Scaling Speech Technology to 1,000+ Languages
Paper • 2305.13516 • Published • 13
mms-tts-asm
Assamese voice converted for bolo, a speech server that runs without PyTorch (sherpa-onnx).
Use it
bolo pull mms-tts-asm
Licence
Weights: CC-BY-NC-4.0 - non-commercial use only.
The full licence text is in LICENSE.
Attribution and provenance
mms-tts-asm: Assamese voice from Meta's Massively Multilingual Speech (MMS) project
Model: https://huggingface.co/facebook/mms-tts/tree/44cc7fb408064ef9ea6e7c59130d88cac1274671/models/asm
Authors: Vineel Pratap, Andros Tjandra, Bowen Shi, et al. (Meta AI), "Scaling Speech Technology to
1,000+ Languages", 2023, https://arxiv.org/abs/2305.13516
Licence: Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC-4.0),
https://creativecommons.org/licenses/by-nc/4.0/ (full text in LICENSE). NON-COMMERCIAL USE ONLY.
Changes: format conversion only (PyTorch checkpoint to ONNX with the original VITS inference code,
mms-meta/MMS at revision 65d863f41196654aa8b8f3dc586474a4c8f30934), made with tools/voices/export_mms.py of the
bolo project (https://github.com/cb-Sigma/bolo). The weights were not retrained or otherwise changed.
Inference code: VITS, MIT licence, Copyright (c) 2021 Jaehyeon Kim.
Source: https://huggingface.co/facebook/mms-tts/tree/44cc7fb408064ef9ea6e7c59130d88cac1274671/models/asm
This repository holds a format conversion only; the weights were not retrained.
Measured by the bolo maintainers
| round-trip WER (5 sentences) | 0.133 |
| round-trip CER | 0.032 |
| recogniser | indicconformer-as-community |
| real-time factor (2 threads, shared 4 vCPU) | 0.341 |
| time to first audio (ms) | 633 |
| date | 2026-11-13 |