure it ends with ". "
+    if not ref_text.endswith(". ") and not ref_text.endswith("。"):
+        if ref_text.endswith("."):
+            ref_text += " "
+        else:
+            ref_text += ". "
 
     # Split the input text into batches
     audio, sr = torchaudio.load(ref_audio)
-    max_chars = int(len(ref_text.encode('utf-8')) / (audio.shape[-1] / sr) * (30 - audio.shape[-1] / sr))
-    gen_text_batches = split_text_into_batches(gen_text, max_chars=max_chars)
+    max_chars = int(len(ref_text.encode('utf-8')) / (audio.shape[-1] / sr) * (25 - audio.shape[-1] / sr))
+    gen_text_batches = chunk_text(gen_text, max_chars=max_chars)
     print('ref_text', ref_text)
     for i, gen_text in enumerate(gen_text_batches):
         print(f'gen_text {i}', gen_text)
     
     print(f"Generating audio using {model} in {len(gen_text_batches)} batches, loading models...")
-    return infer_batch((audio, sr), ref_text, gen_text_batches, model, remove_silence)
+    return infer_batch((audio, sr), ref_text, gen_text_batches, model, remove_silence, cross_fade_duration)
     
 
-infer(ref_audio, ref_text, gen_text, model, remove_silence, ",".join(SPLIT_WORDS))
+def process(ref_audio, ref_text, text_gen, model, remove_silence):
+    main_voice = {"ref_audio":ref_audio, "ref_text":ref_text}
+    if "voices" not in config:
+        voices = {"main": main_voice}
+    else:
+        voices = config["voices"]
+        voices["main"] = main_voice
+    for voice in voices:
+        voices[voice]['ref_audio'], voices[voice]['ref_text'] = process_voice(voices[voice]['ref_audio'], voices[voice]['ref_text'])
+
+    generated_audio_segments = []
+    reg1 = r'(?=\[\w+\])'
+    chunks = re.split(reg1, text_gen)
+    reg2 = r'\[(\w+)\]'
+    for text in chunks:
+        match = re.match(reg2, text)
+        if not match or voice not in voices:
+            voice = "main"
+        else:
+            voice = match[1]
+        text = re.sub(reg2, "", text)
+        gen_text = text.strip()
+        ref_audio = voices[voice]['ref_audio']
+        ref_text = voices[voice]['ref_text']
+        print(f"Voice: {voice}")
+        audio, spectragram = infer(ref_audio, ref_text, gen_text, model, remove_silence)
+        generated_audio_segments.append(audio)
+
+    if generated_audio_segments:
+        final_wave = np.concatenate(generated_audio_segments)
+        with open(wave_path, "wb") as f:
+            sf.write(f.name, final_wave, target_sample_rate)
+            # Remove silence
+            if remove_silence:
+                aseg = AudioSegment.from_file(f.name)
+                non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
+                non_silent_wave = AudioSegment.silent(duration=0)
+                for non_silent_seg in non_silent_segs:
+                    non_silent_wave += non_silent_seg
+                aseg = non_silent_wave
+                aseg.export(f.name, format="wav")
+            print(f.name)
+
+process(ref_audio, ref_text, gen_text, model, remove_silence)

inference-cli.toml

@@ -4,5 +4,7 @@ ref_audio = "tests/ref_audio/test_en_1_ref_short.wav"
 # If an empty "", transcribes the reference audio automatically.
 ref_text = "Some call me nature, others call me mother nature."
 gen_text = "I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring. Respect me and I'll nurture you; ignore me and you shall face the consequences."
-remove_silence = true
+# File with text to generate. Ignores the text above.
+gen_file = ""
+remove_silence = false
 output_dir = "tests"

model/dataset.py

@@ -184,11 +184,15 @@ class DynamicBatchSampler(Sampler[list[int]]):
 
 def load_dataset(
         dataset_name: str,
-        tokenizer: str,
+        tokenizer: str = "pinyin",
         dataset_type: str = "CustomDataset", 
         audio_type: str = "raw", 
         mel_spec_kwargs: dict = dict()
-        ) -> CustomDataset:
+        ) -> CustomDataset | HFDataset:
+    '''
+    dataset_type    - "CustomDataset" if you want to use tokenizer name and default data path to load for train_dataset
+                    - "CustomDatasetPath" if you just want to pass the full path to a preprocessed dataset without relying on tokenizer
+    '''
     
     print("Loading dataset ...")
 
