New multispeaker model

.gitattributes

@@ -40,3 +40,27 @@ checkpoint_epoch=599.ckpt filter=lfs diff=lfs merge=lfs -text
 checkpoint_epoch=649.ckpt filter=lfs diff=lfs merge=lfs -text
 g_00140000_m filter=lfs diff=lfs merge=lfs -text
 checkpoints/g_00120000 filter=lfs diff=lfs merge=lfs -text
+checkpoints/checkpoint_epoch=100.ckpt filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_16.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_23.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_7.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_22.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_3.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_5.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_9.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_13.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_14.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_15.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_17.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_1.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_12.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_19.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_21.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_20.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_24.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_6.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_8.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_10.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_11.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_18.wav filter=lfs diff=lfs merge=lfs -text
+prompts/speaker_2.wav filter=lfs diff=lfs merge=lfs -text

app.py

@@ -1,49 +1,34 @@
 import os
-from pathlib import Path
 import torchaudio
-import gradio as gr
-
-import numpy as np
-
 import torch
-import json
-
-
-from hifigan.config import v1
-from hifigan.denoiser import Denoiser
-from hifigan.env import AttrDict
-from hifigan.models import Generator as HiFiGAN
+import gradio as gr
 
 
 from pflow.models.pflow_tts import pflowTTS
-from pflow.text import  text_to_sequence, sequence_to_text
+from pflow.text import  text_to_sequence, sequence_to_text, cleaned_text_to_sequence
 from pflow.utils.utils import intersperse
 from pflow.data.text_mel_datamodule import mel_spectrogram
 from pflow.utils.model import normalize
 from vocos import Vocos
 
 
-
-
-
-PFLOW_MODEL_PATH = 'checkpoints/checkpoint_epoch=649.ckpt'
-#PFLOW_MODEL_PATH = 'checkpoint_m_epoch=054.ckpt'
+PFLOW_MODEL_PATH = 'checkpoints/checkpoint_epoch=100.ckpt'
 VOCODER22_MODEL_PATH = 'BSC-LT/vocos-mel-22khz'
 VOCODER44_MODEL_PATH = 'patriotyk/vocos-mel-hifigan-compat-44100khz'
 
-HIFIGAN_MODEL_PATH = 'checkpoints/g_00120000'
-
-
-volnorm = torchaudio.transforms.Vol(gain=-32, gain_type="db")
-
 
+volnorm = torchaudio.transforms.Vol(gain=-15, gain_type="db")
 prompts_dir = 'prompts'
-print(os.listdir(prompts_dir))
 prompts_list = sorted(os.listdir(prompts_dir), key=lambda x: x.split('.')[0])
 
-def process_text(text: str, device: torch.device):
+def process_text(text: str, device: torch.device, ipa=False):
+    if ipa:
+        seq = cleaned_text_to_sequence(text)
+    else:
+        seq = text_to_sequence(text, ["ukr_cleaners"])
+    
     x = torch.tensor(
-        intersperse(text_to_sequence(text, ["ukr_cleaners"]), 0),
+        intersperse(seq, 0),
         dtype=torch.long,
         device=device,
     )[None]
@@ -53,18 +38,6 @@ def process_text(text: str, device: torch.device):
 
 
 
-
-def load_hifigan(checkpoint_path, device):
-    h = AttrDict(v1)
-    hifigan = HiFiGAN(h).to(device)
-    hifigan.load_state_dict(torch.load(checkpoint_path, map_location=device)["generator"])
-    _ = hifigan.eval()
-    hifigan.remove_weight_norm()
-    return hifigan
-
-
-
-
 def load_vocos(checkpoint_path, config_path, device):
     model = Vocos.from_hparams(config_path).to(device)
 
@@ -75,9 +48,6 @@ def load_vocos(checkpoint_path, config_path, device):
     return model
 
