Rui Cheng

Beamforming method can effectively remove background noise, even in the complex environment, so it is widely used in speech enhancement. We propose a novel Generalized Eigenvalue (GEV) beamforming with Blind Analytic Normalization (BAN) method. In this method, the GEV beamformer coefficients are constructed by estimating logarithmic power spectrum (LPS), which are used to filter multichannel speech signals, and post filter technology is used to further remove noise in the beamformed signals. Firstly, in order to estimate the LPS of speech signal in each channel, we use the data-driven method to train the deep neural network (DNN) model. Then, we use the well trained DNN model to estimate LPS, which is used to calculate the power spectral density (PSD) matrix of speech, and further obtain the coefficients of the GEV beamformer. Since the GEV beamformer will cause speech distortion, the BAN is employed to post-process the beamformed signal. Furthermore, single channel speech enhancement is used to reduce residual noise. Our experiment is conducted in 8-channel simulation data set. The experimental results show that, compared with some existing speech enhancement methods, the proposed method can effectively remove background noise and achieve better speech enhancement effect.

Multi-channel speech coding and enhancement is an indispensable technology in speech communication. In order to verify the effectiveness of multi-channel speech coding and enhancement methods in the research and development, a microphone array speech simulator (MASS) used in room acoustic environment is proposed. The proposed MASS is the improvement and extension of the existing multi-channel speech simulator. It aims to simulate clean speech, noisy speech, clean speech with reverberation, noisy speech with reverberation, and noise signals by microphone array used for multi-channel coding and enhancement of speech signal in room acoustic environment. The experimental results of the multi-channel speech coding and enhancement prove that the MASS could well simulate the signals used in real room acoustic environment and can be applied to the research of the related fields.

Speech enhancement is a vital technology for reducing the noise in speech communication. Most speech enhancement methods only estimate magnitude spectrum of clean speech from noisy speech and combine noisy phase spectrum to recover the enhanced speech. In this paper, considering the importance of recovering the phase of clean speech in speech enhancement, a phase recovery method of speech is proposed by combining phase unwrapping and deep neural network (DNN). By integrating the recovered phase of clean speech into conventional magnitude enhancement methods, the performance is improved effectively. The verification is conducted by several types of noises at different signal-to-noise ratio (SNR) levels. The experimental results also confirmed that the recovered phase of clean speech resulted in an obvious improvement on the speech quality and intelligibility compared to the noisy phase.

At present, the technology of using microphone arrays for speech enhancement has been widely concerned, and the enhancement effect is excellent. The widely used Generalized Sidelobe Canceller (GSC) method can achieve good noise reduction for noisy speech in the additive noise acoustic environment, and achieve better intelligibility improvement. But there are also areas for improvement. In the lower branch of GSC, signal leakage caused by the estimation of the incident angle or the slight change of the position of the microphone array may cause the self-cancellation of target speech signal, thereby the severe speech distortion is caused. In this paper, the Generative Adversarial Network (GAN), which has broad application prospects in deep learning technology, replaces the lower branch of the traditional GSC structure, thus the self-cancellation of speech signals is avoided and improving the anti-error ability of the enhancement system is improved effectively.

Deep neural network (DNN) has become a popular means for separating target speech from noisy speech in the supervised speech enhancement due to its good performance for learning higher-level information. For DNN-based methods, the training target and acoustic features have a significant impact on the performance of speech restoration. The ideal ratio mask (IRM) is commonly used as the training target. But, generally it does not take into account phase information. The recent studies have revealed that incorporating phase information into the mask can effectively help improve speech quality of the enhanced speech. In this paper, a bounded IRM with phase parameterization is presented and used as the training target of the DNN model. In addition, some acoustic features with harmonic preservation are incorporated into the input of DNN model, which are considered as additional information to improve quality of the enhanced speech. The experiments are performed under various noise environments and signal to noise ratio (SNR) conditions. The results show that the proposed method can outperform reference methods.

In this paper, a multi-channel speech enhancement method with the minimum variance distortionless response (MVDR) beamforming method based on the time-frequency (T-F) masking is proposed. In this study, First, the logarithmic power spectrum (LPS) features of multichannel signals are used as input features to estimate a T-F mask of the reference microphone by the deep neural network (DNN) model. Then, the estimated mask is utilized to calculate speech covariance matrix that is used to estimate a steering vector for constructing the MVDR beamformer. The steering vector is estimated by the generalized eigenvalue decomposition (GEVD) method. Finally, the output speech of the beamformer is processed by the DNN-based IRM model. In order to prove the effectiveness of the proposed method, the perceptual evaluation of speech quality (PESQ) and the segment signal-to-noise ratio (SSNR) are employed. The experimental results show that the proposed method effectively increased the PESQ and SSNR.

Speech enhancement in the noisy and reverberant environment remains a challenging task. Acoustic beamforming algorithm with minimum variance distortionless response (MVDR) has shown to be effective for this case. The crucial issue in MVDR-based speech enhancement is to get accurate estimates of the speech and noise spatial covariance matrices (SCMs). On this way, time-frequency mask-based method which is a reliable method to estimate the SCMs can improve the performance of the MVDR beamformer in speech enhancement. In this paper, an optimal ratio mask-based method used for MVDR beamforming is proposed. Specifically, the convolutional neural networks (CNNs) is used in the proposed method, which operates on the magnitude and phase components of the short-time Fourier transform (STFT) of microphones to estimate the optimal ratio masks, and these masks are used to get the SCMs for constructing MVDR beamformer. Experiments are conducted by using simulated data. The results show that the proposed method is more robust than the reference methods against the terrible acoustic conditions.