We propose deep neural network (DNN)-based speech enhancement that reduces musical noise and achieves better auditory impressions. The musical noise is an artifact generated by nonlinear signal processing and negatively affects the auditory impressions. We aim to develop musical-noise-free speech enhancement methods that suppress the musical noise generation and produce perceptually-comfortable enhanced speech. DNN-based speech enhancement using a soft mask achieves high noise reduction but generates musical noise in non-speech regions. Therefore, first, we define kurtosis matching for DNN-based low-musical-noise speech enhancement. Kurtosis is the fourth-order moment and is known to correlate with the amount of musical noise. The kurtosis matching is a penalty term of the DNN training and works to reduce the amount of musical noise. We further extend this scheme to standardized-moment matching. The extended scheme involves using moments whose orders are higher than kurtosis and generalizes the conventional musical-noise-free method based on kurtosis matching. We formulate standardized-moment matching and explore how effectively the higher-order moments reduce the amount of musical noise. Experimental evaluation results 1) demonstrate that kurtosis matching can reduce musical noise without negatively affecting noise suppression and 2) newly reveal that the sixth-moment matching also achieves low-musical-noise speech enhancement as well as kurtosis matching.

We propose a structure-generalized blind spatial subtraction array (BSSA), and the theoretical analysis of the amounts of musical noise and speech distortion. The structure of BSSA should be selected according to the application, i.e., a channelwise BSSA is recommended for listening but a conventional BSSA is suitable for speech recognition.

This paper focuses on the underdetermined blind source separation (BSS) of three speech signals mixed in a real environment from measurements provided by two sensors. To date, solutions to the underdetermined BSS problem have mainly been based on the assumption that the speech signals are sufficiently sparse. They involve designing binary masks that extract signals at time-frequency points where only one signal was assumed to exist. The major issue encountered in previous work relates to the occurrence of distortion, which affects a separated signal with loud musical noise. To overcome this problem, we propose combining sparseness with the use of an estimated mixing matrix. First, we use a geometrical approach to detect when only one source is active and to perform a preliminary separation with a time-frequency mask. This information is then used to estimate the mixing matrix, which allows us to improve our separation. Experimental results show that this combination of time-frequency mask and mixing matrix estimation provides separated signals of better quality (less distortion, less musical noise) than those extracted without using the estimated mixing matrix in reverberant conditions where the reverberant time (TR) was 130 ms and 200 ms. Furthermore, informal listening tests clearly show that musical noise is deeply lowered by the proposed method comparatively to the classical approaches.

This paper addresses the single channel speech enhancement method which utilizes the mean value and variance of the logarithmic noise power spectra. An important issue for single channel speech enhancement algorithm is to determine the trade-off point for the spectral distortion and residual noise. Thus the accurate discrimination between speech spectral and noise components is required. The conventional methods determine the trade-off point using parameters obtained experimentally. As a result spectral discrimination is not adequate. And the enhanced speech is deteriorated by spectral distortion or residual noise. Therefore, a criteria to determine the point is necessary. The proposed method determines the trade-off point of spectral distortion and residual noise level by discrimination between speech spectral and noise components based on statistical criteria. The spectral discrimination is performed using hypothesis testing that utilizes means and variances of the logarithmic power spectra. The discriminated spectral components are divided into speech-dominant spectral components and noise-dominant ones. For the speech-dominant ones, spectral subtraction is performed to minimize the spectral distortion. For the noise-dominant ones, attenuation is performed to reduce the noise level. The performance of the method is confirmed in terms of waveform, spectrogram, noise reduction level and speech recognition task. As a result, the noise reduction level and speech recognition rate are improved so that the method reduces the musical noise effectively and improves the enhanced speech quality.