train_vits.py

import os

from trainer import Trainer, TrainerArgs

from TTS.config.shared_configs import BaseAudioConfig
from TTS.tts.configs.shared_configs import BaseDatasetConfig
from TTS.tts.configs.vits_config import VitsConfig
from TTS.tts.datasets import load_tts_samples
from TTS.tts.models.vits import Vits, VitsArgs
from TTS.tts.utils.speakers import SpeakerManager
from TTS.tts.utils.text.tokenizer import TTSTokenizer
from TTS.utils.audio import AudioProcessor

# to read tsv files from common voice
import pandas as pd

# output_path = '/media/popos/Barracuda/Models/TTS_new/trained_common_voice'
# dataset_path = "/media/popos/Barracuda/Datasets/CommonVoiceMozillaIta/it_29-03-2021/cv-corpus-6.1-2020-12-11/it"
output_path = '/run/media/opensuse/Barracuda/Models/TTS_new/trained_common_voice'
dataset_path = "/run/media/opensuse/Barracuda/Datasets/CommonVoiceMozillaIta/cv-corpus-9.0-2022-04-27/it"

pretrained_path = '/run/media/opensuse/Barracuda/Models/TTS_new/trained_common_voice/vits_vctk-June-05-2022_03+45PM-0cf3265a/'

dataset_config = BaseDatasetConfig(
    name="vctk", meta_file_train="", language="it-it", path=dataset_path
)

# custom formatter implementation


def commonvoice_formatter(root_path, manifest_file, **kwargs):
    # from root path we have train.tsv, test.tsv and val.tsv or use validated.tsv that contains all
    txt_file = os.path.join(root_path, 'train.tsv')
    df = pd.read_csv(txt_file, sep='\t')
    items = []
    for i, data in df.iterrows():
        items.append({
            "text": data['sentence'],
            "audio_file": os.path.join(root_path, 'clips', data['path']),
            "speaker_name": data['client_id']
        })
    return items


audio_config = BaseAudioConfig(
    sample_rate=22050,
    win_length=1024,
    hop_length=256,
    num_mels=80,
    preemphasis=0.0,
    ref_level_db=20,
    log_func="np.log",
    do_trim_silence=True,
    trim_db=23.0,
    mel_fmin=0,
    mel_fmax=None,
    spec_gain=1.0,
    signal_norm=False,
    do_amp_to_db_linear=False,
    resample=True,
)

vitsArgs = VitsArgs(
    use_speaker_embedding=True,
)

config = VitsConfig(
    model_args=vitsArgs,
    audio=audio_config,
    run_name="vits_vctk",
    batch_size=32,
    eval_batch_size=16,
    batch_group_size=5,
    num_loader_workers=4,
    num_eval_loader_workers=4,
    run_eval=True,
    test_delay_epochs=-1,
    epochs=1000,
    text_cleaner="english_cleaners",
    use_phonemes=False,
    phoneme_cache_path=os.path.join(output_path, "phoneme_cache"),
    compute_input_seq_cache=True,
    print_step=25,
    print_eval=False,
    mixed_precision=True,
    max_text_len=325,  # change this if you have a larger VRAM than 16GB
    output_path=output_path,
    datasets=[dataset_config],
)

# INITIALIZE THE AUDIO PROCESSOR
# Audio processor is used for feature extraction and audio I/O.
# It mainly serves to the dataloader and the training loggers.
ap = AudioProcessor.init_from_config(config)

# INITIALIZE THE TOKENIZER
# Tokenizer is used to convert text to sequences of token IDs.
# config is updated with the default characters if not defined in the config.
tokenizer, config = TTSTokenizer.init_from_config(config)

# LOAD DATA SAMPLES
# Each sample is a list of ```[text, audio_file_path, speaker_name]```
# You can define your custom sample loader returning the list of samples.
# Or define your custom formatter and pass it to the `load_tts_samples`.
# Check `TTS.tts.datasets.load_tts_samples` for more details.
train_samples, eval_samples = load_tts_samples(
    dataset_config, eval_split=True, formatter=commonvoice_formatter)

# init speaker manager for multi-speaker training
# it maps speaker-id to speaker-name in the model and data-loader
speaker_manager = SpeakerManager()
speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
config.model_args.num_speakers = speaker_manager.num_speakers

# init model
model = Vits(config, ap, tokenizer, speaker_manager)

# init the trainer and 🚀
if pretrained_path:
    trainer = Trainer(
        TrainerArgs(
            continue_path=pretrained_path,
        ),
        config,
        output_path,
        model=model,
        train_samples=train_samples,
        eval_samples=eval_samples,
    )
else:
    trainer = Trainer(
        TrainerArgs(),
        config,
        output_path,
        model=model,
        train_samples=train_samples,
        eval_samples=eval_samples,
    )
trainer.fit()