Hello, what's the current progress of this PR? Thanks!

Hello, what's the current progress of this PR? Thanks!

Developing the runtime first, see k2-fsa/sherpa-onnx#505 , will clean up this PR soon.

Would it be possible to implement a KWS system using the output from the CTC branch, transforming it into a lattice to utilize Kaldi's decoders? Similar to what is done with kaldi-decoder/faster-decoder.h and kaldi-decoder/decodable-ctc.h? What part of the kaldi code must be implemented in kaldi-decoder in order to achieve that could you give me some direction?

@alucassch I indeed have the plan to use ctc branch, but I think I won't use the kaldi decoders. As for using the kaldi decoders, you can compile the keywords into a lattice, than decode the audios with this lattice (faster decoder is enough, I think), then for each frame (or chunk) you can match the suffix of decoded results with keywords candidates, if matching and the logprob is larger than given threshold the corresponding keyword is triggered. Sorry, I don't have much experience in this direction, so here is just my thought, you can try it yourself.

Here are some results of this PR, you can find more details in the RESULTS.md of each recipe.

English

The positive set is from https://github.com/pkufool/open-commands the negative set is the test set of gigaspeech.

Each metric has two columns, one for original model trained on gigaspeech, the other for finetune model.

small

large

Chinese

The positive set is from https://github.com/pkufool/open-commands the negative set is the test-net set of wenetspeech.

Each metric has two columns, one for original model trained on wenetspeech, the other for finetune model.

small

large and others

麻烦问下这有支持python接口使用吗，看文档https://k2-fsa.github.io/sherpa/onnx/kws/pretrained_models/index.html#sherpa-onnx-kws-pre-trained-models不是很清楚

麻烦问下这有支持python接口使用吗，看文档https://k2-fsa.github.io/sherpa/onnx/kws/pretrained_models/index.html#sherpa-onnx-kws-pre-trained-models不是很清楚

已添加，k2-fsa/sherpa-onnx#576

  --decoder-dim 320 \
  --joiner-dim 320 \
  --num-encoder-layers 1,1,1,1,1,1 \
  --feedforward-dim 192,192,192,192,192,192 \
  --encoder-dim 128,128,128,128,128,128 \
  --encoder-unmasked-dim 128,128,128,128,128,128 \

Are these numbers the result of extensive search, or chosen with some intuition? Thanks!

  --decoder-dim 320 \
  --joiner-dim 320 \
  --num-encoder-layers 1,1,1,1,1,1 \
  --feedforward-dim 192,192,192,192,192,192 \
  --encoder-dim 128,128,128,128,128,128 \
  --encoder-unmasked-dim 128,128,128,128,128,128 \

Are these numbers the result of extensive search, or chosen with some intuition? Thanks!

No, just with some intuition. We are searching for better models also smaller models.

您好，请问有提供kws微调前的预训练pt模型吗？只看到了onnx的。

您好，请问有提供kws微调前的预训练pt模型吗？只看到了onnx的。

看 results.md, 里面有链接。

There is a small bug in egs/wenetspeech/ASR/prepare.sh （line164 172 180）
it use a fixed num_splits 1000 , should be pieces=$(find data/fbank/M_split_${num_splits} -name "cuts_M.*.jsonl.gz")

@KIM7AZEN Thanks! Could you make a PR to fix it.

@KIM7AZEN Thanks! Could you make a PR to fix it.

ok. wait a moment

Chinese

small

Commands	FN in positive set	FN in positive set	Recall	Recall	FP in negative set	FP in negative set	False alarm (time / hour) 23 hours	False alarm (time / hour) 23 hours
	original	finetune	original	finetune	original	finetune	original	finetune
All	426 / 985	40/985	56.8%	95.9%	7	1	0.3	0.04

large and others

Commands	FN in positive set	FN in positive set	Recall	Recall	FP in negative set	FP in negative set	False alarm (time / hour) 23 hours	False alarm (time / hour) 23 hours
---	original	finetune	original	finetune	original	finetune	original	finetune
large	2429/4505	477 / 4505	46.1%	89.4%	50	41	2.17	1.78

What do 'small' and 'large and others' mean here? Are they referring to the size of the models or the sizes of different test sets? Why does the larger one seem to perform worse than the smaller one?

@zhuangweiji Test sets, you can see https://github.com/k2-fsa/icefall/blob/master/egs/wenetspeech/KWS/RESULTS.md for more details.