| Speech enhancement aims at suppressing the noise in noisy speech and guaranteeing the quality of enhanced speech. Traditional speech enhancement algorithms such as specral subtraction, Wiener filtering and so on, are not suitable for dealing with non-statinary noise. It is mainly because they have no a priori information about signals. In order to solve the problem, some algorthms using a priori information of speech and noise have been proposed. The representative ones are the speech enhancement algorithm based on HMM(Hidden Markov Model) and the speech enhancement algorithm based on codebook. These algothms use the HMM and codebook to store the a priori information about the auto-regressive(AR) spectral shapes of speech and noise offline, and apply some estimators to estimate the AR model parameters(AR spectral shapes and gains) of speech and noise online. Then the obtained paramters are used to construct a Winer filter to enhance the noisy speech. Since the estimated AR spectral gain can track the nosie energy online, these alogorithm can better deal with non-stationary nosie. However, there are still some problems to be addressed. For example, for the traditional speech enhancement algorithm based on parallel HMM, it does not consider the mismatching problem of spectral energy between the training and test sets. For the traditional codebook-driven speech enhancement algorithms, one is the problem of noise classification. Another is the lower spectrum parameter estimation accuracy. Moreover, traditional codebook-driven speech enhancement algorithms can not suppress the noise between the harmonics of speech. Focusing on the above problems, we propose corresponding solutions in this paper.The contribution of this thesis is composed of the following three parts:Firstly, based on the traditional PHMM speech enhancement algorithm, this paper proposes a gain adaptive PHMM for speech enhancement. The algorithm use AR(Auto-Regressive) spectral coefficients and MFS(Mel-Frequency Spectral) coefficient as a parallel feature for training PHMM,which is constituted by a AR-HMM and a MFS-HMM. The AR-HMM is used to contrcut Wiener filter, and the MFS-HMM is applied to estimating weighting value of the Wiener filter. In order to solve the mismatching problem between the training and test sets, the proposed algorithm introduces two energy gain factor for adaptively adjusting the energy of speech and noise online, respectively. Meanwhile, the robustness of the proposed algorithm is improved.Secondly, based on the traditional codebook-driven speech enhancement algorithm, this paper presents a codebook-driven speech enhancement algorithm using Markov process and speech presence probability(SPP). The algorithm uses the Markov process to model the correlation between the adjacent code-vectors in the codebook for optimizing Bayesian minimum mean squared error(MMSE) estimator. The accuracy of parameter estimation is improved. Meanwhile, this algorithm combines the SPP with codebook-driven Wiener filter to suppress the noise between the harmonics of noisy speech. The perceptual quality of the enhanced speech is guaranteed.Finally, for the problem of lower estimated accuracy of AR spectral gains and the problem of noise classification in conventional codebook-driven speech enhancement algorithm, we propose a novel codebook-driven speech enhancement algorithm based on a multiplicative update of AR spectral gains. This algorithm uses a nosie estimation module to obtain the spectral shape of noise online instead of training a codebook offline for solving the problem of noise classification. And a multiplicative update rule is ultilized to estimate the AR gains of speech and noise more accurately. This algorithm can reserve more speech compoments in the obtained enhanced speech, which contains less residual noise. |
