SenseVoiceSmall-INT8-ONNX
sherpa-onnx distribution of SenseVoiceSmall (int8), used as a local, non-autoregressive speech recognizer.
Why a repackaging exists at all
The official FunASR ONNX export
(iic/SenseVoiceSmall-onnx) cannot be loaded by sherpa-onnx, and not because
of the quantization — model_quant.onnx there is already int8. Two format
differences block it:
- the graph carries no ONNX metadata, and sherpa-onnx's SenseVoice loader
requires at least
vocab_size(loading fails with'vocab_size' does not exist in the metadata); - it ships
tokens.json(a JSON list), while sherpa-onnx reads a token table with one symbol per line.
This repository is the sherpa-onnx packaging of the same weights: identical model,
with the metadata sherpa-onnx expects and a tokens.txt in its format. am.mvn
is embedded in the graph, so only two files are needed.
Files
| file | bytes | sha256 |
|---|---|---|
model.int8.onnx |
239233841 | c71f0ce00bec95b07744e116345e33d8cbbe08cef896382cf907bf4b51a2cd51 |
tokens.txt |
315894 | f449eb28dc567533d7fa59be34e2abca8784f771850c78a47fb731a31429a1dc |
Total 228.5 MB — against 941 MB and 2293 MB for the two autoregressive tiers this is intended to complement.
Measured
On an i5-12600KF (16 logical cores, 6P+4E), decoding the same 11 × 6 s segments of the same real English recording:
| engine | ms per 6 s segment | RTF |
|---|---|---|
| Qwen3-ASR 0.6B int8 (autoregressive) | 2744 | 0.46 |
| SenseVoiceSmall int8 (this package) | 226 @4 threads / 206 @8 threads | 0.04 |
Roughly 12x, with an essentially identical transcript (996 vs 1002 characters over the eleven segments). The cost is also flat in output length: being encoder-only + CTC, it does not pay a per-token decode cost, which is what makes the autoregressive tier slow on smaller machines.
Languages
auto, zh, yue, en, ja, ko — five languages. This is not a
drop-in replacement for a multilingual tier; it is meant to sit alongside one.
License
Apache License 2.0, as declared by the model's ModelScope page
(iic/SenseVoiceSmall).
This repository only changes the packaging — the same weights in sherpa-onnx's container format — not the licensing terms.