 
-def to_waveform(mel, vocoder, denoiser=None):
-    return  vocoder.decode(mel).clamp(-1, 1).cpu().squeeze()
-
 
 def get_device():
     if torch.cuda.is_available():
@@ -93,51 +63,54 @@ device = get_device()
 model = pflowTTS.load_from_checkpoint(PFLOW_MODEL_PATH, map_location=device)
 _ = model.eval()
 
-
-hifigan = load_hifigan(HIFIGAN_MODEL_PATH, device)
-vocos_22050 = Vocos.from_pretrained(VOCODER22_MODEL_PATH).to(device)
-#vocos_44100 = load_vocos('checkpoints/vocos_checkpoint_epoch=209_step=3924480_val_loss=3.7036_44100_11.ckpt', 'vocos.yaml', device)
 vocos_44100 = Vocos.from_pretrained(VOCODER44_MODEL_PATH).to(device)
-denoiser = None#Denoiser(vocoder, mode="zeros")
+vocos_22050 = Vocos.from_pretrained(VOCODER22_MODEL_PATH).to(device)
 
 
 @torch.inference_mode()
-def synthesise(text, prompt_selection, speed):
+def synthesise(text, ipa, prompt_selection, audio_prompt, temperature, speed):
+    print(text, prompt_selection, temperature, speed)
     if len(text) > 1000:
         raise gr.Error("Текст повинен бути коротшим за 1000 символів.")
 
-    text_processed = process_text(text.strip(), device)
-    prompt_audio_path = os.path.join(prompts_dir, prompt_selection)
-    wav, sr = torchaudio.load(prompt_audio_path)
-    prompt = mel_spectrogram(volnorm(wav), 1024, 80, 22050, 256, 1024, 0, 8000, center=False)[:,:,:264]
+    if audio_prompt:
+        wav, sr = torchaudio.load(audio_prompt)
+        wav = torchaudio.functional.resample(wav, orig_freq=sr, new_freq=44100)
+    else:
+        prompt_audio_path = os.path.join(prompts_dir, prompt_selection)
+        wav, _ = torchaudio.load(prompt_audio_path)
 
+    if ipa:
+        text_processed = process_text(ipa, device, ipa=True)
+    else:
+        text_processed = process_text(text.strip(), device, ipa=False)
+
+    prompt = mel_spectrogram(volnorm(wav), 2048, 80, 44100, 512, 2048, 0, 8000, center=False)[:,:,:264]
+
+    
     output = model.synthesise(
         text_processed["x"].to(device),
         text_processed["x_lengths"].to(device),
         n_timesteps=40,
-        temperature=0.0,
+        temperature=temperature,
         length_scale=1/speed,
         prompt=normalize(prompt, model.mel_mean, model.mel_std).to(device),
 
-        guidance_scale=2.0
+        guidance_scale=1.8
 
     )
-    waveform_vocos = vocos_22050.decode(output["mel"]).cpu().squeeze()
+    
     waveform_vocos_44100 = vocos_44100.decode(output["mel"]).cpu().squeeze()
-    waveform_hifigan = hifigan(output["mel"]).clamp(-1, 1).cpu().squeeze()
-    transform = torchaudio.transforms.Vol(gain=-18, gain_type="db")
-
-
-    return text_processed['x_phones'][1::2], (44100, waveform_vocos_44100.numpy()), (22050, waveform_vocos.numpy()), (22050, transform(waveform_hifigan).numpy())
+    waveform_vocos = vocos_22050.decode(output["mel"]).cpu().squeeze()
+    return text_processed['x_phones'][1::2], (44100, waveform_vocos_44100.numpy()), (22050, waveform_vocos.numpy())
 
