Communicated by Michael Lewicki
The Cocktail Party Problem
Adaptive Systems Lab, McMaster University, Hamilton, Ontario, Canada L8S 4K1
This review presents an overview of a challenging problem in auditory
perception, the cocktail party phenomenon, the delineation of which goes
back to a classic paper by Cherry in 1953. In this review, we address the
following issues: (1) human auditory scene analysis, which is a general
process carried out by the auditory system of a human listener; (2) in-
sight into auditory perception, which is derived from Marr’s vision the-
ory; (3) computational auditory scene analysis, which focuses on specific
approaches aimed at solving the machine cocktail party problem; (4) ac-
tive audition, the proposal for which is motivated by analogy with active
vision, and (5) discussion of brain theory and independent component
analysis, on the one hand, and correlative neural firing, on the other.
One of our most important faculties is our ability to listen to, and follow,one speaker in the presence of others. This is such a common experiencethat we may take it for granted; we may call it “the cocktail party prob-lem.” No machine has been constructed to do just this, to filter out oneconversation from a number jumbled together.
The cocktail party problem
(CPP), first proposed by Colin Cherry, is a psychoa-coustic phenomenon that refers to the remarkable human ability to selec-tively attend to and recognize one source of auditory input in a noisy envi-ronment, where the hearing interference is produced by competing speechsounds or a variety of noises that are often assumed to be independentof each other (Cherry, 1953). Following the early pioneering work (Cherry,1953, 1957, 1961; Cherry & Taylor, 1954; Cherry & Sayers, 1956, 1959; Sayers& Cherry, 1957), numerous efforts have been dedicated to the CPP in diversefields: physiology, neurobiology, psychophysiology, cognitive psychology,biophysics, computer science, and engineering. Due to its multidisciplinarynature, it is almost impossible to completely cover this problem in a single
Neural Computation 17, 1875–1902
2005 Massachusetts Institute of Technology
article.1 Some early partial treatment and reviews of this problem are foundin different disciplinary publications (Bregman, 1990; Arons, 1992; Wood& Cowan, 1995; Yost, 1997; Feng & Ratnam, 2000; Bronkhorst, 2000). Halfa century after Cherry’s seminal work, however, it seems fair to say that acomplete understanding of the cocktail party phenomenon is still missing,and the story is far from being complete; the enigma about the marvelous au-ditory perception capability of human beings remains a mystery. To unveilthe mystery and imitate the human performance with a machine, compu-tational neuroscientists, computer scientists, and engineers have attemptedto view and simplify this complex perceptual task as a learning problem, forwhich a tractable computational solution is sought. Despite their obvioussimplicity and distinction from reality, the efforts seeking the computationalsolutions to imitate a human’s unbeatable audition capability have revealedthat we require a deep understanding of the human auditory system andthe underlying neural mechanisms. Bearing such a goal in mind, it doesnot mean that we must duplicate every aspect of the human auditory sys-tem in solving the machine cocktail party problem. Rather, it is our beliefthat seeking the ultimate answer to the CPP requires deep understandingof many fundamental issues that are deemed to be of theoretical and tech-nical importance. In addition to its obvious theoretical values in differentdisciplines, the tackling of the CPP will certainly be beneficial to ongoingresearch on human-machine interfaces.
There are three fundamental questions pertaining to CPP:
3. Is it possible to build a machine capable of solving it in a satisfactory
The first two questions are human oriented and mainly involve the disci-plines of neuroscience, cognitive psychology, and psychoacoustics; the lastquestion is rooted in machine learning, which involves computer scienceand engineering disciplines.
In addressing the cocktail party problem, we are interested in three un-
gation, which refers to the segmentation of an incoming auditory signalto individual channels or streams.3 Among the heuristics used by a
1A recently edited volume by Divenyi (2004) discusses several aspects of the cocktail
party problem that are complementary to the material covered in this review.
2 Categorization of these three neural processes, done essentially for research-related
studies, is somewhat artificial; the boundary between them is fuzzy in that the brain doesnot necessarily distinguish between them as defined here.
3 For an early discussion of the segmentation process, see Moray (1959).
listener to do the segmentation, spatial location is perhaps the mostimportant. Specifically, sounds coming from the same location aregrouped together, while sounds originating from other different di-rections are segregated.
