Custom microWakeWord models needed — Vance, Hex, Mochi — personal voice samples included

Configuration Voice Assistant
,
1

Hi all,

I’m running 4 Home Assistant Voice PE units and have custom wake words set up (Vance, Hex, Mochi) using models built from synthetic TTS only. False trigger rate is really bad so I want to retrain with real voice samples included.

I’ve recorded ~30 samples each for:

  • Vance — my voice (Geordie accent)
  • Hex — my voice + my partner’s voice (Burton accent)
  • Mochi — my daughter’s voice (age 11)

I’ve tried training locally using the TaterTotterson Docker trainer but my server is a Beelink Mini S with no GPU and it keeps running out of memory mid-training.

Would anyone with a GPU machine be willing to run the training with my personal samples included? Happy to share the WAV files via Google Drive or similar.

These are for Home Assistant Voice PE units running ESPHome so need to be microWakeWord v2 .tflite + .json output.

Thanks in advance

2

I could give it a try, right now I'm running the v3 release Releases · TaterTotterson/microWakeWord-Trainer-Nvidia-Docker · GitHub

But the voice samples need to be 16khz mono pcm wav iirc for it to work (I had issues getting ffmpeg running for whatever reason).

Oops forgot to tag @swanky