SenseVoiceSmall is Alibaba's multi-task encoder-only ASR: one forward pass through a 70-block SANM encoder emits the full transcript plus the spoken language ID, emotion, and audio-event tags through a single CTC head. Non-autoregressive design → 15× faster than Whisper-Large (70 ms for 10 s of audio in upstream's measurements).
70-block SenseVoiceEncoderSmall (1 entry block @ 560→512 + 49 main blocks + 20 tp blocks, all 512-dim, 4 heads, FSMN k=11 depthwise convolution branch — the same encoder body Fun-ASR-Nano-2512 ships, just here paired with a CTC head instead of an LLM decoder)
4 query embeddings (language / event / emotion / textnorm) prepended to the LFR fbank features so the encoder can emit rich annotations at those positions
CTC head (ctc.ctc_lo, 25055 SentencePiece pieces)
50+ languages with native LID (no whisper-tiny pre-step needed)
Three quants shipped (May 2026): F16 (448 MB), Q8_0 (240 MB), Q4_K (129 MB — recommended default). All three produce byte-identical transcripts on English (JFK) and Japanese (JSUT) clips end-to-end on M1 Metal. 72 tensors stay F16 in the Q4_K/Q8_0 quants because their leading dim isn't quant-block-aligned: 70× attn.fsmn.w (kernel=11 depthwise convolution) and 2× attn.qkv.w (560-dim input from the SANM context concat); the other ~280 weight matrices quantize cleanly.
What you get in the output
By default, stdout shows the clean transcript:
And so my fellow Americans ask not what your country can do for you, ask what you can do for your country.
With -oj the JSON output exposes the four rich-annotation tags as
explicit fields:
json
1{2"text":"And so my fellow Americans...",3"language":"en",4"audio_event":"Speech",5"emotion":"ANGRY",6"itn_flag":"withitn"7}
The legacy sensevoice_transcribe() C ABI still returns the original
prefixed string for callers that want it that way:
text
1<|en|><|HAPPY|><|Speech|><|withitn|>And so my fellow Americans...
2<|zh|><|NEUTRAL|><|Speech|><|withitn|>开饭时间早上9点至下午5点。
New callers should use sensevoice_transcribe_structured() which
returns the same six fields as a struct sensevoice_result.
Tag value sets:
Languages: zh / en / yue / ja / ko / nospeech
Emotions: HAPPY / SAD / ANGRY / NEUTRAL / EMO_UNKNOWN
Audio events: Speech / Music / Applause / Laughter / Cry / BGM (and more — the upstream set is open-ended)
Text norm: withitn (Arabic digits, punctuation) or woitn (raw)
Files
File
Size
Notes
sensevoice-small-q4_k.gguf
129 MB
Recommended default. 2× faster on M1 vs F16; byte-identical transcript on tested clips. Auto-download target for --backend sensevoice -m auto.
sensevoice-small-q8_0.gguf
240 MB
Larger but slightly closer to F16 numerically on borderline emotion-tag argmax cases.
sensevoice-small-f16.gguf
448 MB
F16 reference weights. Use when you want bit-stability against the upstream PyTorch reference for diff testing.
Quick Start
bash
1git clone https://github.com/CrispStrobe/CrispASR
2cd CrispASR
3cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
4cmake --build build --target crispasr-cli
56./build/bin/crispasr \7 --backend sensevoice \8 -m /path/to/sensevoice-small-q4_k.gguf \9 -f samples/jfk.wav -l en
1011# Or auto-download (resolves to Q4_K by default):12./build/bin/crispasr --backend sensevoice -m auto -f samples/jfk.wav -l en
Verification
crispasr-diff sensevoice is 76/76 PASS, byte-identical generated_text,
on Alibaba's own example zh.mp3; 75/76 PASS on samples/jfk.wav with
the single difference being the emotion-tag argmax flipping between
<|ANGRY|> and <|EMO_UNKNOWN|> (F16/op-order pushes that one slot
across a near-tied boundary; the transcript itself is byte-identical
in both runs). On Apple M1 Metal the runtime hits 15-22× realtime.
These GGUF files are a quantised / repackaged distribution of the upstream weights and inherit the FunASR Model License v1.1. Please attribute Alibaba / FunAudioLLM in downstream products.
If you use this model, please also cite the upstream FunASR work.
See the upstream model card for the canonical citation.
Upstream licence:other. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.
