search

cmusphinx / pocketsphinx

A small speech recognizer
Other
3.84k stars 710 forks source link

Q/A: pocketsphinx_continuous on GPU #72

Closed novitoll closed 7 years ago

novitoll commented 7 years ago

Is it possible to run pocketsphinx_continuous command on GPU? Is there any utils or prepared layers for that? Thanks

gorinars commented 7 years ago

Just curious, why would you need that and which prepared layers do you mean? Pocketsphinx has nothing to do with neural networks. Am I missing something?

novitoll commented 7 years ago

@gorinars, Right, pocketsphinx has nothing to do with NN, it has "plain" HMM model for acoustics. But I'm just curious how to speed it up. For example, 1:26:31 hour .wav file (166.1 MB disk memory, 16 KHz sample rate, pcm-16-le, monochannel) took ~ 37 min of ASR. Maybe there are some options where we could allocate heavy calculation (MFCCs for example) in GPU etc.

gorinars commented 7 years ago

MFCC calculation is not as heavy as model computations and search. I am not sure if some gain from GPU can be achieved now.

Check http://cmusphinx.sourceforge.net/wiki/decodertuning for some hints on how to make decoding faster.

You can also consider cutting long wavs on segments and processing parts on different threads/machines if you do batch decoding.

novitoll commented 7 years ago

Got it, thanks