Speaker-Independent Continuous Mandarin Speech Recognition Under Telephone Environments

This paper presents a study on speaker-independent continuous Mandarin speech recognition over the telephone. A comparison of several cepstral bias removal techniques such as cepstral mean subtraction (CMS), signal bias removal (SBR) and stochastic matching (SM) for telephone channel compensation was first investigated. Then some modifications and combinations of these techniques were developed for further improvement of the environmental robustness under telephone environments. To better estimate the contextual acoustics and coarticulation in spontaneous telephone speech, the between-syllable context-dependent phone-like units (such as triphones, hipbones and demiphones) were used to train the speech models. In addition, the discriminative capabilities of the speech models were further enhanced using the minimum classification error (MCE) algorithms. Experimental results showed that the achieved recognition rates for Mandarin syllables were as high as 59.53%, which indicated a 27.81% of error rate reduction.