Hydramus/Simba-TTS-tsn-onnx

Simba-TTS-tsn (ONNX for Transformers.js)

ONNX export of UBC-NLP/Simba-TTS-tsn, a VITS neural Setswana (Tswana) text-to-speech model, packaged for in-browser synthesis with 🤗 Transformers.js.

Runs entirely client-side (WebAssembly) — no server, no audio leaving the device.

Usage

import { pipeline } from '@huggingface/transformers';

const tts = await pipeline('text-to-speech', 'Hydramus/Simba-TTS-tsn-onnx', { dtype: 'fp32' });
const { audio, sampling_rate } = await tts('Dumela rra, o tsogile jang?');
// `audio` is a Float32Array at `sampling_rate` (16000 Hz)

Files

• onnx/model.onnx — fp32 (~109 MB). VITS is Conv-dominated and onnxruntime-web has no ConvInteger, so int8 quantization saves almost nothing here; fp32 is shipped for simplicity and quality.
• config.json, vocab.json, tokenizer_config.json, etc. — VITS char tokenizer.

Attribution & license

• Base model: UBC-NLP/Simba-TTS-tsn (The University of British Columbia — Natural Language Processing group), trained on the SimbaBench dataset.
• License: CC-BY-4.0 — same as the base model. Attribution required.
• This repository only re-packages the original weights as ONNX; all model credit belongs to UBC-NLP.