Fixed-point implementation of isolated sub-word level speech recognition using hidden Markov models.

This paper presents a limited vocabulary isolated-word speech recognition system based on Hidden Markov Model (HMM) that involves two stage classification and is implemented on Texas Instruments' (TI) DaVinci embedded platform for a home infotainment system. A methodology using simple metric has been proposed for segmenting the words into sub-word units and these sub-words are used in the second stage to improve recognition accuracy. Also, a simple and efficient way of handling the out-of-vocabulary words using an additional HMM model is presented. We have achieved recognition accuracy of around 89% for a fixed point implementation on the TI DaVinci platform, demonstrating its suitability for embedded systems.