Catálogo de publicaciones - libros

Although a number of dereverberation methods have been studied, dereverberation is still a challenging problem especially when using a single microphone. An important aspect of single-channel speech enhancement is the characteristic feature of speech signals that allows us to restore the quality of source signals. In this chapter, we describe a single-channel speech dereverberation method based on the harmonicity of speech signals. We show that a filter that enhances the harmonic structure of reverberant speech signals approximates the inverse filter of the reverberation process, thus enabling us to achieve high quality blind dereverberation. The presented method is referred to as the harmonicity based dereverberation method, HERB. Simulation experiments show that HERB can work effectively to dereverberate speech signals in terms of energy decay curves of room impulse responses and automatic speech recognition performance even when the reverberation time is as long as 1.0 sec provided a sufficiently large number of observed signals are available. Further discussions on several future directions are also provided with a view to extending HERB so that it can cope with more realistic situations.

Speech enhancement was not and should not be examined solely with the tool of time-frequency analysis. Approaching this problem from different perspectives or incorporating other knowledges helps to expand the number of options open to us when developing a speech enhancement system. Using multiple microphones at different locations makes it possible to develop more sophisticated source separation and dereverberation technologies for speech enhancement, which enable man-made systems to extract a speech signal of interest in a noisy environment with competing speech and/or noise sources. This phenomenon is referred to as the cocktail party effect demonstrated by human beings and many other creatures with few efforts. However, separating and dereverberating speech signals is a very difficult problem in reverberant environments and the state-of-the-art algorithms are still unsatisfactory. The challenge lies in the coexistence of spatial interference from competing sources and temporal echoes due to room reverberation in the observed microphone signals. Focusing only on optimizing the signal-to-interference ratio is inadequate for most speech processing systems where source separation and speech dereverberation are two fully-integrated problems. In this chapter, we study these two problems in a unified framework. We deduce that spatial interference and temporal reverberation can be separated and a SIMO system with the speech signal of interest as input is extracted from the MIMO system. Furthermore, this interference-free SIMO system is dereverberated using the MINT theorem. Such a two-stage procedure leads to a novel sequential source separation and speech dereverberation algorithm based on blind multichannel identification. Simulations with measurements obtained in the varechoic chamber at Bell Labs verified the proposed algorithm.

This chapter discusses the frequency-domain approach to the blind source separation (BSS) of convolutively mixed acoustic signals. In this approach, independent component analysis (ICA) is employed in each frequency bin to calculate the frequency responses of separation filters. Since convolutive mixtures in the time domain can be approximated as multiple instantaneous mixtures in the frequency domain, the advantage of this approach is that ICA is applied just for instantaneous mixtures, which is very simple. However, the permutation ambiguity of ICA solutions then becomes a problem. This chapter mainly deals with a method for solving the permutation problem. The method utilizes the source location information that can be estimated from the ICA solutions. We also discuss other important topics for frequency-domain BSS, such as complex-valued ICA, scaling alignment and spectral smoothing. To show the effectiveness of this frequency-domain approach, we report experimental results for separating up to four sources with a 4-element linear array, and also six sources with an 8-element planar array.

In this chapter, we address subband-based blind source separation (BSS) for convolutive mixtures of speech by reporting a large number of experimental results. The subband-based BSS approach offers a compromise between time-domain and frequency-domain techniques. The former is usually difficult and slow with many separation filter coefficients to estimate. With the latter it is difficult to estimate statistics when the adaptation data length is insufficient. With subband-based BSS, a sufficient number of samples for estimating statistics can be held in each subband by using a moderate number of subbands. Moreover, by using FIR filters in each subband, which are shorter than the filters used for time-domain BSS, we can handle long reverberation. In addition, subband-based BSS allows us to select the separation method suited to each subband. Using this advantage, we introduce efficient separation procedures that take both the frequency characteristics of the room reverberation and speech signals into consideration. In concrete terms, longer separation filters and an overlap-blockshift in BSS’s batch adaptation in low frequency bands improve the separation performance. Consequently, frequency-dependent subband processing is successfully realized with subband-based BSS.

In this chapter, we present a method for the real-time blind source separation (BSS) of moving speech signals in a room. The method employs frequen-cydomain independent component analysis (ICA) using a blockwise batch algorithm in the first stage, and the separated signals are refined by postprocessing using crosstalk component estimation and non-stationary crosstalk cancellation in the second stage. The blockwise batch algorithm achieves better performance than an online algorithm when sources are stationary, and the postprocessing compensates for performance degradation caused by source movement. Experimental results using speech signals recorded in a real room show that our method realizes robust real-time separation for moving sources.

The term auditory scene analysis (ASA) refers to the ability of human listeners to form perceptual representations of the constituent sources in an acoustic mixture, as in the well-known ‘cocktail party’ effect. Accordingly, computational auditory scene analysis (CASA) is the field of study which attempts to replicate ASA in machines. Some CASA systems are closely modelled on the known stages of auditory processing, whereas others adopt a more functional approach. However, all are broadly based on the principles underlying the perception and organization of sound by human listeners, and in this respect they differ from ICA and other approaches to sound separation. In this chapter, we review the principles underlying ASA and show how they can be implemented in CASA systems. We also consider the link between CASA and automatic speech recognition, and draw distinctions between the CASA and ICA approaches.