Speech enhancement and separation are core problems in audio signal processing, with commercial applications in devices as diverse as mobile phones, conference call systems, hands-free systems, or hearing aids. In addition, they are crucial preprocessing steps for noise-robust automatic speech and speaker recognition. Many devices now have two to eight microphones. The enhancement and separation capabilities offered by these multichannel interfaces are usually greater than those of single-channel interfaces. Research in speech enhancement and separation has followed two convergent paths, starting with microphone array processing and blind source separation, respectively. These communities are now strongly interrelated and routinely borrow ideas from each other. Yet, a comprehensive overview of the common foundations and the differences between these approaches is lacking at present. In this paper, we propose to fill this gap by analyzing a large number of established and recent techniques according to four transverse axes: 1 the acoustic impulse response model, 2 the spatial filter design criterion, 3 the parameter estimation algorithm, and 4 optional postfiltering. We conclude this overview paper by providing a list of software and data resources and by discussing perspectives and future trends in the field.

/pdf/a-consolidated-perspective-on-multimicrophone-speech-qnbaqxmtvq.pdf

A Consolidated Perspective on Multimicrophone Speech Enhancement and Source Separation

In speech communication systems, such as voice-controlled systems, hands-free mobile telephones, and hearing aids, the received microphone signals are degraded by room reverberation, background noise, and other interferences This signal degradation may lead to total unintelligibility of the speech and decreases the performance of automatic speech recognition systems In the context of this work reverberation is the process of multi-path propagation of an acoustic sound from its source to one or more microphones The received microphone signal generally consists of a direct sound, reflections that arrive shortly after the direct sound (commonly called early reverberation), and reflections that arrive after the early reverberation (commonly called late reverberation) Reverberant speech can be described as sounding distant with noticeable echo and colouration These detrimental perceptual effects are primarily caused by late reverberation, and generally increase with increasing distance between the source and microphone Conversely, early reverberations tend to improve the intelligibility of speech In combination with the direct sound it is sometimes referred to as the early speech component Reduction of the detrimental effects of reflections is evidently of considerable practical importance, and is the focus of this dissertation More specifically the dissertation deals with dereverberation techniques, ie, signal processing techniques to reduce the detrimental effects of reflections In the dissertation, novel single- and multimicrophone speech dereverberation algorithms are developed that aim at the suppression of late reverberation, ie, at estimation of the early speech component This is done via so-called spectral enhancement techniques that require a specific measure of the late reverberant signal This measure, called spectral variance, can be estimated directly from the received (possibly noisy) reverberant signal(s) using a statistical reverberation model and a limited amount of a priori knowledge about the acoustic channel(s) between the source and the microphone(s) In our work an existing single-channel statistical reverberation model serves as a starting point The model is characterized by one parameter that depends on the acoustic characteristics of the environment We show that the spectral variance estimator that is based on this model, can only be used when the source-microphone distance is larger than the so-called critical distance This is, crudely speaking, the distance where the direct sound power is equal to the total reflective power A generalization of the statistical reverberation model in which the direct sound is incorporated is developed This model requires one additional parameter that is related to the ratio between the direct sound energy and the sound energy of all reflections The generalized model is used to derive a novel spectral variance estimator When the novel estimator is used for dereverberation rather than the existing estimator, and the source-microphone distance is smaller than the critical distance, the dereverberation performance is significantly increased Single-microphone systems only exploit the temporal and spectral diversity of the received signal Reverberation, of course, also induces spatial diversity To additionally exploit this diversity, multiple microphones must be used, and their outputs must be combined by a suitable spatial processor such as the so-called delay and sum beamformer It is not a priori evident whether spectral enhancement is best done before or after the spatial processor For this reason we investigate both possibilities, as well as a merge of the spatial processor and the spectral enhancement technique An advantage of the latter option is that the spectral variance estimator can be further improved Our experiments show that the use of multiple microphones affords a significant improvement of the perceptual speech quality The applicability of the theory developed in this dissertation is demonstrated using a hands-free communication system Since hands-free systems are often used in a noisy and reverberant environment, the received microphone signal does not only contain the desired signal but also interferences such as room reverberation that is caused by the desired source, background noise, and a far-end echo signal that results from a sound that is produced by the loudspeaker Usually an acoustic echo canceller is used to cancel the far-end echo Additionally a post-processor is used to suppress background noise and residual echo, ie, echo which could not be cancelled by the echo canceller In this work a novel structure and post-processor for an acoustic echo canceller are developed The post-processor suppresses late reverberation caused by the desired source, residual echo, and background noise The late reverberation and late residual echo are estimated using the generalized statistical reverberation model Experimental results convincingly demonstrate the benefits of the proposed system for suppressing late reverberation, residual echo and background noise The proposed structure and post-processor have a low computational complexity, a highly modular structure, can be seamlessly integrated into existing hands-free communication systems, and affords a significant increase of the listening comfort and speech intelligibility

/pdf/single-and-multi-microphone-speech-dereverberation-using-81pl8h93qv.pdf

Single- and multi-microphone speech dereverberation using spectral enhancement

Acoustic data provide scientific and engineering insights in fields ranging from biology and communications to ocean and Earth science. We survey the recent advances and transformative potential of machine learning (ML), including deep learning, in the field of acoustics. ML is a broad family of techniques, which are often based in statistics, for automatically detecting and utilizing patterns in data. Relative to conventional acoustics and signal processing, ML is data-driven. Given sufficient training data, ML can discover complex relationships between features and desired labels or actions, or between features themselves. With large volumes of training data, ML can discover models describing complex acoustic phenomena such as human speech and reverberation. ML in acoustics is rapidly developing with compelling results and significant future promise. We first introduce ML, then highlight ML developments in four acoustics research areas: source localization in speech processing, source localization in ocean acoustics, bioacoustics, and environmental sounds in everyday scenes.

Machine learning in acoustics: theory and applications

This paper reviews a statistical time-frequency model of late reverberation decays, and describes an associated analysis procedure for deriving the time-frequency envelope of the reverberation from a measured impulse response, based on the notion of

Analysis and Synthesis of Room Reverberation Based on a Statistical Time-Frequency Model

Principles of binaural room simulation

Playing billiards in the concert hall: The mathematical foundations of geometrical room acoustics

On the variability of room acoustical parameters : reproducibility and statistical validity

In Room Acoustics, the quantity that fully describes the hall is a set of room impulse responses (RIRs), which are composed of the succession of arrivals (i.e., some sound rays which have undergone one or more reflections on their way from the source to the receiver). The mixing time is defined as the time it takes for initially adjacent sound rays to spread uniformly across the room. This paper proposes to investigate the temporal distribution of arrivals and the estimation of mixing time. A method based on maxima of correlations (Matching Pursuit) between the source impulse and the RIR allows to estimate in practice arrivals. This paper compares the cumulative distribution function of arrivals of experimental and synthesized RIRs (using a stochastic model). The mixing time is estimated when the arrival density becomes constant. The dependence of mixing time upon the distance source/receiver is investigated with measured and synthesized RIRs.

/pdf/using-matching-pursuit-for-estimating-mixing-time-within-4g5xclsuh3.pdf

J.-D. Polack

Papers

Playing billiards in the concert hall: The mathematical foundations of geometrical room acoustics

On the variability of room acoustical parameters : reproducibility and statistical validity

Using Matching Pursuit for Estimating Mixing Time Within Room Impulse Responses

Estimating the mixing time of concert halls using the eXtensible Fourier Transform