The recognition process involves analyzing the statisti-
cal structure of the patterns contained in a sound stream that are help-ful in recognizing the patterns. The goal of recognition is to uncoverthe neurobiological mechanisms through which humans are able toidentify a segregated sound from multiple streams with relative ease.
The synthesis process involves the reconstruction of indi-
vidual sound waveforms from the separated sound streams. Whilesynthesis is an important process carried out in the brain (Warren,1970; Warren, Obusek, & Ackroff, 1972; Bregman, 1990), the synthesisproblem is primarily of interest to the machine CPP.
Note also that recognition does not require the analysis process to be per-fect; and by the same token, an accurate synthesis does not necessarily meanhaving solved the analysis and recognition problems, although extra infor-mation might provide more hints for the synthesis process.
From an engineering viewpoint, in a loose sense, synthesis may be re-
garded as the inverse of the combination of analysis and recognition in thatit attempts to uncover the relevant attributes of the speech production mech-anism. The aim of synthesis is to build a machine that offers the capabilitiesof operating on a convolved mixture of multiple sources of sounds and tofocus attention on the extraction from the convolved mixture a stream ofsounds that is of particular interest to an observer; the convolution men-tioned here refers to reverberation in a confined environment, which is ahallmark of real-life cocktail party phenomena.
The main theme of this review4 is philosophical and didactic; hence,
no detailed mathematical analysis is presented. The rest of the review isorganized as follows. Section 2 discusses human auditory scene analysis.
Section 3 discusses the impact of Marr’s work in vision (Marr, 1982) onauditory scene analysis (Bregman, 1990). Section 4 presents an overviewof computational approaches for solving the cocktail party problem, withan emphasis on independent component analysis, temporal binding andoscillatory correlation, and cortronic network. Section 5 discusses activeaudition. Discussion of some basic issues pertaining to the cocktail partyproblem in section 6 concludes the review.
4 The early version of this review appeared as a presentation made by the first au-
thor (Haykin, 2003) and a more lengthy technical report by the second author (Chen,2003). In this latter report, we presented a comprehensive overview of the cocktail partyproblem, including a historical account, auditory perceptual processes, relations to visualperception, and detailed descriptions of various related computational approaches.
2 Human Auditory Scene Analysis
Human auditory scene analysis (ASA) is a general process carried out bythe auditory system of a human listener for the purpose of extracting in-formation pertaining to a sound source of interest, which is embedded in abackground of noise interference.
The auditory system is made up of two ears (constituting the organs of
hearing) and auditory pathways. In more specific terms, it is a sophisti-cated information processing system that enables us to detect not only thefrequency composition of an incoming sound but also to locate the soundsources (Kandel, Schwartz, & Jessell, 2000). This is all the more remarkable,given the fact that the energy in the incoming sound waves is exceedinglysmall and the frequency composition of most sounds is rather complicated.
In this review, our primary concern is with the cocktail party problem. In
this context, a complete understanding of the hearing process must includea delineation of where the sounds are located, what sounds are perceived,as well as an explanation of how their perception is accomplished.
2.1 “Where” and “What.”
The mechanisms in auditory perception es-
sentially involve two processes: sound localization (“where”) and soundrecognition (“what”). It is well known that (e.g., Blauert, 1983; Yost & Goure-vitch, 1987; Yost, 2000) for localizing sound sources in the azimuthal plane,interaural time difference is the main acoustic cue for sound location at lowfrequencies, and for complex stimuli with low-frequency repetition, interau-ral level is the main cue for sound localization at high frequencies. Spectraldifferences provided by the head-related transfer function (HRTF) are themain cues used for vertical localization. Loudness (intensity) and early re-flections are the probable cues for localization as a function of distance. Inhearing, the precedence effect refers to the phenomenon that occurs duringauditory fusion when two sounds of the same order of magnitude are pre-sented dichotically and produce localization of the secondary sound wavestoward the outer ear receiving the first sound stimulus (Yost, 2000); theprecedence effect stresses the importance of the first wave in determiningthe sound location.
