Noam R. Shabtai

Generalized Spherical Array Beamforming for Binaural Speech Reproduction

Microphone arrays are used in speech signal processing applications such as teleconferencing and telepresence, in order to enhance a desired speech signal in the presence of speech signals from other speakers, reverberation and background noise. These arrays usually provide a single-channel output, so that no spatial information is available in the output signal. However, spatial information on the sound sources may increase the intelligibility of a speech signal perceived by a human listener. This work presents a mathematical framework for generalized spherical array beamforming that in addition to suppressing noise and reverberation, is aiming to preserve spatial information on the sources in the recording venue. The generalized beamforming, formulated in the spherical harmonics domain, is based on binaural sound reproduction where the head-related transfer functions are incorporated into a headphones presentation. The performance of the proposed generalized beamformer is compared to that of a single-channel output maximum-directivity beamformer. Listening tests with human subjects show that when the generalized beamformer is used the intelligibility is improved at low input SNRs.

Feature selection for room volume identification from room impulse response

The room impulse response (RIR) can be used to calculate many room acoustical parameters, such as the reverberation time (RT). However, estimating the room volume, another important room parameter, from the RIR is typically a more difficult task requiring extraction of other features from the RIR. Most of the existing fully-blind methods for estimating the room volume from the RIR do not combine features from different feature sets. This can be one reason to the fact that these methods are sensitive to differences in source-to-receiver distance and wall reflection coefficients. We propose a new approach in which hypothetical-volume room models are trained with room volume features from different feature sets. Estimation is performed by identifying the hypothesis with maximum-likelihood (ML) using background model normalization. The different feature sets are compared using equal error rate (EER) of hypothesis verification. A combination of features from the different feature sets is selected so that minimum EER is achieved. Using the selected features, we achieve average detection rate of 98.8% with a standard deviation (STD) of 1.5% for eight rooms with different volumes, source-to-receiver distances, and wall reflection coefficients.

Optimization of the directivity in binaural sound reproduction beamforming

Microphone arrays usually combine multiple input signals into one output signal, such that spatial information on the sound sources is not included in the output signal. Since spatial information on the sound sources may increase the intelligibility of a speech signal that is perceived by a human listener, recent works aim to include this spatial information in the output of the microphone array by utilizing binaural cues preservation. More current works apply binaural sound reproduction (BSR) using spherical microphone arrays by incorporating the head related transfer functions (HRTFs) in the weight function of a conventional maximum-directivity beamformer. However, the HRTFs may affect the optimality of beamformers that were already designed to provide a maximal directivity without the HRTFs. This work presents a more general mathematical framework than previously presented for the incorporation of HRTFs in the weight function, which allows the optimization of the weight function using an averaged maximum-directivity criterion under the condition that the HRTFs are already incorporated. It is shown that the proposed optimized BSR beamformer achieves higher directivity index.

The effect of room parameters on speaker verification using reverberant speech

The performance of speaker verification (SVR) systems degrades with the presence of room reverberation. Reverberation results in mismatched conditions between target models and test segments. Reverberation time (RT) is commonly used as a room parameter that represents reverberation. We investigate the effect of other room parameters such as room dimensions and reflection coefficients of the walls on SVR. Equal error rate (EER) is calculated by using room dimensions and reflection coefficients as parameters. Results of SVR with reverberant speech of the same RT are shown to be essentially different, when other room parameters are different. Feature normalization techniques are tested with reverberant speech of the same RT, and shown to be either improving or degrading the performance of SVR when other room parameters are different. This stands in contradiction to the approach in the literature towards RT as a dominant room parameter.

