VITS-OpenBible-Luganda

VITS Open Bible — Luganda

A multispeaker text-to-speech model for Luganda, trained from scratch on the Open Bible corpus using the VITS architecture (end-to-end TTS with adversarial learning, 22,050 Hz output) via the Coqui TTS framework.

Unlike zero-shot TTS models, VITS is conditioned on speaker embeddings learned during training. A speaker name from the training set must be supplied at inference time.

Files

Intended use

• Multispeaker TTS for Luganda using one of the training-set speaker voices.
• Research on multilingual TTS, low-resource TTS evaluation, and listening studies on Open Bible–style read-speech.

How to use

Install Coqui TTS:

pip install TTS

Download the checkpoint and run inference:

import torch
from huggingface_hub import hf_hub_download
from TTS.tts.utils.speakers import SpeakerManager
from TTS.utils.synthesizer import Synthesizer

repo_id  = "multilingual-tts/VITS-OpenBible-Luganda"
ckpt     = hf_hub_download(repo_id, "model_last.pth")
config   = hf_hub_download(repo_id, "config.json")
speakers = hf_hub_download(repo_id, "speakers.pth")

use_cuda = torch.cuda.is_available()
synthesizer = Synthesizer(
    tts_checkpoint=ckpt,
    tts_config_path=config,
    tts_speakers_file=speakers,
    use_cuda=use_cuda,
)

# Coqui's Synthesizer may not inject the speakers file into the model config
# automatically — restore the SpeakerManager manually when needed.
if synthesizer.tts_model.speaker_manager is None:
    synthesizer.tts_model.speaker_manager = SpeakerManager(
        speaker_id_file_path=speakers
    )

# List available speaker names
print(sorted(synthesizer.tts_model.speaker_manager.speaker_names))

wav = synthesizer.tts(
    text="...",          # text to synthesise in Luganda
    speaker_name="...",  # one of the speaker names printed above
    split_sentences=True,
)

Training data

• Source: davidguzmanr/open-bible-resources, config Luganda
• Size: approximately 21,553 utterances
• Speakers: multispeaker; speaker identity is fixed to one of the training-set voices and selected by name at inference time
• Sample rate: 22,050 Hz

Training procedure

• Architecture: VITS (Conditional Variational Autoencoder + adversarial training).
• Grapheme-level tokenizer, built from the training transcripts.
• Optimizer: AdamW, learning rate 2e-4.
• Training budget: 500,000 optimizer updates on 2 GPUs with mixed precision (bf16).

Audio preprocessing and training are reproducible via the upstream open-bible-models repo.

Evaluation

Evaluated alongside other Open-Bible TTS systems on character/word error rate (via Meta's Omnilingual ASR) and UTMOSv2 naturalness scores. See the open-bible-models repository for the evaluation pipeline and the open-bible-surveys repository for the human-listening survey methodology.