@@ -206,7 +210,18 @@ def load_dataset(
             data_dict = json.load(f)
         durations = data_dict["duration"]
         train_dataset = CustomDataset(train_dataset, durations=durations, preprocessed_mel=preprocessed_mel, **mel_spec_kwargs)
-
+        
+    elif dataset_type == "CustomDatasetPath":
+        try:
+            train_dataset = load_from_disk(f"{dataset_name}/raw")
+        except:
+            train_dataset = Dataset_.from_file(f"{dataset_name}/raw.arrow")
+            
+        with open(f"{dataset_name}/duration.json", 'r', encoding='utf-8') as f:
+            data_dict = json.load(f)
+        durations = data_dict["duration"]
+        train_dataset = CustomDataset(train_dataset, durations=durations, preprocessed_mel=preprocessed_mel, **mel_spec_kwargs)
+            
     elif dataset_type == "HFDataset":
         print("Should manually modify the path of huggingface dataset to your need.\n" +
               "May also the corresponding script cuz different dataset may have different format.")

model/trainer.py

@@ -140,7 +140,7 @@ class Trainer:
         else:
             latest_checkpoint = sorted([f for f in os.listdir(self.checkpoint_path) if f.endswith('.pt')], key=lambda x: int(''.join(filter(str.isdigit, x))))[-1]
         # checkpoint = torch.load(f"{self.checkpoint_path}/{latest_checkpoint}", map_location=self.accelerator.device)  # rather use accelerator.load_state ಥ_ಥ
-        checkpoint = torch.load(f"{self.checkpoint_path}/{latest_checkpoint}", map_location="cpu")
+        checkpoint = torch.load(f"{self.checkpoint_path}/{latest_checkpoint}", weights_only=True, map_location="cpu")
 
         if self.is_main:
             self.ema_model.load_state_dict(checkpoint['ema_model_state_dict'])

model/utils.py

@@ -22,12 +22,6 @@ from einops import rearrange, reduce
 
 import jieba
 from pypinyin import lazy_pinyin, Style
-import zhconv
-from zhon.hanzi import punctuation
-from jiwer import compute_measures
-
-from funasr import AutoModel
-from faster_whisper import WhisperModel
 
 from model.ecapa_tdnn import ECAPA_TDNN_SMALL
 from model.modules import MelSpec
@@ -129,6 +123,7 @@ def get_tokenizer(dataset_name, tokenizer: str = "pinyin"):
     tokenizer   - "pinyin" do g2p for only chinese characters, need .txt vocab_file
                 - "char" for char-wise tokenizer, need .txt vocab_file
                 - "byte" for utf-8 tokenizer
+                - "custom" if you're directly passing in a path to the vocab.txt you want to use
     vocab_size  - if use "pinyin", all available pinyin types, common alphabets (also those with accent) and symbols
                 - if use "char", derived from unfiltered character & symbol counts of custom dataset
                 - if use "byte", set to 256 (unicode byte range) 
@@ -144,6 +139,12 @@ def get_tokenizer(dataset_name, tokenizer: str = "pinyin"):
     elif tokenizer == "byte":
         vocab_char_map = None
         vocab_size = 256
+    elif tokenizer == "custom":
+        with open (dataset_name, "r", encoding="utf-8") as f:
+            vocab_char_map = {}
+            for i, char in enumerate(f):
+                vocab_char_map[char[:-1]] = i
+        vocab_size = len(vocab_char_map)
 
     return vocab_char_map, vocab_size
 
@@ -425,6 +426,7 @@ def get_librispeech_test(metalst, gen_wav_dir, gpus, librispeech_test_clean_path
 
 def load_asr_model(lang, ckpt_dir = ""):
     if lang == "zh":
+        from funasr import AutoModel
         model = AutoModel(
             model = os.path.join(ckpt_dir, "paraformer-zh"), 
             # vad_model = os.path.join(ckpt_dir, "fsmn-vad"), 
@@ -433,6 +435,7 @@ def load_asr_model(lang, ckpt_dir = ""):
             disable_update=True,
             )  # following seed-tts setting
     elif lang == "en":
+        from faster_whisper import WhisperModel
         model_size = "large-v3" if ckpt_dir == "" else ckpt_dir
         model = WhisperModel(model_size, device="cuda", compute_type="float16")
     return model
@@ -444,6 +447,7 @@ def run_asr_wer(args):
     rank, lang, test_set, ckpt_dir = args
 
