[$200 bounty] Voice entry demo

[adeebshihadeh]

We want to use voice entry for setting nav destinations and more. The goal of this is to get a small, self-contained demo working of detecting a phrase like "Hey comma, navigate home". The script will run forever: detect start of speech, detect end of speech, do speech to text, then print it out.

Requirements:

• wake word detection needs to be local and fast
• speech to text can use an API
• must be low latency (<1s)
• new dependencies must be used only when necessary
• doesn't use any GPU; CPU and DSP are available though
• has to work reliably in the expected use case
  • mounted on the windshield and speaking
  • while openpilot is onroad (everything running)
  • can't make the rest of openpilot lag

See how we're running other models on the device: https://github.com/commaai/openpilot/tree/master/selfdrive/modeld

[singh13apoorv]

Can you explain this part? ""The goal of this is to get a small, self-contained demo working of detecting a phrase like "Hey comma, navigate home"." Do you want standalone script with demo or do you want the script to be in the repo but work separately for a demo?

[meirdev]

https://github.com/meirdev/commaai-assistant (with Whisper)

[jakethesnake420]

https://github.com/commaai/openpilot/pull/31010