Reduce time taken by STT / TTS

[celian-garcia]

As per the previous study #30 , we can identify time taking tasks

WUW process

The trace starts when the program starts to listen for the Wake Up Word. That's why we have a lot of WUW processes
Once the keyword is detected, it instantly start the STT, so we can eliminate WUW from the cause.

STT process

Even deducing the time during which we are talking, 5 sec is still a lot of time. It can perhaps be improved using a local library, or using directly Google API instead of SpeechRecognition that does't seems to be very maintained.

HTTP POST

We can probably improve the backend's /talk/text API, to do that we need to have finer grain details of each span in the trace to identify big consumer(s)

TTS process

Not sure if the root cause is the TTS or the play function on the Rasp. To be checked with finer grain.

Summary

It can be improved in every aspect, but generally, a finer grain tracing would be required.

[celian-garcia]

Might be interesting to follow this and retry https://wiki.linuxaudio.org/wiki/raspberrypi