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

This paper provides a thorough description of the implementation of automatic speech recognition (ASR) algorithms on a fixed-point digital signal processor (DSP). It is intended to serve as a useful self-contained reference for DSP engineers to follow when developing similar applications. The work is based on a detailed analysis of hidden Markov model (HMM) based ASR algorithms. The computationally critical steps are clearly identified, and for each of them, different ways of optimization for real-time computation are suggested and evaluated. The trade-off among computational efficiency, memory requirements and recognition performance is illustrated quantitatively via three example systems, one for the recognition of isolated Chinese words and the other two for the recognition of English and Chinese digit strings, respectively. The paper also discusses about other techniques that can be implemented to further improve the recognition performance in real-world applications.