Axiveri/Renpiper-mms-onnx-V1

Renpiper MMS ONNX V1

This is an ONNX conversion of Meta AI's facebook/mms-tts checkpoints, converted using the approach documented by the sherpa-onnx project (https://k2-fsa.github.io/sherpa/onnx/tts/mms.html), for on-device mobile inference via the sherpa_onnx runtime. The weights are unmodified from Meta AI's originals -- no fine-tuning has been applied, only a format conversion (PyTorch checkpoint -> ONNX) and vocabulary re-export (vocab.txt -> tokens.txt).

Language status

This table reflects exactly what this repo's own build notebook found and produced -- not an assumption about MMS's general language coverage.

Language	Status	ONNX size	Sample rate
Yoruba (yor)	Converted	108.8 MB	16000 Hz
Hausa (hau)	Converted	108.8 MB	16000 Hz
Igbo (ibo)	No raw checkpoint found (tried models/ibo, models/ig, models/igbo, models/ib, full_models/ibo, full_models/ig). A separate transformers-wrapped facebook/mms-tts-ibo repo does not exist, for reference -- not usable for this notebook's ONNX pipeline either way.	--	--
Nigerian Pidgin (pcm)	Converted	108.8 MB	16000 Hz

Each converted language lives in its own subfolder (yor/, hau/, ibo/, pcm/ -- whichever succeeded) containing model.onnx and tokens.txt.

What this is NOT

• Not an original or independently trained model.
• Not fine-tuned or adapted beyond the format conversion described above.
• Not available for commercial use or commercial relicensing under any name -- see License below.
• Not a claim that every language listed in Meta's general "1107 languages" MMS coverage is available here -- only what this notebook actually verified and converted.

License

CC-BY-NC 4.0, inherited unchanged from the source checkpoints. Non-commercial use only. This means this repo cannot be used as the basis for a paid product, client deliverable, or commercial deployment (including under a different product name) without a separate commercial license from Meta AI.

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}
}

Model developed by Vineel Pratap et al., Meta AI. All credit for the underlying model belongs to Meta AI / the MMS project, not to this repo's maintainer.

Inference (sherpa-onnx)

# pip install sherpa-onnx
import sherpa_onnx

tts = sherpa_onnx.OfflineTts(
    sherpa_onnx.OfflineTtsConfig(
        model=sherpa_onnx.OfflineTtsModelConfig(
            vits=sherpa_onnx.OfflineTtsVitsModelConfig(
                model="yor/model.onnx",
                tokens="yor/tokens.txt",
            ),
        ),
    )
)
audio = tts.generate("Your Yoruba text here")