Abstract
The matching of environmental conditions between the test speech and the training data is crucial to the automatic speech recognizer. If they are mismatched, the recognition accuracy often degrades seriously. Since the environmental condition is not fixed, it is necessary to adapt the reference model of speech so as to keep the match requirement. In this dissertation, some adaptation techniques are developed to adapt the continuous hidden Markov model based speech recognizer. These adaptation method can be classified into two categories; one is proposed by taking advantage of the reaction of speech feature under the corruption of noise, and the other tries to obtain the models of noisy speech through a combination of the noise model and the speech models. The norm of feature vector of the cepstral coefficients was shown to become smaller when the speech signal is corrupted by white noise. The projection method was proposed to adapt the cepstral mean by multiplying a scaling factor so as to compensate the noise effect. This method is effective for the corruption of white noise, but its performance could be limited for some types of noises. Investigating the behavior of the cepstral coefficients under various types of noises, it is found that the variation of cepstral vector is noise dependent and may be quite different to the white noise. In this study, we propose a representation of cepstral coefficients under the noise effect so as to adapt the cepstral mean with the reference of variation direction of cepstral vector. The variation direction accounts for the change of cepstral vector with the increase of noise power and can be determined once the noise spectrum is given. With this directional information, this method outperforms the projection method as shown in our experiments. In addition to the adaptation of cepstral mean, a joint adaptation of mean and variance is proposed by correlating the shrinkage of the mean and the variance with a simple function. Using the relationship function, an optimal shrinkage can be computed based on the framework of the projection method. Four functions are examined in this work to search for an optimal one so that the recognition accuracy is the best. The main idea in the second class of the adaptation technique is to use the models of noisy speech that matches the test condition. The noisy models can be obtained from the database of noisy speech. However, the re-training process is time-consuming and is impractical in a real application. The parallel model combination (PMC) method was proposed to obtain the cepstral mean and variance of noisy speech through a combination in the linear spectral domain with the noise. For the purpose, a domain mapping for the mean and variance is necessary and its closed form formulation can be derived by assuming that the combined signal is normal distributed in the log-spectral domain. In this dissertation, the mapping process is modified so that the discriminative ability of the adapted model is improved. The modified PMC method, termed weighted PMC, is shown useful to improve the recognition rate of noisy speech in low SNR. The PMC method is efficient compared with the re-training method. However, its computational requirement is still large, especially to a large set of model parameters. To speed up the adaptation process, a direct adaptation of cepstral variance is proposed by taking approximation on the adapted variances in the process of the WPMC method. The proposed method is fast because a great deal of computation for the domain mapping such as the matrix multiplication of DCT can be removed. The recognition results of the WPMC and its approximate version, i.e., the fast algorithm, are quite similar and the adaptation speed is at least 33 times faster as shown in our experiments.