The “what” question mainly addresses the processes of sound segrega-
tion (streaming) and sound determination (identification). Although it hasa critical role in sound localization, spatial separation is not considered astrong acoustic cue for streaming or segregation (Bregman, 1990). Accord-ing to Bregman’s studies, sound segregation consists of a two-stage process:feature selection and feature grouping. Feature selection invokes processingthe auditory stimuli into a collection of favorable (e.g., frequency sensitive,pitch-related, temporal-spectral-like) features. Feature grouping is responsi-ble for combining similar elements of incoming sounds according to certainprinciples into one or more coherent streams, with each stream correspond-ing to one informative sound source. Sound determination is more specific
than segregation in that it not only involves segmentation of the incomingsound into different streams, but also identifies the content of the soundsource in question. We will revisit Bregman’s viewpoint of human auditoryscene analysis in section 3.
2.2 Spatial Hearing.
From a communication perspective, our two outer
ears act as receive-antennae for acoustic signals from a speaker or audiosource. In the presence of one (or fewer) competing or masking soundsource, the human ability to detect and understand the source of interest (i.e.,target) is degraded. However, the influence of the masking source generallydecreases when the target and masker are spatially separated, compared towhen the target and masker are in the same location; this effect is creditedto spatial hearing (filtering).
In his classic paper, Cherry (1953) suggested that spatial hearing plays
a major role in the auditory system’s ability to separate sound sources in amultiple-source acoustic environment. Many subsequent experiments haveverified Cherry’s conjecture. On the other hand, spatial hearing is viewedas one of the important cues that are exploited in solving the CPP (Yost &Gourevitch, 1987) and enhancing speech intelligibility (Hawley, Litovsky,& Colburn, 1999). Spatial separation of the sound sources is also believedto be more beneficial to localization than segregation (Bregman, 1990). Butin some cases, spatial hearing is crucial to the sound determination task(Yost, 1991, 1997). Specifically, spatial unmasking produces three effects:(1) pure acoustic effects due to the way sound impinges on the listener’shead and body, (2) binaural processing that improves the target signal-to-masker ratio, and (3) central attention whereby the listener can selectivelyfocus attention on a source at a particular direction and block out the com-peting sources in the unattended directions. The classic book by Blauert(1983) presents a comprehensive treatment of the psychophysical aspect ofhuman sound localization. Given multiple sound sources in an enclosedspace (such as a conference room), spatial hearing helps the brain to takefull advantage of the slight difference (timing, intensity) between the sig-nals that reach the two outer ears to perform monaural (autocorrelation)and binaural (cross-correlation) processing for specific tasks (such as co-incidence detection, precedence detection, localization, and fusion), basedon which auditory events are identified and followed by higher-level audi-tory processing (e.g., attention, streaming, cognition). Figure 1 illustrates afunctional diagram of the binaural spatial hearing process.
2.3 Binaural Processing.
One of the key observations derived from
Cherry’s classic experiment (Cherry, 1953) was that it is easier to separatethe sources heard binaurally than when they are heard monaurally. Quot-ing from Cherry and Taylor (1954): “One of the most striking facts aboutour ears is that we have two of them—and yet we hear one acoustic world;only one voice per speaker.” We believe that nature gives us two ears for a
formation of the
Figure 1: Functional diagram of binaural hearing, which consists of physical,psychophysical, and psychological aspects of auditory perception. (Adaptedfrom Blauert, 1983, with permission.)
reason just like it gives us two eyes. It is the binocular vision (stereovision)and binaural hearing (stereausis) that enable us to perceive the dynamicouter world and provide the main sensory information sources. Binocu-lar/binaural processing is considered to be crucial in certain perceptualactivities (e.g., binocular/binaural fusion, depth perception, localization).5Given one sound source, the two ears receive slightly different sound pat-terns due to a finite delay produced by their physically separated locations.
The brain is known to be extremely efficient in using varieties of acousticcues, such as interaural time difference (ITD), interaural intensity difference
5 We can view sound localization as binaural depth perception, representing the coun-
terpart to binocular depth perception in vision.
(IID), and interaural phase difference (IPD), to perform specific auditiontasks. The slight differences in these cues are sufficient to identify the loca-tion and direction of the incoming sound waves.
