The speaking rate adaptation algorithm in Putonghua continuous speech recognition

In continuous speech, the difference of speaking rates is big among speakers in different speaking environment. The variation of the speaking rates can cause recognition errors and affect the performance of LVCSR(Large Vocabulary Continuous Speech Recognition) systems. It is noted that the duration of neighboring speech units, which is affected by speaking rates, increases or decreases synchronously and a strong correlation exits between them. Based on the framework of DDBHMM (Duration Distribution Based HMM), a speaking rate adaptation algorithm is proposed. For utilizing the correlation information between duration of neighboring speech units. The experiments on connected digit and large vocabulary continuous speech show that the new algorithm is effective.