train_vocoder.py

# train_vocoder.py
import os
from trainer import Trainer, TrainerArgs
from TTS.utils.audio import AudioProcessor
from TTS.config.shared_configs import BaseAudioConfig
from TTS.vocoder.configs import HifiganConfig
from TTS.vocoder.datasets.preprocess import load_wav_data
from TTS.vocoder.models.gan import GAN


def main():
    output_path = os.path.dirname(os.path.abspath(__file__))
    data_path = os.path.join(output_path, "LJSpeech-1.1/wavs/")

    audio_config = BaseAudioConfig(
        sample_rate=22050,
        resample=False,
        do_trim_silence=True,
        trim_db=45,

        fft_size=1024,
        win_length=1024,
        hop_length=256,
        frame_shift_ms=None,
        frame_length_ms=None,

        num_mels=80,
        mel_fmin=0.0,
        mel_fmax=None,

        signal_norm=True,
        symmetric_norm=True,
        max_norm=4.0,
        clip_norm=True,
        ref_level_db=20,
        min_level_db=-100,
        spec_gain=20.0,
        log_func="np.log10",
        preemphasis=0.0,

        stats_path=None,
    )

    config = HifiganConfig(
        run_name="hifigan_ljspeech",
        run_description="HiFi-GAN v1 from scratch, GlowTTS-compatible mels",

        data_path=data_path,
        output_path=output_path,
        eval_split_size=10,

        audio=audio_config,

        epochs=2000,
        batch_size=64,
        eval_batch_size=16,
        num_loader_workers=4,
        num_eval_loader_workers=2,
        run_eval=True,
        test_delay_epochs=5,
        mixed_precision=True,

        seq_len=8192,
        pad_short=2000,
        use_noise_augment=True,

        lr_gen=2e-4,
        lr_disc=2e-4,

        print_step=50,
        print_eval=False,
        save_step=5000,
        save_n_checkpoints=5,
        save_checkpoints=True,
        log_model_step=10000,
        plot_step=500,
    )

    ap = AudioProcessor(config=config.audio)

    eval_samples, train_samples = load_wav_data(
        config.data_path,
        config.eval_split_size,
    )

    model = GAN(config)

    trainer = Trainer(
        TrainerArgs(),
        config,
        output_path,
        model=model,
        train_samples=train_samples,
        eval_samples=eval_samples,
        training_assets={"audio_processor": ap},
    )

    trainer.fit()


if __name__ == "__main__":
    main()