 
 description = f'''
-# Експериментальна апка для генерації аудіо з тексту.
-
-    pflow checkpoint {PFLOW_MODEL_PATH}
-    Vocos 44100 аудіо - {VOCODER44_MODEL_PATH}
-    Vocos 22050 аудіо - {VOCODER22_MODEL_PATH}
-    HIFIGAN 22050 аудіо - {HIFIGAN_MODEL_PATH}
+Модель натренована на приватному датасеті з аудіо книжок створненому за допомогою програми
+[narizaka](https://github.com/patriotyk/narizaka).
+Програма може не коректно визначати деякі наголоси і не дуже добре перетворює цифри, акроніми і різні скорочення в словесну форму.
+На даний момент, відкритого рішення для української мови для цих проблем нема, тому якщо у вас є запитання,
+чи ви хочете допомогти їх вирішити приєднуйтесь до нашого чату в [телеграм](https://t.me/speech_synthesis_uk) або [discord](https://discord.gg/yVAjkBgmt4) 
 '''
 
 
@@ -147,7 +120,10 @@ if __name__ == "__main__":
         description=description,
         inputs=[
             gr.Text(label='Текст для синтезу:', lines=5, max_lines=10),
-            gr.Dropdown(label="Prompt audio", choices=prompts_list, value=prompts_list[0]),
+            gr.Text(label='Aбо IPA:', lines=5, max_lines=10),
+            gr.Dropdown(label="Виберіть промт", choices=prompts_list, value=prompts_list[0]),
+            gr.Audio(label="Або завантажте свій:", interactive=True, type='filepath', max_length=300, waveform_options={'waveform_progress_color': '#3C82F6'}),
+            gr.Slider(minimum=0.0, maximum=1.0, label="Шум", value=0.7),
             gr.Slider(minimum=0.6, maximum=2.0, label="Швидкість", value=1.1)
         ],
         outputs=[
@@ -164,21 +140,17 @@ if __name__ == "__main__":
                         streaming=False,
                         type="numpy",
                     ),
-            gr.Audio(
-                        label="HIFIGAN 22050 аудіо:",
-                        autoplay=False,
-                        streaming=False,
-                        type="numpy",
-                    )
             
         ],
         allow_flagging ='manual',
-        #flagging_options=[("Якщо дуже погоне аудіо, тисни цю кнопку.", "negative")],
         cache_examples=True,
-        title='',
-        # description=description,
-        # article=article,
-        # examples=examples,
+        title='Генерація української мови за допомогою pflowtts.',
+        examples=[ 
+            ['Мені тринадцятий минало. Я пас ягнята за селом. Чи то так сонечко сіяло, Чи так мені чого було? Мені так любо, любо стало, Неначе в бога. Уже прокликали до паю, А я собі у бур\'яні Молюся богу І не знаю, Чого маленькому мені Тоді так приязно молилось, Чого так весело було?', "meˈnʲi trɪˈnad͡zʲt͡sʲɐtɪi̯ mɪˈnaɫɔ. jɐ pɐs jɐɦˈnʲatɐ zɐ seˈɫɔm. t͡ʃɪ tɔ tɐk ˈsɔnet͡ʃkɔ sʲiˈjɐɫɔ, t͡ʃɪ tɐk meˈnʲi t͡ʃɔˈɦɔ bʊˈɫɔ? meˈnʲi tɐk ˈlʲubɔ, ˈlʲubɔ ˈstaɫɔ, neˈnat͡ʃe ʋ ˈbɔɦɐ.  ʊˈʒɛ prɔkɫɪkɐɫɪ dɔ ˈpajʊ, ɐ jɐ soˈbʲi ʊ bur-jɐˈnʲi moˈlʲusʲɐ ˈbɔɦʊ i ne ˈznajʊ, t͡ʃɔˈɦɔ mɐˈɫɛnʲkɔmʊ meˈnʲi toˈdʲi tɐk ˈprɪjɐznɔ mɔˈɫɪɫɔsʲ, t͡ʃɔˈɦɔ tɐk ˈʋɛseɫɔ bʊˈɫɔ?", '', 'prompts/speaker_22.wav', 0.6, 1.1],
+            ['Ти, малий, скажи малому, хай малий малому скаже, хай малий теля прив\'яже.', '',  '', 'prompts/speaker_11.wav', 0.4, 1.1 ],
+            ['По мірі розвитку клубу зростатиме і кількість його членів, а отже, команда менеджменту теж буде пропорційно збільшуватись. Яка ж команда потрібна клубу, що налічує, скажімо, сто осіб, і які компетенції повинна мати?', '',  '', 'prompts/speaker_20.wav', 0.7, 1.1],
+            ['Да ти дєтка гоніш! один рік? І що? Як ви задрали нити, рік вона не може, в когось діти мруть в день народження, викидні, а вона, бляха, рік не може, купи собі рожеве поні і реви побільше, дурепа.', 'dɐ tɪ dʲetkɐ ɦ��nʲiʃ! ɔˈdɪn rʲik? i ʃt͡ʃɔ? jɐk ʋɪ zɐˈdraɫɪ ˈnɪtɪ, rʲik wɔˈna ne ˈmɔʒe, ʋ kɔɦɔsʲ ˈdʲitɪ mrʊtʲ ʋ denʲ nɐˈrɔd͡ʒenʲːɐ, ˈʋɪkɪdʲnʲi, ɐ wɔˈna, ˈblʲaxɐ, rʲik ne ˈmɔʒe, kʊpɪ soˈbʲi rɔˈʒɛʋe ˈpɔnʲi i reʋɪ poˈbʲilʲʃe, dʊˈrɛpɐ.', '', 'prompts/speaker_5.wav', 0.7, 1.2]
+        ],
     )
     i.queue(max_size=20, default_concurrency_limit=4)
     i.launch(share=False, server_name="0.0.0.0")

