Minimize kaldi nnet3 chain decoder, it can recognize domain words in embedded platform.
I used it in CPU: 1GHz, ram:16m, flash can ignore, because this decoder executable file is very small, contains models(based on your model), my demo is 484K.
About 2s sound wave, nosiy envirenment time cost is 300~400ms.
I also try porting it to STM32, STM32 memory resource is crisis, chain model decode need a lot of memory(Token,ForwardLinks), so you can refer online-wav-gmm-decode-faster.cc
where ProcessEmitting() caculate LogLikelihood() instead of chain neural networks compute result. I finished this and can work on STM32.
Others: This decoder not include vad , ns , agc ..., if these done, The recognize result will be better.
Moved to tools folder, then run below command:
./install_extern_libs.sh
if message "All libs Compiled Done!!!" shows that compile alsa, openblas successfully!
Notice: openblas just run a general compile, if you need optimize it, should specify the CPU target.
You can know more about in this github
Also support generic matrix operation, you can choose it when compile!
audio capture code is in src/audio/ path, there is only one api void Read(std::vector &data)
Default capture audio data is 2 seconds, if need modify, you can change AUDIO_LEN macro;
Please refer my blog: https://blog.csdn.net/cj1989111/article/details/102954071
Please refer my blog: https://blog.csdn.net/cj1989111/article/details/103265334
// TODO
Before step 1 need done.
cd src/build/
| Compile types | compile command |
|---|---|
| use openblas | make USE_OPEN_BLAS=true |
| without openblas | make |
then you can see decoder.bin executable file, in this demo, you can say wakeup word "智能管家", it can recognized and very few misidentifications.
Notice: models should follows as these ---> features has no ivector , no pitch, and mfcc 40 dims or 13 dims
The models I trained under aishell2
The trained scripts currently not open source.
Please refer to convert_models