A new feature extraction method for noisy speech recognition based on masking model

The performance of traditional speech recognition system degrades seriously in noisy environment. This paper presents a new speech feature extraction method based on masking properties of the human auditory system. We derive MFCC from masking model of noisy speech. The new method is evaluated by a task on TIMIT digit database (from 0 to 9, in English). Several types of noises from NoiseX92 database are added to the original speech to simulate noisy speech at different SNR. An average of 152% increase in recognition accuracy rate compared to classical MFCC is obtained in three different kinds of noises at 0dB SNR. The experimental results show that the performance of speech recognition systems can be greatly improved by using the new feature method under noisy environment.