     if lang == "zh":
+        import zhconv
         torch.cuda.set_device(rank)
     elif lang == "en":
         os.environ["CUDA_VISIBLE_DEVICES"] = str(rank)
@@ -451,10 +455,12 @@ def run_asr_wer(args):
         raise NotImplementedError("lang support only 'zh' (funasr paraformer-zh), 'en' (faster-whisper-large-v3), for now.")
 
     asr_model = load_asr_model(lang, ckpt_dir = ckpt_dir)
-
+    
+    from zhon.hanzi import punctuation
     punctuation_all = punctuation + string.punctuation
     wers = []
 
+    from jiwer import compute_measures
     for gen_wav, prompt_wav, truth in tqdm(test_set):
         if lang == "zh":
             res = asr_model.generate(input=gen_wav, batch_size_s=300, disable_pbar=True)
@@ -503,7 +509,7 @@ def run_sim(args):
     device = f"cuda:{rank}"
 
     model = ECAPA_TDNN_SMALL(feat_dim=1024, feat_type='wavlm_large', config_path=None)
-    state_dict = torch.load(ckpt_dir, map_location=lambda storage, loc: storage)
+    state_dict = torch.load(ckpt_dir, weights_only=True, map_location=lambda storage, loc: storage)
     model.load_state_dict(state_dict['model'], strict=False)
 
     use_gpu=True if torch.cuda.is_available() else False
@@ -559,7 +565,7 @@ def load_checkpoint(model, ckpt_path, device, use_ema = True):
         from safetensors.torch import load_file
         checkpoint = load_file(ckpt_path, device=device)
     else:
-        checkpoint = torch.load(ckpt_path, map_location=device)
+        checkpoint = torch.load(ckpt_path, weights_only=True, map_location=device)
 
     if use_ema == True:
         ema_model = EMA(model, include_online_model = False).to(device)

requirements.txt

@@ -5,25 +5,19 @@ datasets
 einops>=0.8.0
 einx>=0.3.0
 ema_pytorch>=0.5.2
-faster_whisper
-funasr
 gradio
 jieba
-jiwer
 librosa
 matplotlib
-numpy==1.23.5
+numpy<=1.26.4
 pydub
 pypinyin
 safetensors
 soundfile
-# torch>=2.0
-# torchaudio>=2.3.0
+tomli
 torchdiffeq
 tqdm>=4.65.0
 transformers
 vocos
 wandb
 x_transformers>=1.31.14
-zhconv
-zhon

requirements_eval.txt

@@ -0,0 +1,5 @@
+faster_whisper
+funasr
+jiwer
+zhconv
+zhon

samples/story.toml

@@ -0,0 +1,19 @@
+# F5-TTS | E2-TTS
+model = "F5-TTS"
+ref_audio = "samples/main.flac"
+# If an empty "", transcribes the reference audio automatically.
+ref_text = ""
+gen_text = ""
+# File with text to generate. Ignores the text above.
+gen_file = "samples/story.txt"
+remove_silence = true
+output_dir = "samples"
+
+[voices.town]
+ref_audio = "samples/town.flac"
+ref_text = ""
+
+[voices.country]
+ref_audio = "samples/country.flac"
+ref_text = ""
+

samples/story.txt

@@ -0,0 +1 @@
+A Town Mouse and a Country Mouse were acquaintances, and the Country Mouse one day invited his friend to come and see him at his home in the fields. The Town Mouse came, and they sat down to a dinner of barleycorns and roots, the latter of which had a distinctly earthy flavour. The fare was not much to the taste of the guest, and presently he broke out with [town] “My poor dear friend, you live here no better than the ants. Now, you should just see how I fare! My larder is a regular horn of plenty. You must come and stay with me, and I promise you you shall live on the fat of the land.” [main] So when he returned to town he took the Country Mouse with him, and showed him into a larder containing flour and oatmeal and figs and honey and dates. The Country Mouse had never seen anything like it, and sat down to enjoy the luxuries his friend provided: but before they had well begun, the door of the larder opened and someone came in. The two Mice scampered off and hid themselves in a narrow and exceedingly uncomfortable hole. Presently, when all was quiet, they ventured out again; but someone else came in, and off they scuttled again. This was too much for the visitor. [country] “Goodbye,” [main] said he, [country] “I’m off. You live in the lap of luxury, I can see, but you are surrounded by dangers; whereas at home I can enjoy my simple dinner of roots and corn in peace.”

