openWakeWord Stack Seeks Pi 5 Tuning

The bottleneck here is probably not just compute, but recognition strategy: short, fixed commands usually improve more from better front-end filtering, constrained intent handling, and custom data than from a larger general-purpose ASR model.

• –openWakeWord itself points to practical knobs for noisy rooms, including Speex noise suppression, VAD gating, threshold tuning, and custom verifier models
• –whisper.cpp supports quantized models and has ongoing Arm-oriented performance work, so Pi 5 should be viable, but tiny/freeform transcription may still be overkill for a 50-command vocabulary
• –For a bounded command set, a command classifier or grammar-constrained pipeline can beat open-ended ASR on latency and false accepts
• –Piper is only the TTS leg here; the accuracy work should focus upstream on microphone quality, denoising, VAD, and training data
• –If the environment is genuinely noisy, better audio capture hardware or a mic array may buy more reliability than another model swap