An influential binaural phenomenon is the so-called binaural masking
(e.g., Durlach & Colburn, 1978; Moore, 1997; Yost, 2000). The threshold ofdetecting a signal masked in noise can sometimes be lower when listeningwith two ears compared to listening with only one, which is demonstratedby a phenomenon called binaural masking level difference (BMLD). It isknown (Yost, 2000) that the masked threshold of a signal is the same whenthe stimuli are presented in a monotic or diotic condition; when the maskerand the signal are presented in a dichotic situation, the signal has a lowerthreshold than in either monotic or diotic conditions. Similarly, many exper-iments have also verified that binaural hearing increases speech intelligi-bility when the speech signal and noise are presented dichotically. Anotherimportant binaural phenomenon is binaural fusion. Fusion is the essence ofdirectional hearing; the fusion mechanism is often modeled as performingsome kind of correlation analysis (Cherry & Sayers, 1956; Cherry, 1961), inwhich a binaural fusion model based on the autocorrelogram and cross-correlogram was proposed (see Figure 1).
Another function basic to human auditory analysis is
that of attention, which is a dynamic cognitive process. According to James(1890), the effects of attention include five types of cognitive behavior:(1) perceive, (2) conceive, (3) distinguish, (4) remember, and (5) shortenthe reaction time of perceiving and conceiving.
In the context of the cocktail party problem, attention pertains to the
ability of a listener to focus attention on one channel while ignoring otherirrelevant channels. In particular, two kinds of attention processes are ofteninvolved in the cocktail party phenomenon (Jones & Yee, 1993; Yost, 1997):
r Selective attention, in which the listener attends to one particular
Once the attention mechanism (either selective or divided) is initiated, ahuman subject is capable of maintaining attention for a short period oftime, hence the term maintained attention.
Another intrinsic mechanism relating to attention is switched attention,
which involves the ability of the human brain to switch attention from onechannel to another; switched attention is probably mediated in a top-downmanner by “gating” the incoming auditory signal (Wood & Cowan, 1995).
In this context, a matter of particular interest is the fact that unlike thevisual system, whose cortical top-down feedback goes only as far down asthe thalamus, the cortical feedback in the auditory system exerts its effect
Figure 2: Schematic diagram of an auditory attention circuit.
all the way down to the outer hair cells in the cochlea via the midbrainstructure (Wood & Cowan, 1995). Accordingly, the potential for the earlyselection process of a speech signal of interest in audition is large. Figure 2is a schematic diagram of the auditory attention circuit. As depicted in thefigure, the thalamus acts mainly as a relay station between the sensory haircells and the auditory cortex.6
The bottom-up signals received from the hair cells are sent to medial
geniculate nuclei (MGN) in the thalamus and farther up to the auditory cor-tex through the thalamocortical pathways. The top-down signals from thecortex are sent back to the hair cells through the corticothalamic pathways,to reinforce the signal stream of interest and maximize expectation throughfeedback.
In addition to auditory scene inputs, visual scene inputs are believed to
influence the attention mechanism (Jones & Yee, 1993). For instance, lip-reading is known to be beneficial to speech perception. The beneficial effectis made possible by virtue of the fact that the attention circuit also encom-passes cortico-cortical loops between the auditory and visual cortices.
2.5 Feature Binding.
One other important cognitive process involved
in the cocktail party phenomenon is that of feature binding, which refers tothe problem of representing conjunctions of features. According to von der
6 This is also consistent with the postulate of visual attention mechanism (Crick, 1984;
Malsburg (1999), binding is a very general problem that applies to all typesof knowledge representations, from the most basic perceptual representa-tion to the most complex cognitive representation. Feature binding may beeither static or dynamic. Static feature binding involves a representationalunit that stands for a specific conjunction of properties, whereas dynamicfeature binding involves conjunctions of properties as the binding of unitsin the representation of an auditory scene, an idea traced back to Treismanand Gelade (1980). The most popular dynamic binding mechanism is basedon temporal synchrony, hence the reference to it as “temporal binding.”K ¨onig, Engel, and Singer (1996) suggested that synchronous firing of neu-rons plays an important role in information processing within the cortex.
