lichunnan / larynx2

A fast, local neural text to speech system

Larynx

A fast, local neural text to speech system.

echo 'Welcome to the world of speech synthesis!' | \
  ./larynx --model en-us-blizzard_lessac-medium.onnx --output_file welcome.wav

Voices

• U.S. English (16Khz, single speaker)
• German (16Khz, single speaker)
• Danish (22Khz, multispeaker)
• Norwegian (22Khz, single speaker)
• Nepali (16Khz, multispeaker)
• Vietnamese (16Khz, multispeaker)

Purpose

Larynx is meant to sound as good as CoquiTTS, but run reasonably fast on the Raspberry Pi 4.

Voices are trained with VITS and exported to the onnxruntime.

Installation

Download a release:

• amd64 (desktop Linux)
• arm64 (Raspberry Pi 4)

If you want to build from source, see the Makefile and C++ source. Last tested with onnxruntime 1.13.1.

Usage

1. Download a voice and extract the .onnx and .onnx.json files
2. Run the larynx binary with text on standard input, --model /path/to/your-voice.onnx, and --output_file output.wav

For example:

echo 'Welcome to the world of speech synthesis!' | \
  ./larynx --model blizzard_lessac-medium.onnx --output_file welcome.wav

For multi-speaker models, use --speaker <number> to change speakers (default: 0).

See larynx --help for more options.

Training

See src/python

Start by creating a virtual environment:

python3 -m venv .venv
source .venv/bin/activate
pip3 install --upgrade pip
pip3 install --upgrade wheel setuptools
pip3 install -r requirements.txt

Ensure you have espeak-ng installed (sudo apt-get install espeak-ng).

Next, preprocess your dataset:

python3 -m larynx_train.preprocess \
  --language en-us \
  --input-dir /path/to/ljspeech/ \
  --output-dir /path/to/training_dir/ \
  --dataset-format ljspeech \
  --sample-rate 22050

Datasets must either be in the LJSpeech format or from Mimic Recording Studio (--dataset-format mycroft).

Finally, you can train:

python3 -m larynx_train \
    --dataset-dir /path/to/training_dir/ \
    --accelerator 'gpu' \
    --devices 1 \
    --batch-size 32 \
    --validation-split 0.05 \
    --num-test-examples 5 \
    --max_epochs 10000 \
    --precision 32

Training uses PyTorch Lightning. Run tensorboard --logdir /path/to/training_dir/lightning_logs to monitor. See python3 -m larynx_train --help for many additional options.

It is highly recommended to train with the following Dockerfile:

FROM nvcr.io/nvidia/pytorch:22.03-py3

RUN pip3 install \
    'pytorch-lightning'

ENV NUMBA_CACHE_DIR=.numba_cache

See the various infer_* and export_* scripts in src/python/larynx_train to test and export your voice from the checkpoint in lightning_logs. The dataset.jsonl file in your training directory can be used with python3 -m larynx_train.infer for quick testing:

head -n5 /path/to/training_dir/dataset.jsonl | \
  python3 -m larynx_train.infer \
    --checkpoint lightning_logs/path/to/checkpoint.ckpt \
    --sample-rate 22050 \
    --output-dir wavs