pflow/data/text_mel_datamodule.py

@@ -39,6 +39,7 @@ class TextMelDataModule(LightningDataModule):
         f_max,
         data_statistics,
         seed,
+        min_sample_size,
     ):
         super().__init__()
 
@@ -68,6 +69,7 @@ class TextMelDataModule(LightningDataModule):
             self.hparams.f_max,
             self.hparams.data_statistics,
             self.hparams.seed,
+            self.hparams.min_sample_size,
         )
         self.validset = TextMelDataset(  # pylint: disable=attribute-defined-outside-init
             self.hparams.valid_filelist_path,
@@ -83,6 +85,7 @@ class TextMelDataModule(LightningDataModule):
             self.hparams.f_max,
             self.hparams.data_statistics,
             self.hparams.seed,
+            self.hparams.min_sample_size,
         )
 
     def train_dataloader(self):
@@ -134,6 +137,7 @@ class TextMelDataset(torch.utils.data.Dataset):
         f_max=8000,
         data_parameters=None,
         seed=None,
+        min_sample_size=4,
     ):
         self.filepaths_and_text = parse_filelist(filelist_path)
         self.n_spks = n_spks
@@ -146,6 +150,7 @@ class TextMelDataset(torch.utils.data.Dataset):
         self.win_length = win_length
         self.f_min = f_min
         self.f_max = f_max
+        self.min_sample_size = min_sample_size
         if data_parameters is not None:
             self.data_parameters = data_parameters
         else:
@@ -196,9 +201,9 @@ class TextMelDataset(torch.utils.data.Dataset):
 
     def __getitem__(self, index):
         datapoint = self.get_datapoint(self.filepaths_and_text[index])
-        if datapoint["wav"].shape[1] <= 66150:
+        if datapoint["wav"].shape[1] <= self.min_sample_size * self.sample_rate:
             ''' 
-            skip datapoint if too short (3s) 
+            skip datapoint if too short (<4s , prompt is 3s)
             TODO To not waste data, we can concatenate wavs less than 3s and use them
             TODO as a hyperparameter; multispeaker dataset can use another wav of same speaker
             '''

pflow/models/pflow_tts.py

@@ -5,7 +5,7 @@ import random
 import torch
 import torch.nn.functional as F
 
-
+from pflow import utils
 from pflow.models.baselightningmodule import BaseLightningClass
 from pflow.models.components.flow_matching import CFM
 from pflow.models.components.speech_prompt_encoder import TextEncoder
@@ -19,7 +19,7 @@ from pflow.utils.model import (
 from pflow.models.components import commons
 from pflow.models.components.aligner import Aligner, ForwardSumLoss, BinLoss
 
