phowhisper-salony-onnx

PhoWhisper-small — ONNX export for sherpa-onnx

An ONNX conversion of vinai/PhoWhisper-small, int8-quantised, packaged for sherpa-onnx.

No weights were retrained, fine-tuned or altered beyond format conversion and quantisation. This repository exists because the conversion output has no upstream home — VinAI publishes the PyTorch checkpoint, not an ONNX build — and an application that ships on-device recognition needs somewhere immutable to fetch one from.

Contents

One archive, phowhisper-small-onnx.tar.gz, holding exactly three files:

The three names are read literally by the consuming application. Renaming one breaks loading after a successful download, which is the worst place for it to fail — so if you republish, keep the names.

How it was produced

The dynamo=False part matters to anyone reproducing this: PyTorch's torch.onnx.export defaults to the dynamo exporter, which ignores dynamic_axes. Exported without it, the graph carries fixed shapes and sherpa-onnx fails at a Reshape node the moment the audio length differs from the tracing input.

Intended use, and what it is not

Built for short spoken commands in a Vietnamese salon — names, service names, counts — recognised entirely on-device.

Published WER for PhoWhisper-small, from the paper: 6.33% VIVOS (clean read speech), 11.08% Common Voice, 32.96% VLSP Task-2 (spontaneous). Spontaneous speech is the closest analogue to real use, and proper nouns are the worst case for any recogniser. Treat these as a floor.

Quantisation to int8 is not free. If you need the reference quality, export fp32 from the same checkpoint rather than assuming these files match it.

Licence

This artifact is a derivative of PhoWhisper and is distributed under the BSD 3-Clause licence, the licence of the original model. The copyright notice, licence text and disclaimer below travel with these files as a condition of redistribution.

Copy the LICENSE file verbatim from vinai/PhoWhisper-small into this section rather than reconstructing it from a template. BSD-3 texts vary in the copyright line and occasionally elsewhere, and the obligation is to reproduce their notice, not a generic one.

PhoWhisper itself is a fine-tune of OpenAI Whisper, which is MIT-licensed. That lineage is acknowledged here for accuracy; the licence governing this artifact is the BSD-3-Clause one above.

Per BSD-3 clause 3, neither the name VinAI Research nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. Describing a product as running on PhoWhisper is a statement of fact; describing it as endorsed by or in partnership with VinAI is not, and is not permitted here.

Citation

The authors ask that PhoWhisper be cited when it is incorporated into other software. It is:

@inproceedings{PhoWhisper,
  title     = {{PhoWhisper: Automatic Speech Recognition for Vietnamese}},
  author    = {Thanh-Thien Le and Linh The Nguyen and Dat Quoc Nguyen},
  booktitle = {Proceedings of the ICLR 2024 Tiny Papers track},
  year      = {2024}
}

Acknowledgements

• VinAI Research for PhoWhisper.
• OpenAI for Whisper, which PhoWhisper fine-tunes.
• k2-fsa for sherpa-onnx and its export tooling.