Automatic speech recognition in Mandarin for embedded platforms

In this paper, we describe a real-time automatic speech recognition system for Mandarin for low-cost embedded platforms using fixed-point digital signal processors. The hands-free, speaker-independent speech recognition system employs 41 mono-phone models for representing the sounds in Mandarin Chinese and 11 whole-word models for connected digit recognition. The system achieves greater than 98% recognition accuracy on our hands-free test database of 46 distinct command phrases. The system achieves 95.9% digit accuracy on a 14 speaker, hands-free, connected digit recognition database. The analysis of the results shows that for speakers without dialect, the digit recognition accuracy is almost 98%. We present a detailed analysis of the digit recognition results and propose further improvements. A realtime platform based upon Lucent’s DSP1627 fixed-point digital signal processor has been developed.