Codebook-based Speech Enhancement Using Speech Presence Uncertainty

Through nearly four decades of development, researchers have proposed several speech enhancement algorithms including spectral-subtractive algorithm, Wiener filtering and statistical-model-based method. The performance of these approaches performs well in stationary noise but reduces severely while dealing with non-stationary noise. Codebook-based methods overcame this boundedness and worked well even in highly non-stationary noise. These methods stored the offline trained speech and noise Auto-Regressive(AR) coefficients in the respective priori codebooks. The corresponding AR gains are obtained online based on the spectral distortion measure. Then we exploit the obtained AR parameters to generate a Wiener filter which is applied to enhance the noisy speech. However, there are still several problems in the codebook-based methods. Firstly, due to the inaccuracy estimation of AR parameters, a large amount of background noise is remained in silence segments. Secondly, the codebook-based methods only focus on improving amplitude spectrum estimation of the noisy speech, and neglect its phase spectrum estimation. Finally, the traditional priori codebooks only model the spectral envelopes of speech and noise rather than their fine structure, which results in the background noise remained in the voiced segments of the enhanced speech. For the aforementioned three problems, we propose improved algorithms given as follows:1. For the problem of inaccuracy estimation of AR parameters in codebook-based speech enhancement, we propose a codebook-based speech enhancement using speech presence uncertainty. This method considers two hypotheses under the original Bayesian framework, i.e., speech presence hypothesis and speech absence hypothesis. The Bayesian estimation of speech and noise AR parameters, which is different from traditional ones, is the weighted sum of the obtained AR parameters under the two hypotheses. The corresponding weighted coefficients are speech presence probability(SPP) and speech absence probability(SAP) respectively, and the probabilities vary frame by frame online. The enhanced amplitude spectrum is still obtained by the reconstructed Wiener filter.2. For the traditional codebook-based speech enhancement neglected the phase spectrum estimation of clean speech, we proposed a novel phase estimation method. The method exploits the vector relationships of speech, noise and noisy signal spectra to obtain the cosine expression of the phase difference between speech and noisy signal. Then the clean speech phase is estimated using the arc-cosine function and noisy phase. Instead of the traditional noisy phase, the estimated phase and the enhanced amplitude are utilized to recover the clean speech signal. The perceptual quality of speech enhancement is further improved at low signal to noise ratio.3. For the aforementioned codebook-based speech enhancement using speech presence uncertainty, one problem is the lower estimation accuracy of SPP and SAP. Another is the remained noise between the harmonics of voiced speech. We propose a codebook-based speech enhancement with Bayesian AR parameters estimation for the above problems. In order to improve the estimation accuracy of SPP and SAP, this method estimate the AR parameters of speech and noise based on the current and past frames of noisy speech. Meanwhile, hidden Markov model(HMM) is applied to drive SPP and SAP, and we exploit the normalized cross-correction to adjust the transition probabilities between speech-presence and speech-absence states of HMM. In addition, we employ the a posteriori SPP in each time-frequency point to modify the Wiener filter, which can remove the residual noise between the harmonics of enhanced speech effectively. Our experiments demonstrate that the proposed method is superior to the reference methods.