DigitalUmuganda/lingala_vits_tts

Lingala Text-to-Speech

This model was trained on the OpenSLR's 71.6 hours aligned lingala bible dataset.

Model description

A Conditional Variational Autoencoder with Adversarial Learning(VITS), which is an end-to-end approach to the text-to-speech task. To train the model, we used the espnet2 toolkit.

Usage

First install espnet2

pip install espnet

Download the model and the config files from this repo. To generate a wav file using this model, run the following:

from espnet2.bin.tts_inference import Text2Speech
import soundfile as sf

text2speech = Text2Speech(train_config="config.yaml",model_file="train.total_count.best.pth")
wav = text2speech("oyo kati na Ye ozwi lisiko mpe bolimbisi ya masumu")["wav"]
sf.write("outfile.wav", wav.numpy(), text2speech.fs, "PCM_16")