Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
34 lines
1.6 KiB
Markdown
34 lines
1.6 KiB
Markdown
# sherpa-onnx Zipformer keyword spotting
|
|
|
|
The `sherpa_kws` family runs streaming Zipformer transducer keyword-spotting
|
|
exports. Convert a model directory with:
|
|
|
|
```bash
|
|
conda run -n qwen3-tts python tools/models/convert_sherpa_kws_gguf.py \
|
|
/path/to/sherpa-onnx-kws-model /path/to/sherpa-kws-f32.gguf
|
|
```
|
|
|
|
The converter defaults to the original F32 encoder, decoder, and joiner ONNX
|
|
files. The default self-contained package is
|
|
`Sherpa-KWS-GGUF/sherpa-kws-zipformer-zh-en-3m-f32.gguf` under
|
|
`audio-cpp/audio.cpp-gguf`; it embeds configuration, tokens, keywords, and the
|
|
model spec. Only F32 is packaged.
|
|
|
|
The default checkpoint is
|
|
[sherpa-onnx-kws-zipformer-zh-en-3M-2025-12-20](https://modelscope.cn/models/pkufool/icefall-kws-zipformer-zh-en-3M-2025-12-20).
|
|
Its model author [confirmed Apache-2.0 licensing](https://github.com/k2-fsa/sherpa-onnx/issues/3802#issuecomment-6052663364).
|
|
Other sherpa-onnx exports may have different model-weight licenses.
|
|
|
|
```bash
|
|
build/debug/bin/audiocpp_cli --family sherpa_kws --task wake \
|
|
--model /path/to/sherpa-kws-zipformer-zh-en-3m-f32.gguf \
|
|
--backend cpu --threads 4 --audio input.wav --segments-out detections.json --log
|
|
```
|
|
|
|
Use `--mode streaming` for incremental input. Detections are speech segments
|
|
with the keyword in `text`, its confidence, and absolute start/end sample positions.
|
|
|
|
The default `keywords.txt` uses sherpa-onnx tokenized keyword rows. A request
|
|
can replace it with `keywords=...`; separate rows with newlines or `/`. Each
|
|
row contains token symbols and can include `:score`, `#threshold`, and
|
|
`@label`, matching sherpa-onnx keyword files.
|