s60sc / esp-box_firmware_alexa Goto Github PK

Hi, @s60sc I’m new to esp devices and even programming c.
I have a esp32-s3-box and esp-korvo.
I was able to install/update firmware but I search for a starting point to add Mqtt in addition to demo functions.
It seem that platformio, audrino and esphome Doesn’t support this board. Even if then I’m not able by now to get this demo functions developed.
so customizing Demos is the way to go.
How you did it?

Tried Alexa and works fine never did get round to trying if the 34mm mic spacing makes much difference.
Speed of sound is 343 m / s so divide by 16k is 21.5mm so likely the 16mm diff (less than a sample) will make little diff is not exact.

Dunno if you noticed but the 4 channel ADC provides ref channel sync as guess there must be latency in the i2s stream as AEC can be very clock sensitive.
I noticed this with speex as using different soundcards and clocks kills speex aec and the dev notes of speex highlight that pretty strongly so guess MASE is the same.
In the forum it was suggested it was due to DAC distortion that the input I2S could not be used for ref, I doubt that very much as the amp, speaker & room distortion will be huge in comparison.

Prob is with compiled blob libs is it stops much tinkering as if the clock latency of mic & ref could be matched as presume its just a few clocks, then maybe a hardware loopback would not be needed as pretty sure its there for sync and not dac distortion.

I have been wondering if tflite-micro is going to be optimised for the s3 as it seems to have more layer capacity and xtensa even have a LSTM that I think would run on S3.
I think there are better models that likely would run DS-CNN from https://github.com/google-research/google-research/tree/master/kws_streaming is one but really if LSTM is an option then CRNN like multinet is a good option for streaming KWS and having control of the dataset is really advantageous as if you save the signature and collate local data you can run additional epochs of actual use to really up accuracy.
Its strange multinet is CRNN based as ESP-DL seems to lack both LSTM or GRU or do the convert tools work even if layers are not implicit in the ESP32-DL layers?

Thinking about it if the mic i2s stream was an incrementing number then the 1st number 1 at the dac would have the clock count at mic input.
Dunno sure there is a way to measure the input stream clock latency as really struggling to source cost effective quad channel ADC's and pretty sure the necessary sync can be worked out in software.