Skip to content

Hark v0.3.0-alpha

Pre-release
Pre-release

Choose a tag to compare

@pioborgelt pioborgelt released this 13 Nov 16:43
· 5 commits to main since this release

The first public alpha of Hark: a wake word engine for English and German where you just type the phrase. No recordings, no training per word.

What's in it

  • Home Assistant app. Add https://github.com/pioborgelt/hark as an app repository, install Hark Wake Word, type your phrases. Home Assistant finds it on its own.
  • Docker image for Home Assistant Container, Core or another machine: ghcr.io/pioborgelt/hark:alpha (amd64 and arm64).
  • Python package and CLI with hark listen to try it on your microphone and wyoming-hark to run the server yourself.

Setup steps for all three are in the README, and the FAQ covers everything I tripped over while testing it in fresh Home Assistant installs.

How well it works

These numbers are for exactly the two model files attached to this release, all zero-shot.

On the Picovoice wake word benchmark (six words, real noise at 10 dB, at most 0.1 false alarms per hour) Hark gets 78.1%. sherpa-onnx, the best open-source zero-shot engine I could find, gets 66.0%. On clean audio Hark reaches 91.0%.

On my own "Okay PAI" benchmark with 28 hours of speech and music as negatives, Hark catches 85.9% in English without a single false alarm (92.0% if you allow two). German is clearly weaker: 57.0% with zero false alarms and 62.8% with two. The positives in that benchmark are synthetic voices, so take it as a sanity check rather than a promise for your microphone.

All the tables are in docs/BENCHMARKS.md.

Model files

Hark downloads these automatically and checks their SHA256, so you only need them if you want to look inside. They're also on Hugging Face.

  • hark-stage1-r221-h384.pt (9.1 MB, 2.3M parameters): the small model that listens all the time.
    SHA256 c095fc5431e07a939eb482d18125e446a917864a8398edc15200bff316ebe0de
  • hark-stage2-r222-v384.pt (22.1 MB, 5.5M parameters): the second model that double-checks every candidate.
    SHA256 bbdb7928802bd04cfc0b399bfea6433dbf8ee1cf48cb59c6edf272deabcc3743

The weights are CC BY-NC 4.0 (free for personal use, Home Assistant setups, tinkering and research). The code is MIT.

Good to know

  • English and German only for now.
  • Hark confirms a wake word about 0.3 seconds after you finish it. Pause briefly before your command, otherwise the first word gets cut off.
  • Linux satellites in Home Assistant can go quiet after a reboot or a failed command. The FAQ has a small automation that fixes it.
  • I haven't measured it on a Raspberry Pi yet. If you run it on one, I'd love to hear how it goes.

This is an early alpha that I build in my free time. If it misses you or goes off when it shouldn't, please open an issue with your phrase, the language and your microphone setup.