scripts/eval_infer_batch.py

@@ -127,7 +127,7 @@ local = False
 if local:
     vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
     vocos = Vocos.from_hparams(f"{vocos_local_path}/config.yaml")
-    state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", map_location=device)
+    state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", weights_only=True, map_location=device)
     vocos.load_state_dict(state_dict)
     vocos.eval()
 else:

scripts/prepare_csv_wavs.py

@@ -0,0 +1,132 @@
+import sys, os
+sys.path.append(os.getcwd())
+
+from pathlib import Path
+import json
+import shutil
+import argparse
+
+import csv
+import torchaudio
+from tqdm import tqdm
+from datasets.arrow_writer import ArrowWriter
+
+from model.utils import (
+    convert_char_to_pinyin,
+)
+
+PRETRAINED_VOCAB_PATH = Path(__file__).parent.parent / "data/Emilia_ZH_EN_pinyin/vocab.txt"
+
+def is_csv_wavs_format(input_dataset_dir):
+    fpath = Path(input_dataset_dir)
+    metadata = fpath / "metadata.csv"
+    wavs = fpath / 'wavs'
+    return metadata.exists() and metadata.is_file() and wavs.exists() and wavs.is_dir()
+
+
+def prepare_csv_wavs_dir(input_dir):
+    assert is_csv_wavs_format(input_dir), f"not csv_wavs format: {input_dir}"
+    input_dir = Path(input_dir)
+    metadata_path = input_dir / "metadata.csv"
+    audio_path_text_pairs = read_audio_text_pairs(metadata_path.as_posix())
+
+    sub_result, durations = [], []
+    vocab_set = set()
+    polyphone = True
+    for audio_path, text in audio_path_text_pairs:
+        if not Path(audio_path).exists():
+            print(f"audio {audio_path} not found, skipping")
+            continue
+        audio_duration = get_audio_duration(audio_path)
+        # assume tokenizer = "pinyin"  ("pinyin" | "char")
+        text = convert_char_to_pinyin([text], polyphone=polyphone)[0]
+        sub_result.append({"audio_path": audio_path, "text": text, "duration": audio_duration})
+        durations.append(audio_duration)
+        vocab_set.update(list(text))
+
+    return sub_result, durations, vocab_set
+
+def get_audio_duration(audio_path):
+    audio, sample_rate = torchaudio.load(audio_path)
+    num_channels = audio.shape[0]
+    return audio.shape[1] / (sample_rate * num_channels)
+
+def read_audio_text_pairs(csv_file_path):
+    audio_text_pairs = []
+
+    parent = Path(csv_file_path).parent
+    with open(csv_file_path, mode='r', newline='', encoding='utf-8') as csvfile:
+        reader = csv.reader(csvfile, delimiter='|')
+        next(reader)  # Skip the header row
+        for row in reader:
+            if len(row) >= 2:
+                audio_file = row[0].strip()  # First column: audio file path
+                text = row[1].strip()          # Second column: text
+                audio_file_path = parent / audio_file
+                audio_text_pairs.append((audio_file_path.as_posix(), text))
+
+    return audio_text_pairs
+
+
+def save_prepped_dataset(out_dir, result, duration_list, text_vocab_set, is_finetune):
+    out_dir = Path(out_dir)
+    # save preprocessed dataset to disk
+    out_dir.mkdir(exist_ok=True, parents=True)
+    print(f"\nSaving to {out_dir} ...")
+
+    # dataset = Dataset.from_dict({"audio_path": audio_path_list, "text": text_list, "duration": duration_list})  # oom
+    # dataset.save_to_disk(f"data/{dataset_name}/raw", max_shard_size="2GB")
+    raw_arrow_path = out_dir / "raw.arrow"
+    with ArrowWriter(path=raw_arrow_path.as_posix(), writer_batch_size=1) as writer:
+        for line in tqdm(result, desc=f"Writing to raw.arrow ..."):
+            writer.write(line)
+
+    # dup a json separately saving duration in case for DynamicBatchSampler ease
+    dur_json_path = out_dir / "duration.json"
+    with open(dur_json_path.as_posix(), 'w', encoding='utf-8') as f:
+        json.dump({"duration": duration_list}, f, ensure_ascii=False)
+
+    # vocab map, i.e. tokenizer
+    # add alphabets and symbols (optional, if plan to ft on de/fr etc.)
+    # if tokenizer == "pinyin":
+    #     text_vocab_set.update([chr(i) for i in range(32, 127)] + [chr(i) for i in range(192, 256)])
+    voca_out_path = out_dir / "vocab.txt"
+    with open(voca_out_path.as_posix(), "w") as f:
+        for vocab in sorted(text_vocab_set):
+            f.write(vocab + "\n")
+
+    if is_finetune:
+        file_vocab_finetune = PRETRAINED_VOCAB_PATH.as_posix()
+        shutil.copy2(file_vocab_finetune, voca_out_path)
+    else:
+        with open(voca_out_path, "w") as f:
+            for vocab in sorted(text_vocab_set):
+                f.write(vocab + "\n")
+
+    dataset_name = out_dir.stem
+    print(f"\nFor {dataset_name}, sample count: {len(result)}")
+    print(f"For {dataset_name}, vocab size is: {len(text_vocab_set)}")
+    print(f"For {dataset_name}, total {sum(duration_list)/3600:.2f} hours")
+
+
+def prepare_and_save_set(inp_dir, out_dir, is_finetune: bool = True):
+    if is_finetune:
+        assert PRETRAINED_VOCAB_PATH.exists(), f"pretrained vocab.txt not found: {PRETRAINED_VOCAB_PATH}"
+    sub_result, durations, vocab_set = prepare_csv_wavs_dir(inp_dir)
+    save_prepped_dataset(out_dir, sub_result, durations, vocab_set, is_finetune)
+
+
+def cli():
+    # finetune: python scripts/prepare_csv_wavs.py /path/to/input_dir /path/to/output_dir_pinyin
+    # pretrain: python scripts/prepare_csv_wavs.py /path/to/output_dir_pinyin --pretrain
+    parser = argparse.ArgumentParser(description="Prepare and save dataset.")
+    parser.add_argument('inp_dir', type=str, help="Input directory containing the data.")
+    parser.add_argument('out_dir', type=str, help="Output directory to save the prepared data.")
+    parser.add_argument('--pretrain', action='store_true', help="Enable for new pretrain, otherwise is a fine-tune")
+
+    args = parser.parse_args()
+
+    prepare_and_save_set(args.inp_dir, args.out_dir, is_finetune=not args.pretrain)
+
+if __name__ == "__main__":
+    cli()