-
+log = utils.get_pylogger(__name__)
 
 class pflowTTS(BaseLightningClass):  # 
     def __init__(
@@ -31,6 +31,7 @@ class pflowTTS(BaseLightningClass):  #
         cfm,
         data_statistics,
         prompt_size=264,
+        dur_p_use_log=False,
         optimizer=None,
         scheduler=None,
         **kwargs,
@@ -42,6 +43,7 @@ class pflowTTS(BaseLightningClass):  #
         self.n_vocab = n_vocab
         self.n_feats = n_feats
         self.prompt_size = prompt_size
+        self.dur_p_use_log = dur_p_use_log
         speech_in_channels = n_feats
 
         self.encoder = TextEncoder(
@@ -151,7 +153,7 @@ class pflowTTS(BaseLightningClass):  #
             )
 
         logw_ = torch.log(1e-8 + attn.sum(2)) * x_mask
-        dur_loss = duration_loss(logw, logw_, x_lengths)
+        dur_loss = duration_loss(logw, logw_, x_lengths, use_log=self.dur_p_use_log)
 
         # aln_hard, aln_soft, aln_log, aln_mask = self.aligner(
         #     mu_x.transpose(1,2), x_mask, y, y_mask

pflow/text/cleaners.py

@@ -3,7 +3,7 @@ from ukrainian_word_stress import Stressifier
 import regex
 import re
 from ipa_uk import ipa
-stressify = Stressifier()
+stressify = Stressifier(stress_symbol="ˈ")
 
 
 _whitespace_re = re.compile(r"\s+")
@@ -15,5 +15,5 @@ def ukr_cleaners(text):
     text = collapse_whitespace(text)
     text = norm(text).lower()
 
-    text = regex.sub(r'[^\p{L}\p{N}\?\!\,\.\-\: ]', '', text)
+    text = regex.sub(r'[^\ˈ\p{L}\p{N}\?\!\,\.\-\: ]', '', text)
     return ipa(stressify(text), False)

pflow/text/symbols.py

@@ -2,14 +2,40 @@
 
 Defines the set of symbols used in text input to the model.
 """
+# _pad = "_"
+# _punctuation = '-´;:,.!?¡¿—…"«»“” '
+# _letters = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
+# _letters_ipa = (
+#     "éýíó'̯'͡ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
+# )
+
+
+# # Export all symbols:
+# symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
+
+# # Special symbol ids
+# SPACE_ID = symbols.index(" ")
+
+# _pad = "_"
+# _punctuation = '()-;:,.!?¡¿—…"«»“” '
+# _letters = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzАБВГҐДЕЄЖЗИІЇЙКЛМНОПРСТУФХЦЧШЩЬЮЯабвгґдеєжзиіїйклмнопрстуфхцчшщьюя"
+# _letters_ipa = (
+#     "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ̯͡"
+# )
+
+# # Export all symbols:
+# symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
+
+# # Special symbol ids
+# SPACE_ID = symbols.index(" ")
+
 _pad = "_"
 _punctuation = '-´;:,.!?¡¿—…"«»“” '
 _letters = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
 _letters_ipa = (
-    "éýíó'̯'͡ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
+    "éýíó'̯'͡ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲ'̩'ᵻ"
 )
 
-
 # Export all symbols:
 symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
 

pflow/utils/model.py

@@ -41,8 +41,11 @@ def generate_path(duration, mask):
     return path
 
 
-def duration_loss(logw, logw_, lengths):
-    loss = torch.sum((logw - logw_) ** 2) / torch.sum(lengths)
+def duration_loss(logw, logw_, lengths, use_log=False):
+    if use_log:
+        loss = torch.sum((logw - logw_) ** 2) / torch.sum(lengths)
+    else:
+        loss = torch.sum((torch.exp(logw) - torch.exp(logw_)) ** 2) / torch.sum(lengths)
     return loss
 
