Noise robust speech recognition selectively using noise adapted HMM set
Hiroyuki Sakuno, Noboru Hayasaka, Youji Iiguni · 2014
This paper describes noise robust isolated word recognition whose target is stand-alone devices. To date, multi-condition model approach is typical for robust speech recognition under noisy conditions. This approach uses an averaging noise-adapted acoustic model set created from various noisy speech. Therefore, it does not consider characteristics of each noise. This paper proposes noise robust speech recognition which considers the characteristics. Our proposed method divides the training-noise data into some clusters by cluster analysis, then creates noise-adapted acoustic model set from clean speech (all words) and noise in each cluster. Besides, our proposed method also creates noise model from each cluster. By comparing noise models with noise part of noisy speech, an appropriate noise-adapted acoustic model set is used for testing. In an isolated word recognition task at SNR=0[dB], the proposed method improves 3.09% average recognition rate for trained noisy speech and 4.89% average recognition rate for untrained noisy speech as compared with multi-condition model approach.