speech_edit.py

@@ -85,8 +85,9 @@ local = False
 if local:
     vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
     vocos = Vocos.from_hparams(f"{vocos_local_path}/config.yaml")
-    state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", map_location=device)
+    state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", weights_only=True, map_location=device)
     vocos.load_state_dict(state_dict)
+    
     vocos.eval()
 else:
     vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")

train.py

@@ -9,10 +9,10 @@ target_sample_rate = 24000
 n_mel_channels = 100
 hop_length = 256
 
-tokenizer = "pinyin"
+tokenizer = "pinyin" # 'pinyin', 'char', or 'custom'
+tokenizer_path = None # if tokenizer = 'custom', define the path to the tokenizer you want to use (should be vocab.txt)
 dataset_name = "Emilia_ZH_EN"
 
-
 # -------------------------- Training Settings -------------------------- #
 
 exp_name = "F5TTS_Base"  # F5TTS_Base | E2TTS_Base
@@ -44,8 +44,11 @@ elif exp_name == "E2TTS_Base":
 # ----------------------------------------------------------------------- #
 
 def main():
-
-    vocab_char_map, vocab_size = get_tokenizer(dataset_name, tokenizer)
+    if tokenizer == "custom":
+        tokenizer_path = tokenizer_path
+    else:
+        tokenizer_path = dataset_name
+    vocab_char_map, vocab_size = get_tokenizer(tokenizer_path, tokenizer)
 
     mel_spec_kwargs = dict(
             target_sample_rate = target_sample_rate,