arjo129/uSpeech

Performs phoneme-based speech recognition through fixed-point signal analysis on resource-constrained microcontrollers, using complexity coefficients (absolute derivative/integral ratios) to classify consonants versus vowels without requiring training data. Directly interfaces with Arduino's `analogRead()` for analog audio input and operates within tight memory constraints (160 bytes per signal object), though accuracy ranges from 30-40% at phoneme level to 80% for word-level matching. The approach trades real-time scheduling capability for processor efficiency, consuming ~3.2ms per sample on ATmega32-based boards.

475 stars. No commits in the last 6 months.