In this thesis work, an isolated word recognition system which employs speaker adaptation is described and the performance of the speaker adaptation algorithm is studied.
In the isolated word recognition system, input speech signal is first processed by a digital filter bank. The digital filter bank is implemented by the fast Fourier transform and its outputs are quantized by a vector quantizer(VQ). Then, the hidden Markov modeling(HMM) is applied to the VQ output sequence for each phoneme-like unit.
Each word model is constructed by concatenating a sequence of corresponding phoneme models. In testing, Viterbi scoring is used to obtain the proper sequence of phoneme models and a word model for a given utterance.
Speaker adaptation is done in two ways. One is VQ codebook adaptation, and the other is HMM parameter adaptation.
In the VQ codebook adaptation process, feature vectors from adaptation data (utterances of a unknown speaker) are quantized according to the reference VQ codebook. Then the codewords in the reference VQ codebook are replaced by centroids of the clusters to make an adapted VQ codebook. With the adaptation VQ codebook, the HMM parameter adaptation is done by estimating the state sequence of the adaptation data using Viterbi algorithm with respect to the refence HMM first, and then the transition and the symbol observation matrices.
Computer simulation was done to obtain the performance of the isolated word recognition system with speaker adaptation. Simulation results show that the VQ codebook adaptation alone is not effective, but much better performance was obtained when both the VQ codebook adaptation and the HMM parameter adaptation were done, simultaneously. The improvements in recognition accuracy obtained for phoneme-like unit recognition and isolated word recognition are 22~39% and 30~67%, respectively, compared with the results of the isolated word recognition system without speaker adaptation.
유사음소를 이용한 인식 시스템에서 feature로 filter bank 출력을 사용하였으며 VQ를 사용하여 data의 감축을 하고 HMM을 이용하여 각 유사음소를 모델링하였다.
유사음소의 분류는 같은 음소일지라도 단어에 나타나는 위치에 따라 변하므로 이음(allophone)을 고려하여 구분하였다. HMM의 구조는 음성의 특성을 잘 나타내는 left-to-right 형태로 모델링하였다. 그리고 test는 viterbi scoring을 이용하였다. 이 때 HMM의 state 수와 data의 frame 수를 이용하여 constraint를 정하여 인식 시간을 줄였다.
이러한 화자 종속 유사음소 인식시스템에서 화자에 따른 인식률 저하를 보상하기 위해 화자적응 알고리즘을 적용한다. 화자적응 알고리즘은 인식시스템의 구조나 특성에 따라 여러가지가 있으나 본 논문에서는 VQ의 codebook과 HMM parameter를 화자에 따라 적응시키는 방식을 취하였다. 적응 결과 codebook adaptation만으로는 적응의 효과가 없지만 HMM parameter adaptation을 같이 수행하면 codebook adaptation의 효과가 나타난다. 이와 같이 두 시스템에 대해서 적응을 한 결과, 유사 음소의 경우 23~39%의 인식률 향상을 얻었고, 격리 단어의 경우는 30~67% 인식률 향상을 이루었다.
