Smoothed unit HMM in Mandarin speech recognition

The base unit in Mandarin speech recognition can be a phoneme, demi-syllable or syllable. The demi-syllable system has fewer HMM models and need less computation, thus it is suitable for real-time systems. But due to the poor description of the acoustic properties of the speech signal, it generally shows a low performance compared to the syllable system. While a system based on the syllable or phoneme (tri-phone or di-phone) has much greater HMM models, and needs massive computation in training and recognition. In this paper, a compromised scheme is proposed. The new system is based on the demi-syllable, but the two demi-syllable HMMs are connected into a full syllable HMM in the training phase, so the data of the whole length of the syllable are used, and smoothing between two demi-syllables is introduced. This can increase the system performance without increasing the HMM models, and it fits into real-time systems with a DSP kernel.