Rather than being a temporal integrator, the cortical neurons might serveas a coincidence detector evidenced by numerous physiological findings.7
Dynamic binding is closely related to the attention mechanism, which
is used to control the synchronized activities of different assemblies ofunits and how the finite binding resource is allocated among the assemblies(Singer, 1993, 1995). Experimental evidence (especially in vision) has shownthat synchronized firing tends to provide the attended stimulus with an en-hanced representation. Temporal binding hypothesis is attractive, thoughnot fully convincing, in interpreting the perceptual (Gestalt) grouping andsensory segmentation, which has also been evidenced by numerous neuro-physiological data (Engel et al., 1991; von der Malsburg, 1999; see also thebibliographies in both works in the special issue on the binding problem).
2.6 Psychophysical and Psychoacoustic Perspectives.
2.6.1 Psychophysical Attributes and Cues.
The psychophysical attributes of
sound mainly involve three forms of information: spatial location, temporalstructure, and spectral characterization. The perception of a sound signal in acocktail party environment is uniquely determined by this kind of collectiveinformation; any difference in any of the three forms of information is be-lieved to be sufficient to discriminate two different sound sources. In soundperception, many acoustic features (cues) are used to perform specific tasks.
Table 1 summarizes the main acoustic features (i.e., the temporal or spectralpatterns) used for a single-stream sound perception. Combination of a fewor more of those acoustic cues is the key to conducting auditory scene analy-sis. Psychophysical evidence also suggests that significant cues may be pro-vided by spectral-temporal correlations (Feng & Ratnam, 2000). It should benoted that the perception ability with respect to different sound objects (e.g.,speech or music) may be different. The fundamental frequencies or tonesof sounds are also crucial to perception sensitivity. Experimental results
7 For detailed discussions of the coincidence detector, see the review papers (Singer,
1993; K ¨onig & Engel, 1995; K ¨onig, Engel, & Singer, 1996).
Table 1: The Features (Cues) Used in Sound Perception.
have confirmed that difficulties occur more often in the presence of compet-ing speech signals than in the presence of a single speech and other acousticsources.
2.6.2 Room Acoustics.
For auditory scene analysis, studying the effect of
room acoustics on the cocktail party environment is important (Sabine, 1953;MacLean, 1959; Blauert, 1983). A conversation occurring in a closed roomoften suffers from the multipath effect—mainly echoes and reverberation,which are almost ubiquitous but are rarely consciously noticed. Accordingto the acoustics of the room, a reflection from one surface (e.g., wall, ground)produces reverberation. In the time domain, the reflection manifests itselfas smaller, delayed replicas (echoes) that are added to the original sound;in the frequency domain, the reflection introduces a comb-filter effect intothe frequency response. When the room is large, echoes can sometimes beconsciously heard. However, the human auditory system is so powerful thatit can take advantage of binaural and spatial hearing to efficiently suppressthe echo, thereby improving the hearing performance.
The acoustic cues listed in Table 1 that are spatially dependent, such as
ITD and IID, are naturally affected by reverberation. The acoustic cues thatare space invariant, such as common onset across frequencies and pitch,are less sensitive to reverberation. On this basis, it is conjectured that theauditory system has the ability to weight the different acoustic cues (priorto their fusion) so as to deal with a reverberant environment in an effectivemanner.
3 Insight from Marr’s Vision Theory
Audition and vision, the most influential perception processes in the humanbrain, enable us to absorb the cyclopean information of the outer world inour daily lives. It is well known to neuroscientists that audition (hearing) and
vision (seeing) share substantial common features in the sensory processingprinciples as well as anatomical and functional organizations in higher-levelcenters in the cortex.
In his landmark book, David Marr first presented three levels of analysis
of information processing systems (Marr, 1982, p. 25):
r Computation: What is the goal of the computation, why is it appro-
priate, and what is the logic of the strategy by which it can be carriedout?
r Representation: How can this computational theory be implemented?
In particular, what is the representation for the input and output, andwhat is the algorithm for the transformation?
r Implementation: How can the representation and algorithm be real-
In many perspectives, Marr’s observations highlight the fundamental ques-tions that need to be addressed in computational neuroscience, not only invision but also audition. As a matter of fact, Marr’s theory has providedmany insights into auditory research (Bregman, 1990; Rosenthal & Okuno,1998).
