NVIDIA Hifigan Vocoder (en-US)

HiFiGAN [1] is a generative adversarial network (GAN) model that generates audio from mel spectrograms. The generator uses transposed convolutions to upsample mel spectrograms to audio.

Usage

The model is available for use in the NeMo toolkit [3] and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.

To train, fine-tune or play with the model you will need to install NVIDIA NeMo. We recommend you install it after you've installed the latest PyTorch version.

git clone https://github.com/NVIDIA/NeMo
cd NeMo
BRANCH = 'main'
python -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[all]

instantiate the model

Note: This model generates only spectrograms and a vocoder is needed to convert the spectrograms to waveforms. In this example HiFiGAN is used.

from huggingface_hub import hf_hub_download
from nemo.collections.tts.models import FastPitchModel
from nemo.collections.tts.models import HifiGanModel

REPO_ID = "Mastering-Python-HF/nvidia_tts_en_fastpitch_multispeaker"
FILENAME = "tts_en_fastpitch_multispeaker.nemo"
path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)

spec_generator = FastPitchModel.restore_from(restore_path=path)

REPO_ID = "Mastering-Python-HF/nvidia_tts_en_hifitts_hifigan_ft_fastpitch"
FILENAME = "tts_en_hifitts_hifigan_ft_fastpitch.nemo"
path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)

model = HifiGanModel.restore_from(restore_path=path)

Generate and save audio

import soundfile as sf
parsed = spec_generator.parse("You can type your sentence here to get nemo to produce speech.")
"""
speaker id:
    92     Cori Samuel
    6097   Phil Benson
    9017   John Van Stan
    6670   Mike Pelton
    6671   Tony Oliva
    8051   Maria Kasper
    9136   Helen Taylor
    11614  Sylviamb
    11697  Celine Major
    12787  LikeManyWaters
"""
spectrogram = spec_generator.generate_spectrogram(tokens=parsed,speaker=92)
audio = model.convert_spectrogram_to_audio(spec=spectrogram)
sf.write("speech.wav", audio.to('cpu').detach().numpy()[0], 44100)

Colab example

LINK : nvidia_tts_en_fastpitch_multispeaker

Input

This model accepts batches of text.

Output

This model generates mel spectrograms.

Model Architecture

FastPitch multispeaker is a fully-parallel text-to-speech model based on FastSpeech, conditioned on fundamental frequency contours. The model predicts pitch contours during inference. By altering these predictions, the generated speech can be more expressive, better match the semantic of the utterance, and in the end more engaging to the listener. FastPitch is based on a fully-parallel Transformer architecture, with a much higher real-time factor than Tacotron2 for the mel-spectrogram synthesis of a typical utterance. It uses an unsupervised speech-text aligner.

Training

The NeMo toolkit [3] was used for training the models for 1000 epochs.

Datasets

This model is trained on HiFiTTS sampled at 44100Hz, and has been tested on generating multispeaker English voices with an American and UK accent.

Performance

No performance information is available at this time.

Limitations

This checkpoint only works well with vocoders that were trained on 44100Hz data. Otherwise, the generated audio may be scratchy or choppy-sounding.

Mastering-Python-HF
/

nvidia_tts_en_hifitts_hifigan_ft_fastpitch