SenseVoice edge artifacts (RKNN / TensorRT) — for OpenVoiceStream
Edge-deployment artifacts for SenseVoice-Small offline ASR (encoder + CTC, zh/en/ja/ko/yue), used by OpenVoiceStream on Rockchip NPU (RKNN) and NVIDIA Jetson (TensorRT).
Attribution
- Model: SenseVoice-Small by FunAudioLLM (Alibaba / ModelScope) — https://github.com/FunAudioLLM/SenseVoice · https://huggingface.co/FunAudioLLM/SenseVoiceSmall
- Model weights license: FunASR Model Open Source License Agreement — https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE (redistribution permitted with attribution; retain source/author/model name)
- Source ONNX:
lovemefan/SenseVoice-onnx· sherpa-onnx export bycsukuangfj(k2-fsa). Decode assets (am.mvn,embedding.npy, sentencepiece model) are from the sherpa-onnx SenseVoice export.
The files here are format conversions / derivatives of the above (ONNX → RKNN / activation-rescaled ONNX for TensorRT). They inherit the FunASR Model License; this repo retains attribution per §2.2. No model code from any GPL/AGPL runtime is included — the OpenVoiceStream runtime + converters are independent.
Files
| file | platform | notes |
|---|---|---|
sense-voice-encoder.rk3576.fp16.rknn |
RK3576 | fp16 |
sense-voice-encoder.rk3588.fp16-scaled.rknn |
RK3588 | fp16 + K=8 activation rescale (fp16-safe on zh) |
sense-voice-encoder.scaled.fixed.onnx |
Jetson | rescaled fixed-shape ONNX; engine built on-device with host TensorRT |
sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17.tar.bz2 |
RPi / CPU | sherpa-onnx SenseVoice package (model.int8.onnx + tokens) |
am.mvn / embedding.npy / chn_jpn_yue_eng_ko_spectok.bpe.model |
RK/Jetson | decode assets |
- Downloads last month
- 6