The Application Of Feature Compensation Method Based On Probability Model In Speech Recognition

Along with the great progress made in the state of the art of speech recognition technology, the speech recognition system is deployed in commercial application recently. However the recognition performance degrades rapidly when there was a mismatch between the testing and the training conditions with the variance of channel and noise. In order to overcome those influence of the negative environment, I tried to investigate the relative research on feature extraction, feature compensation, channel compensation, parameter training etc. The main work is as follows:Additive noise can bring nonlinear distortion of feature parameter so that the performance of recognition system decline. The article introduces a feature compensation method based on probability model. The method utilizes the prior probability distribution information of clean speech, noise and channel excursion. We can gain Minimum Mean Squared Error (MMSE) estimator of clean speech from noisy speech. And then based on the method we introduced dynamic difference character in cepstral domain. The joint distribution of prior probability is estimated using the Expectation Maximization algorithm in detail. And we utilized joint distribution of Gaussian prior probability to compensate feature parameters. The parameter estimation formula of joint Gaussian Mixture Model was also derived. The algorithm was tested in different kinds of noise and Signal Noise Ratio (SNR) .Subjective measure testified that this method can increase the correctness of continuous speech recognition by using the new compensation method. For instance, the correctness of system was 67.68% under the condition of 5dB tank noise without using feature compensation method, while the correctness could reach 72.44% by using feature compensation method under the same condition. But when the new feature compensation method proposed was adopted, the correctness was further improved to 85.23%.The effect of channel made character parameters offset linearly, so the performance of speech recognition system was influenced. We utilized joint distribution of Gaussian prior probability to compensate noise and channel simultaneously, and the result was that the channel mismatch between training environment and testing environment decreased. Experimental result showed that the correctness of system increased to 86.47% when using joint compensation algorithm of noise and channel under the condition of 5dB tank noise.