Aanand P - 2020-06-07

Hello everyone, I was wondering which acoustic model fits well in terms of accuracy and cpu usage for raspberry pi?

I am currently testing continuous, semi-continuous , ptm models on raspberry pi 2,3, and 4 for each of the modes KWS,JSGF,LM.

Right now I have completed testing KWS mode with each of the mentioned raspberry pi models and what I have observed is That continuous models performs badly in pi2 and 3 but performs well in pi4. The PTM model however shows decent accuracy and RTF(below 0.8) on pi2,3 and 4. Also through my observation, semi-continuous models do not perform badly in KWS mode, they give more or less the same accuracy as the PTM model.

My question is, can I conclude that Semi-continuous models can be used for the KWS mode in terms of accuracy and RTF? also, should I continue to test the other modes (LM and JSGF mode) with PTM model only for PI 2,3 and 4? Since continous models also seem to show good performance on PI4, I am thinking of considering it to test for LM and JSGF mode on PI4.

What do you think? Is it true that PTM models performs the best on embedded systems?