Research And Implementation On Speaker Speech Adaptive Technique

Speaker independent speech recognition systems have achievedgreat progress in recent year. However the recognition performancedegrades rapidly when there is a mismatch between the testing andthe training condition, e.g. an outlier speaker. Therefore, adaptivetechniques are critically important to overcome these mismatches andto propel recognition to practical applications.This paper discussed various algorithms of adaptive technique,especially in the field of speaker adaptation. By analyzing theacoustical variations between speakers, the paper discussed andimplemented two speaker adaptation methods: Maximum A Posteriori(MAP) and Eigenvoice(EV).At the end of the paper, Experimentalresults show that these two methods work well in speaker adaptation.Then, a new approach for rapid adaptation is presented in thispaper. While EV method can achieve a fast adaptation rate when onlyfew data is available, MAP has desirable asymptotic properties. Byintroducing a EV module to MAP processing, the new approachintegrates these two methods to make use of their advantages andoffset their disadvantages.At the end of this paper, these three adaptive methods areimplemented by digital recognition. Experimental results show thatthe EV method performs very well for a small amounts of adaptationdata but the MAP method outperforms EV method as more data isavailable. Experimental results also show that the proposed newapproach not only achieves a good performance for a small amountsof adaptation data but also guarantees a consistent estimate as thedata size grows, It can effectively deal with the speaker variations,and is well suited for the robust speech recognition.