xiao esp32-s3 from seed + respeaker for a voice agent solution - audio and wakeword issues

[sunwukong]

Hello,

I'm new to the espressif hardware and in general to embedded, I come from traditional backend dev, but I've been really wishing for an actually smart alexa or google dot and decided to try building one myself that uses modern LLM.

I picked up the respeaker with a xiao esp32-s3 because it made sense to try to offload all the audio processing to a dediacted board instead of doing it in software and the respeaker is substantially cheap than a korvo2 and in theory equally able if not better with the XMOS XU316 audio chip.

I tried to break down the problem and built a small repo to test audio recording and playback, another to test the wakeword, but I've ran into a wall and spent days trying to debug to no avail.

The recording repo works, but I have an issue where when reproducing the recorded audio the volume gets louder and louder and nothing I tried to prevent that work. The wakeword instead never worked. I tried to install the skainet project and that sort of worked (I can see the wakework being detected but then it's only static from the speaker). Bottom line I know the hardware works and is capable of doing what I want, but somehow I can't pull it all together.

Has anybody worked with the respeaker board from seeed studio and can advise me on how to get something going?

thank you in advance.

[ahsrabrifat]

Use Porcupine or Snowboy on a Pi/PC backend. Much easier and more robust. ESP just streams mic audio.