Study On Signal Subspace Approach For Speech Enhancement

The application of modern voice communication system can no longer be used in an ideal environment. Indeed, the currently existing systems are being operated under very adverse conditions. However, people still expect to receive satisfactory services. Therefore, the research of speech enhancement technology becomes more and more important.Speech enhancement means using signal processing tools to improve the intelligibility and the quality of a degraded speech signal. According to available number of speech signals, speech enhancement algorithms can be classified into two groups: single-channel (one microphone) case and multi-channel (more than one microphone) case. Most of today's techniques use the single-channel speech enhancement algorithm. Now it seems that speech enhancement is still a relatively difficult problem for two reasons: First, the nature and characteristics of the noise signals can change dramatically in time and application to application. It is therefore difficult to find versatile algorithms achieve better noise reduction effect in different practical environments. Second, there are two criteria widely used to measure the performance: intelligibility and quality. The former is objective, the latter is subjective. No speech enhancement system can simultaneously improve both.In the first part of Chapter 3, we introduced the signal subspace approach. This approach is decomposing the vector space of the noisy signal into two orthogonal subspaces: the signal-plus-noise subspace and the noise subspace. This is possible because it is well accepted that the clean speech can be modeled with a low-rank model. This approach is based on eigenvalue decomposition (EVD) of the covariance matrix of the input speech vector. The signal coefficients in the signal subspace can be processed individually using a diagonal gain matrix in order to suppress any remaining noise components. Then, the enhanced speech can be reconstructed in the time domain.The masking effects of the human ear is a very interesting phenomenon that has attracted the interest of speech enhancement scholars. Study found that using masking properties of the human ear may offer a good solution to trade-off between signal distortion and residual noise level. Experiments also showed that the use of masking in frequency-domain methods can improve the quality of the enhanced speech. However, the use of masking in conjunction with the signal subspace approach has not been attempted; one apparent reason is that the human perceptual properties work in the frequency domain. Consequently, all available masking models are developed in the frequency domain. Lately, a few solutions to this problem start to emerge. In the second part of Chapter 3, we introduced a frequency domain to eigendomain transformation. Based on this transformation, we presented a new signal subspace approach for noise reduction which combines the advantage of hearing masking effect and the subspace method. The simulation results show that this approach can obtain good speech enhancement effect.In most speech enhancement algorithms, the estimation of the noise spectrum is critical for the performance of the algorithms. If the noise estimation is too low, annoying residual noise will be audible, while if the noise estimation is too high, speech will be distorted resulting possibly in intelligibility loss. In Chapter 4, we introduced a statistical method for background noise estimation. This method is update continuously in every frame using time-frequency smoothing factors calculated based on speech presence probability in each frequency bin of the noisy speech spectrum.Finally, we proposed an enhanced system based on a perceptual signal subspace approach and a noise estimation algorithm for non-stationary noises. This algorithm can achieve with better performance than other subspace methods.