Speech recognition on DSP: issues on computational efficiency and performance analysis

This paper presents an investigation on the implementation of automatic speech recognition (ASR) using fixed-point digital signal processors (DSP). It addresses various important issues on computational efficiency and recognition performance. The investigation starts with a thorough and indepth understanding of hidden Markov model (HMM) based ASR algorithms. The algorithms are analyzed in great detail and the computationally critical steps are clearly identified. Effective optimization techniques are suggested to reduce the computation cost and make real-time processing possible. An isolated word as well as a connected-word recognition system have been successfully implemented and evaluated. The experimental results quantitatively reveal the trade-off relationship between resources requirement and recognition performance.