 

pflow/utils/utils.py

@@ -206,7 +206,6 @@ def get_user_data_dir(appname="pflow_tts"):
 
 
 def assert_model_downloaded(checkpoint_path, url, use_wget=False):
-    print(checkpoint_path)
     if Path(checkpoint_path).exists():
         log.debug(f"[+] Model already present at {checkpoint_path}!")
         return

prompts/speaker_1.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0dd649f417994c06fa40481eb41f7356eeb401881f567668c4526dd58567e10
+size 344026

prompts/speaker_10.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:686e88155eef9caac215f4a538be47656d4b992a3ae59f33e026b6f547bef1e8
+size 379046

prompts/speaker_11.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d86982b3e7dba7704088f8459e90c8aacfb9b27a011372c05746c3b906aa88fb
+size 396946

prompts/speaker_12.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1521a2400b70de8ab89dd7083ea7b2d69547dfa0fd44ba98d42c1c16ba959438
+size 458686

prompts/speaker_13.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2ce78bb8e505ad6687139c3fb6ff8ca0d649a139d65e58c0cbfff1b9095efc7
+size 458426

prompts/speaker_14.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:397b9e44669ead0f6baac314ad2f0b32218db9054b20a9c371474b5d49525ef9
+size 480736

prompts/speaker_15.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d614a13acf2f999a715194c3efb4f74a30f33b98823e6bf693bf26bf79d6f653
+size 573086

prompts/speaker_16.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c90db783c7be686c70d1b2d15400f414bc6ede6f59952b9a1b5c6bb1a96a16c7
+size 511346

prompts/speaker_17.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2ce2df72aa34de17f0e833c903a561e650b35d3b03b142c0bd6d0f1d7d4d4e2
+size 635086

prompts/speaker_18.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f0212146531c97752dbdf5615950298518748e655c9cf0d8f07f57e571eaf9a
+size 445198

prompts/speaker_19.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6e3e9226de668b70ab0c61e79b9ab42f3a5fb933fe874565b0a6daea1f0e570
+size 427556

prompts/speaker_2.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22b38f05add585086e84aa5b1ad5cc60902e97d911c7f5c25efa9907e3a3bd32
+size 674514

prompts/speaker_20.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b27c473712a82d638acfba3a1ea54bde3f0efefe111452101f3798c3aaa46f05
+size 485146

prompts/speaker_21.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f24e374c7ffcc4a2ceee696d6ee21996300997bf56a437212455780d2feab34
+size 379046

prompts/speaker_22.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ad69e9dd8c7c4886319730811920c0e4f56f486203b4ef76acab9f145cd9168
+size 1300738

prompts/speaker_23.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e97404bfd2b086ccb387ec5dff7ab5285dcb68ff789be361b03b0f368425ccf1
+size 1243406

prompts/speaker_24.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8dbbf87ac7a9c229ef8ab0cdc446c1e2408eeb7c0dab56784c221313fe2ffb7
+size 626266

prompts/speaker_3.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aef97ae48b0aac98ac3b3eb01494546d3a39fa1d864361efb01878f21b933fc7
+size 582166

prompts/speaker_5.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2430689318110e4730db7d9f83ad404535a1ac6c632b59483c0693d4804b6f9
+size 418738

prompts/speaker_6.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc9155a46c023af54cd0c1154a2d7fbc0a9f91d7aa1a43fef44d8726439d7c77
+size 388126

prompts/speaker_7.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db46d1910a8fe98d6d8e9b8cb680d732a1800cd8241eb7aa7690d30a2d46931d
+size 462836

prompts/speaker_8.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2895ab68d8d8093a3b51aac1202b0eb68e2d97082f399a09332ccd8df31377e0
+size 418736

prompts/speaker_9.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52555d2474f8d67237e11ef4c329c89388743a826767fb8a9431c196c2a5a021
+size 551036