SenseVoiceSmall is a non-autoregressive speech model. One encoder pass and one CTC greedy collapse give the transcript, so there is no decoder loop. It also reports the spoken language, an emotion and an audio event.

Languages: Chinese, Cantonese, English, Japanese, Korean.

Files

File	Size	Notes
`model.onnx` + `model.onnx_data`	941 MB	fp32
`model_int8.onnx`	241 MB	dynamic int8, weights only (`MatMul` and `Gemm`)
`vocab.txt`	25055 tokens	SentencePiece pieces, id 0 renamed to `<blk>`
`config.json`		model type, prompt id tables

Usage

python

import onnx_asr

model = onnx_asr.load_model("sensevoice", "OpenVoiceOS/sensevoice-small-onnx")
print(model.recognize("audio.wav"))

# language detection is automatic, or force one
print(model.recognize("audio.wav", language="zh"))
# inverse text normalization adds punctuation and digits
print(model.recognize("audio.wav", use_itn=True))

The detected language, emotion, audio event and text-normalization mode are the first four entries of result.tokens when you call recognize with timestamps:

python

result = model.recognize("audio.wav", timestamps=True)
print(result.tokens[:4])   # ['<|zh|>', '<|NEUTRAL|>', '<|Speech|>', '<|woitn|>']
print(result.text)         # rich tokens are not in the text

Graph contract

Input	Type	Shape	Meaning
`features`	f32	`(B, T, 80)`	kaldi fbank, the onnx-asr `wespeaker` preprocessor
`features_lens`	i64	`(B,)`	valid frames per item
`language`	i64	`(B,)`	FunASR lid embedding index (`auto` 0, `zh` 3, `en` 4, `yue` 7, `ja` 11, `ko` 12)
`textnorm`	i64	`(B,)`	14 `withitn`, 15 `woitn`

Output	Type	Shape
`logprobs`	f32	`(B, ceil(T/6) + 4, 25055)`
`logprobs_lens`	i64	`(B,)`

The FunASR WavFrontend does three things: a kaldi fbank, a low frame rate stack (lfr_m 7, lfr_n 6) and the am.mvn mean-variance statistics. Only the fbank stays outside the graph, because onnx-asr already computes exactly that fbank for wespeaker (hamming window, snip_edges, dither 0, preemphasis 0.97, 400/160/512, 80 mel bins, log(max(x, eps))). The LFR stack and the CMVN are folded into model.onnx, together with the four prompt frames.

Batch padding is replaced in-graph by each item's last valid frame before the LFR stack, which is what the FunASR right-padding does. Batched and single-clip results are therefore identical.

The graph accepts any frame count T of 1 or more. The LFR stack pads the sequence up to the next multiple of 6 with the last valid frame, and the pad width is clamped at zero, so a frame count that already fills the last window needs no pad.

License and attribution

The weights are not Apache-2.0 and not an OSI-approved open-source license. They are a conversion of FunAudioLLM/SenseVoiceSmall, which is released by Alibaba Group under the FunASR Model Open Source License Agreement, version 1.1 (text). Read it before you use this model. The terms that matter, as written:

You may use, copy, modify and share the model.
You must attribute the source and author and retain the model names. This repository keeps the name SenseVoice and credits Alibaba Group and the FunASR project, as required.
The model is provided "for reference and learning purposes only". Alibaba Group accepts no responsibility for direct or indirect losses, and you carry all risk.
Section 4.2 forfeits your license if you engage in "unjustified denigration, malicious smearing, or baseless insults" against the software. This is a conduct-based termination clause. No OSI-approved license has one, so this model cannot be called open source.
Section 6 lets Alibaba Group revise the agreement at any time by publishing a new version, and continued use counts as acceptance.
Section 7 names the governing law only as "[Country/Region]", an unfilled placeholder in the published text.

Nothing in this export changes those terms. Only the onnx-asr runtime code is Apache-2.0; it lives in the onnx-asr repository, not here.