Noise robust speech recognition using parallel model compensation and voice activity detection methods

The main purpose of this study is to increase the performance of a speech recognition system under unknown noisy environments. Many methods have been proposed to increase the recognition performance of a speech recognition system for noisy speech. One of the most efficient techniques for dealing with the noisy speech is the Parallel Model Compensation (PMC) method. The speech recognition system will give the best results when there is no mismatch between the training and testing condition. Therefore, PMC method tries to minimize the mismatch between the training and testing conditions by estimating the speech recognition system parameters for noisy environment using the clean speech model and noise model. There are many well-known voice activity detection (VAD) methods for classifying the speech signal into speech and non-speech segments. Non-speech segments that are classified by a VAD method can be used to estimate the noise model. In this thesis, we propose to use VAD methods for estimating the noise model, and we also propose to use PMC for estimating the noisy speech model using the clean speech model and noise model which is estimated using a VAD method. In this study, performances of the baseline and three well know VAD methods have been compared for noisy speech recognition. Noise model were estimated using noise for the baseline method. For all method, PMC is used to estimate the noisy speech model. In addition to this, a new VAD method is proposed to estimate parameters of the noise model. The proposed VAD method's speech recognition performance is better than the most of the well-known VAD methods despite less computational requirement of the proposed VAD method compared to these well-known VAD methods. Key Words: Voice Activity Detection, Noise Estimation, Parallel Model Compensation.