Gujarati MMS VITS — sherpa-onnx voice pack

Gujarati (ગુજરાતી) text-to-speech weights in ONNX format for the sherpa-onnx offline TTS runtime. This is a converted, single-speaker pack derived from Meta's Massively Multilingual Speech (MMS) project — specifically the Gujarati (guj) VITS checkpoint of facebook/mms-tts.

License: CC-BY-NC 4.0 (non-commercial) — see Attribution and LICENSE. The weights' only permitted source is Meta's checkpoint under the same license. Do not use for anything primarily intended for commercial advantage or monetary compensation.

Files

file role size
model.onnx VITS generator (opset 13), fp32 — inputs x, x_length, noise_scale, length_scale, noise_scale_w; output y 114 MB (108.8 MiB)
tokens.txt character vocabulary (token id per line), generated by the sherpa-onnx MMS exporter 462 B, 60 tokens

No lexicon, espeak data dir, or dict dir is needed: the model's frontend is a character frontend over the native Gujarati Unicode script (60-token vocab, add_blank handled internally by sherpa-onnx — this is not a uroman model).

Provenance / conversion metadata

property value
Source checkpoint G_100000.pth (generator trained for 100K updates) from facebook/mms-tts/models/guj
Source config / vocab config.json, vocab.txt (same directory)
Sampling rate 16000 Hz
Speakers 1
Frontend characters, add_blank: true
Uroman not required (train.ltr — native Gujarati script)
ONNX export opset 13, via the sherpa-onnx MMS exporter recipe (see below)
Duration predictor stochastic → synthesis is non-deterministic by design

Usage (sherpa-onnx)

sherpa-onnx-offline-tts \
  --vits-model=./model.onnx \
  --vits-tokens=./tokens.txt \
  --debug=1 \
  --output-filename=./guj.wav \
  "ગુજરાતી ભાષાનું ઉદાહરણ"

Default inference knobs (matching the checkpoint config, also the sherpa-onnx data-class defaults): noise_scale = 0.667, noise_scale_w = 0.8, length_scale = 1.0. The generated WAV is 16 kHz single-channel.

This pack is used by Rustle (on-device Android audiobook reader) as an optional, user-downloaded voice pack; Rustle's app code is separately licensed (CC-BY-NC-ND 4.0) and does not cover the weights in this repository.

Attribution

Required attribution under CC-BY-NC-4.0 (§3):

  • Licensed material (source of the weights):
  • Creator / licensor: Meta AI (MMS team; Vineel Pratap et al.)
  • Copyright: © Meta AI — MMS model weights
  • License: Creative Commons Attribution-NonCommercial 4.0 International (full text in LICENSE)
  • Modification notice: modified — the PyTorch checkpoint was converted to ONNX (opset 13) with the sherpa-onnx MMS exporter and re-hosted in this repository; contents otherwise unaltered
  • Warranties: the material is offered AS-IS per Sections 5 and 7 of the public license, without warranty of any kind

The original VITS implementation used to build MMS checkpoints is MIT-licensed (jaywalnut310/vits); the conversion recipe comes from the sherpa-onnx MMS documentation (k2-fsa; Apache-2.0 tooling). Neither upstream component changes the CC-BY-NC 4.0 status of these weights. fairseq MMS model code (git clone huggingface.co/spaces/mms-meta/MMS) is used at conversion time only and is not redistributed here.

Citation

@article{pratap2023mms,
    title={Scaling Speech Technology to 1,000+ Languages},
    author={Vineel Pratap and Andros Tjandra and Bowen Shi and Paden Tomasello
            and Arun Babu and Sayani Kundu and Ali Elkahky and Zhaoheng Ni
            and Apoorv Vyas and Maryam Fazel-Zarandi and Alexei Baevski
            and Yossi Adi and Xiaohui Zhang and Wei-Ning Hsu and Alexis Conneau
            and Michael Auli},
    journal={arXiv},
    year={2023}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for chintan992/rustle-mms-guj-tts

Quantized
(2)
this model

Paper for chintan992/rustle-mms-guj-tts