What's new in Kiosk Satellite 2026.8.38
Wake word detection gets meaningfully cheaper on low-powered devices: the app now runs quantized int8 vsWakeWord models, validated to behave identically to the originals, with the full-precision files one toggle away. Requires Voice Satellite version 2026.8.9 or higher.
New: Quantized int8 wake word models
- When Voice Satellite serves the new int8 model builds, the app now prefers them over the original fp32 files. Inference drops from 8.6ms to 5.9ms per window on an Echo Show 8, cutting listening CPU by 10-30% depending on how busy the device is, and the model download shrinks to about a third of its size.
- Nothing about detection behavior changes: the manifest, thresholds and gates are the same files the fp32 models use. The quantization was validated against the standard wake word benchmark suite on three wake words, with recall identical on every positive set (including real recordings) and false-positive deltas within a single borderline clip across millions of scored windows.
- On a Voice Satellite version without the int8 builds the app falls back to the fp32 files automatically, so nothing is required on your end and nothing breaks if you update one side before the other.
New: Prefer fp32 vsWakeWord models
- For anyone who wants zero quantization drift, a new toggle under Settings > Voice Satellite opts back into the original full-precision models. It costs 10-30% more CPU while listening to avoid about 2% confidence drift.
New: Model precision row
- The Wake word panel, in the device settings and the remote admin alike, now shows whether the models actually loaded are int8 or fp32, so the active variant is always visible rather than guessed.