tts-hifigan-libritts-16kHz

Vocoder with HiFIGAN trained on LibriTTS

This repository provides all the necessary tools for using a HiFIGAN vocoder trained with LibriTTS (with multiple speakers). The sample rate used for the vocoder is 16000 Hz.

The pre-trained model takes in input a spectrogram and produces a waveform in output. Typically, a vocoder is used after a TTS model that converts an input text into a spectrogram.

Alternatives to this models are the following:

• tts-hifigan-libritts-22050Hz (same model trained on the same dataset, but for a sample rate of 22050 Hz)
• tts-hifigan-ljspeech (same model trained on LJSpeech for a sample rate of 22050 Hz).

Install SpeechBrain

pip install speechbrain

Please notice that we encourage you to read our tutorials and learn more about SpeechBrain.

Using the Vocoder

• Basic Usage:

import torch
from speechbrain.inference.vocoders import HIFIGAN
hifi_gan = HIFIGAN.from_hparams(source="speechbrain/tts-hifigan-libritts-16kHz", savedir="pretrained_models/tts-hifigan-libritts-16kHz")
mel_specs = torch.rand(2, 80,298)

# Running Vocoder (spectrogram-to-waveform)
waveforms = hifi_gan.decode_batch(mel_specs)

• Spectrogram to Waveform Conversion:

import torchaudio
from speechbrain.inference.vocoders import HIFIGAN
from speechbrain.lobes.models.FastSpeech2 import mel_spectogram

# Load a pretrained HIFIGAN Vocoder
hifi_gan = HIFIGAN.from_hparams(source="speechbrain/tts-hifigan-libritts-16kHz", savedir="pretrained_models/tts-hifigan-libritts-16kHz")

# Load an audio file (an example file can be found in this repository)
# Ensure that the audio signal is sampled at 16000 Hz; refer to the provided link for a 22050 Hz Vocoder.
signal, rate = torchaudio.load('tests/samples/ASR/spk1_snt1.wav')

# Ensure the audio is sigle channel
signal = signal[0].squeeze()

torchaudio.save('waveform.wav', signal.unsqueeze(0), 16000)

# Compute the mel spectrogram.
# IMPORTANT: Use these specific parameters to match the Vocoder's training settings for optimal results.
spectrogram, _ = mel_spectogram(
    audio=signal.squeeze(),
    sample_rate=16000,
    hop_length=256,
    win_length=1024,
    n_mels=80,
    n_fft=1024,
    f_min=0.0,
    f_max=8000.0,
    power=1,
    normalized=False,
    min_max_energy_norm=True,
    norm="slaney",
    mel_scale="slaney",
    compression=True
)

# Convert the spectrogram to waveform
waveforms = hifi_gan.decode_batch(spectrogram)

# Save the reconstructed audio as a waveform
torchaudio.save('waveform_reconstructed.wav', waveforms.squeeze(1), 16000)

# If everything is set up correctly, the original and reconstructed audio should be nearly indistinguishable

Inference on GPU

To perform inference on the GPU, add run_opts={"device":"cuda"} when calling the from_hparams method.

Training

The model was trained with SpeechBrain. To train it from scratch follow these steps:

1. Clone SpeechBrain:

git clone https://github.com/speechbrain/speechbrain/

1. Install it:

cd speechbrain
pip install -r requirements.txt
pip install -e .

1. Run Training:

cd recipes/LibriTTS/vocoder/hifigan/
python train.py hparams/train.yaml --data_folder=/path/to/LibriTTS_data_destination --sample_rate=16000

To change the sample rate for model training go to the "recipes/LibriTTS/vocoder/hifigan/hparams/train.yaml" file and change the value for sample_rate as required. The training logs and checkpoints are available here.