souba67/mms-tts-ind-ONNX

mms-tts-ind - ONNX for Transformers.js

ONNX export of facebook/mms-tts-ind for in-browser inference.

• Recommended weights: fp32 (114.2 MB)
• Waveform correlation vs the PyTorch reference: 1.00000
• Sample rate: 16000 Hz | noise scales set to 0 for deterministic output

Quantization notes

VITS ends in a HiFi-GAN vocoder. Naive dynamic int8 over the whole graph produces noise, not speech - measured below.

Files that are in this repo. Two, and no others:

file	size	status
onnx/model.onnx	114.2 MB	fp32, correlation +1.00000, load this one
onnx/model_fp16.onnx	58.2 MB	fp16, never passed, do not use

model_fp16.onnx went up in the same commit as the fp32 export. It is in the repo because that commit put it there, not because it was measured and cleared. Every fp16 conversion tried in this export failed to load: onnxconverter_common could not place the fp16/fp32 boundaries consistently across a VITS graph, and three strategies each broke at a different node. There is no correlation figure for it because inference never ran. Do not reach for it because it is half the size. The demo site asks for dtype: 'fp32' on the wasm backend and never loads it.

Variants that were measured and never uploaded. These three were built and scored during Phase A, all three failed parity, and none of them was pushed. No file in the table below exists in this repo. The correlations are real measurements and are kept because they answer the question a reader actually has, which is whether a smaller download is available here. It is not.

variant	size when built	corr	verdict	in this repo
int8_all	38.4 MB	+0.1911	FAIL	no
int8_no_decoder	73.2 MB	+0.2086	FAIL	no
int8_per_channel	38.4 MB	-0.0500	FAIL	no

fp32 is the only file here that passes parity, so fp32 is the only file to load.

Correction history

Earlier versions of this card listed the three int8 variants in the same table as fp32 and closed with a line saying the others "are included for reference only". They were not included. int8_all, int8_no_decoder and int8_per_channel were built, measured, failed parity and were never pushed. The repo has held onnx/model.onnx and onnx/model_fp16.onnx and nothing else since the first upload on 11 August 2026. The measurements were real; the files were never here to download. On 13 August 2026 the section was split in two so that what was measured and what is downloadable are no longer the same table.

Those versions also said nothing at all about onnx/model_fp16.onnx, which is in the repo, was never cleared, and does not load. It is listed now.

Usage

import { pipeline } from '@huggingface/transformers'
const tts = await pipeline('text-to-speech', 'souba67/mms-tts-ind-ONNX',
                           { dtype: 'fp32' })
const out = await tts('Selamat datang')

Original model by Meta AI (Massively Multilingual Speech). Licence CC-BY-NC 4.0 carries over: non-commercial use only.