Evaluation of noisy speech recognition based on noise reduction and acoustic model adaptation on the Aurora2 tasks

In this paper, we have evaluated a noisy speech recognition method based on noise reduction and acoustic model adaptation, on the AURORA2 tasks. For noise reduction method, we employed two noise reduction methods. One is an Adaptive Sub-Band Spectral Subtraction (ASBSS) method which can vary noise subtraction rate according to SNR in frequency bands at each frame. The other is a Kalman filtering estimation method which re-estimates the accurate speech spectra from speech spectra estimated by ASBSS. We estimated the accurate speech spectra by combining these methods. Usually, the noise reduction method has a problem that it deteriorates the recognition rate because of spectral distortion by residual noise of noise reduction and over estimation. To solve the problem in noise reduction method, adaptation of the acoustic models is employed by using unsupervised MLLR adaptation to spectral distortion. In evaluation on the AURORA2 tasks, our method showed the significant improvement in recognition accuracy both clean training condition and multi training condition.