Acoustic centering of sources with high-order radiation patterns

Surrounding spherical microphone arrays have recently been used in order to model the radiation pattern of acoustic sources that are assumed to be at the center of the array. Source centering algorithms are applied to the measurements in order to reduce the negative effect of acoustic source misalignment with regard to the physical center of the microphone array. Recent works aim to minimize the energy that is contained in the high-order coefficients of the radiation pattern in the spherical harmonics domain, in order to directly address the problem of increased order and spatial aliasing resulted by this misalignment. However, objective functions which directly minimize the norm of these coefficients were shown to be convex only when employed on sources with low-order radiation patterns. This work presents a source centering algorithm that operates on plane sections and aims to achieve a convex objective function on every plane section. The results of the proposed algorithm are shown to be more convex than the previous algorithms for sources with higher-order radiation pattern, usually at higher frequencies.

The Effect of GMM Order and CMS on Speaker Recognition with Reverberant Speech

Speaker recognition is used today in a wide range of applications. The presence of reverberation, in hands-free systems for example, results in performance degradation. The effect of reverberation on the feature vectors and its relation to optimal GMM order are investigated. Optimal model order is calculated in terms of minimum BIC and KIC, and tested for EER of a GMM-based speaker recognition system. Experimental results show that for high reverberation time, reducing model order reduces EER values of speaker recognition. The effect of CMS on state of the art GMM and AGMM-based speaker recognition systems is investigated for reverberant speech. Results show that high reverberation time reduces the effectiveness of CMS.

Binaural sound reproduction beamforming using spherical microphone arrays

Currently employed microphone arrays usually have a single-channel output, such that no spatial information can be perceived by a human listener. However, spatial information may trigger spatial mechanisms in the human auditory system which can improve the intelligibility. This work presents a mathematical framework for the binaural beamforming approach for the ideal and order-limited representation in the spherical harmonics domain. The performance of the proposed binaural beamformer is compared to that of a monaural maximum directivity beamformer using objective signal-based measures and subjective listening tests. It is shown that using the binaural beamformer results in higher intelligibility than the monaural beamformer.

Room volume classification from room impulse response using statistical pattern recognition and feature selection

Classification of the room volume from the room impulse response (RIR) can be useful in acoustic scene analysis applications, using RIR that is provided directly, or estimated from audio recordings. Current methods for estimating the room volume from the RIR require the source-to-receiver distance, and may be sensitive to differences in absorption. A room volume classification method is presented that does not require the source-to-receiver distance, and which is potentially robust to differences in absorption. Room volume features are defined that are related to the room volume and may be extracted from the RIR. Gaussian mixture models are trained to model room volume classes. Room volume is classified according to a maximum likelihood criterion that is normalized with a background model. Feature selection is performed with different classification error criteria. Both simulated and measured RIRs were examined, achieving an equal error rate of 0.1% and 19.1%, respectively.

Estimating the room volume from room impulse response via hypothesis verification approach

In order to understand the acoustic behavior of the sound field in a room, it is important to know the volume of the room, as well as other room parameters, like reverberation time (RT). However, estimating the room volume from the room impulse response (RIR) is usually considered a more difficult task than estimating the RT from the RIR. Most of the existing fullyblind methods for estimating the room volume from the RIR do not have satisfactory performances. The reason is that they are sensitive to differences in source-to-receiver distance and wall reflection coefficients. We propose a new approach in which hypothetical-volume room models are trained, and estimation is performed via hypothesis verification using loglikelihood ratio test (LLRT). We achieve low equal error rate (EER) of 3.6% in hypothesis verification for eight rooms with different values of source-to-receiver distance and wall reflection coefficients.

Spherical array beamforming for binaural sound reproduction

In recent years, the employment of microphone arrays has become popular for capturing speech signals of participants in telecommunication systems. In order to achieve improved realism and perceived speaker directions in telepresence applications it may be desired to preserve some of the spatial information of the recorded sound rather than producing a single channel output. As a consequence, spatial hearing mechanisms in the auditory system of a listener may be triggered, which may potentially improve the intelligibility. This work represents a binaural beamforming approach which aims to preserve some of the spatial characteristics of the captured signal, with regard to head related transfer function (HRTF) of the left and right ears. Spherical microphone array configuration is used for the capturing of sound fields. Beam patterns are displayed for the left and right ears, and compared with the beam pattern of a conventional maximum directivity beamformer.