Poor wake word detection on Echo Show (LineageOS)

[daydy16]

Hi @msp1974,

First: I love VACA and View Assist! Thanks for your fantastic work. Unfortunately I stopped using VACA because wake word detection is too unreliable on Echo Show 8/5 1st Gen.

The Problem

• Only triggers very close (<1m), fails completely at 2m+ distance
• Frequent false positives even after threshold tweaking
• Never reliable enough for daily use

Perfect Alternative: Ava App

Found this by chance: https://github.com/knoop7/Ava

Reddit post with same experience: https://www.reddit.com/r/homeassistant/comments/1q78qvw/turn_any_android_716_device_into_a_smart_home/ [web:8]

Ava works crazy good:

• Triggers every time from 3-4m distance
• Zero false positives (even with TV/music)
• No tuning needed - just works

My Request/Suggestion

Could you take a look at the Ava source code? The developer has clearly cracked the audio pipeline + wake word detection challenge:

Code comparison (I'm not a dev - had AI compare both repos, maybe this helps):

Ava findings:

• MicrophoneInput.kt
  → Always VOICE_RECOGNITION + explicit AEC/AGC/NoiseSuppressor effects

• MicroFrontend.cpp
  → TensorFlow MicroFrontend (noise reduction, PCAN gain, log scale)

• WakeWordDetector.kt
  → TFLite models + sliding window averaging

• MicroWakeWord.kt
  → Per-model probabilityCutoff + window smoothing

VACA (dev branch):

• AudioInput.kt
  → VOICE_RECOGNITION only if "Voice Enhancer" enabled, otherwise MIC
  → No system AEC/AGC/NS effects

• BackgroundTask.kt
  → OpenWakeWord .onnx models + single threshold (no sliding window)

The Ava repo was private for a long time but got pushed publicly ~1 month ago. Not sure if that's the current production version - maybe worth reaching out to @knoop7 directly?

Questions

1. Is there a roadmap to improve wake word reliability on Echo Show devices?
2. Happy to test any audio pipeline improvements!

I'd love to return to VACA - the View Assist integration is superior to anything else. Just needs that rock-solid wake word detection like Ava delivers.

Thanks for the amazing work :)

[msp1974]

Hi, aware of this fork of brownards original Ava project work and actively looking at his work of migrating microwakeword to kotlin instead of our current use of openwakeword.

There is also feedback that openwakeword has fewer false detects than microwakeword (what the HAVPE is using), so likely first iteration will be a selection option.

[daydy16]

Is there a way for me to simply build my own APK and play around with it to test the best settings for my device? Is the source code in this repository up-to-date?