Speaker-independent embedded speech recognition using Hidden Markov Models

This paper describes the implementation of a realtime, speaker-independent isolated speech recognition system using Hidden Markov Models on a 32-bit ARM Cortex-M4F microcontroller. We introduce the theory, the requirements and details of the embedded implementation. The evaluation was made using a multi-speaker isolated-digit corpus of Argentinian Spanish, and its performance in terms of accuracy, speed and required memory was compared against a baseline Dynamic Time Warping recognizer, implemented previously on the same architecture. Test results show that the proposed HMM system outperforms the baseline system, and exhibits a recognition accuracy of 96.21% under a clean acoustic environment.