In a similar vein to visual scene analysis (e.g., Julesz & Hirsh, 1972),
auditory scene analysis (Bregman, 1990) attempts to identify the content(“what”) and the location (“where”) of the sounds and speech in the envi-ronment. In specific terms, auditory scene analysis consists of two stages.
In the first stage, the segmentation process decomposes a complex acous-tic scene into a collection of distinct sensory elements; in the second stage,the grouping process combines these elements into a stream according tosome principles; the streams can be interpreted by a higher-level processfor recognition and scene understanding. Motivated by Gestalt psychology,Bregman (1990) has proposed five grouping principles for auditory sceneanalysis:
1. Proximity, which characterizes the distances between the auditory
features with respect to their onsets, pitch, and intensity (loudness)
2. Similarity, which usually depends on the properties of a sound signal,
3. Continuity, which features the smoothly varying spectra of a time-
4. Closure, which completes fragmentary features that have a good
Gestalt; the completion can be understood as an auditory compen-sation for masking
5. Common fate, which groups together activities (onset, glides, or vi-
Moreover, Bregman (1990) has distinguished at least two levels of audi-tory organization: primitive streaming and schema-based segregation, withschemas provided by the collected phonetic, prosodic, syntactic, and seman-tic information. While being applicable to general sound (speech and music)scene analysis, Bregman’s work focused mainly on primitive stream segre-gation. As discussed earlier, auditory scene analysis attempts to solve theanalysis and recognition aspects of the CPP.
4 Computational Auditory Scene Analysis
Computational auditory scene analysis (CASA) relies on the developmentof a computational model of the auditory scene with one of two goals inmind, depending on the application of interest:
1. The design of a machine, which by itself is able to automatically extract
and track a sound signal of interest in a cocktail party environment
2. The design of an adaptive hearing system, which automatically com-
putes the perceptual grouping process missing from the auditorysystem of a hearing-impaired individual, thereby enabling that in-dividual to attend to a sound signal of interest in a cocktail partyenvironment.
Naturally, CASA is motivated by or builds on the understanding we haveof human auditory scene analysis.
Following Bregman’s seminal work, a number of researchers (Cooke,
1993; Brown, 1992; Ellis, 1996; Cooke & Brown, 1993; Brown & Cooke, 1994;Cooke & Ellis, 2001) have exploited the CASA.8 In the literature, there aretwo representative kinds of CASA systems: data-driven system (Cooke,1993) and prediction-driven system (Ellis, 1996). The common feature inthese two systems is to integrate low-level (bottom-up, primitive) acousticcues for potential grouping. The main differences between them are (Cooke,2002) that data-driven CASA aims to decompose the auditory scene intotime-frequency elements (so-called strands), and then runs the groupingprocedure, while prediction-driven CASA regards prediction as the pri-mary goal. It requires only a world model that is consistent with the stim-ulus; it contains integration of top-down and bottom-up cues and can dealwith incomplete or masked data (i.e., speech signal with missing informa-tion). However, as emphasized by Bregman (1998), it is important for CASAmodelers to take into account psychological data as well as the way humanscarry out auditory scene analysis (ASA). For instance, to model the stabilityof human ASA, the computational system must allow different cues to col-laborate and compete and must account for the propagation of constraintsacross the frequency-by-time field. It is noteworthy that the performances
8 For review, see Rosenthal and Okuno (1998) and Cooke and Ellis, (2001).
of the CASA approaches are quite dependent on the conditions of noise orinterference (such as the spatial location and overlapping time-frequencymap), which may therefore be a practical limitation for solving a machinecocktail party problem.
In what follows, we restrict our attention on three major categorized
computational approaches aimed at solving the cocktail party problem:9(1) blind source separation (BSS) and independent component analysis(ICA), (2) temporal binding and oscillatory correlation, and (3) cortronicnetwork. The first approach has gained a great deal of popularity in theliterature, and there is no doubt it will continue to play an important rolein neuroscience and signal-processing research; however, the basic ICA ap-proach is limited by its assumptions, and it is arguably biologically implau-sible in the context of CPP. The second approach is biologically inspiredand potentially powerful. The third approach is biologically motivated andknowledge based, and it is configured to solve a realistic machine CPP in areal-life environment.
4.1 Independent Component Analysis and Blind Source Separation.
The essence of independent component analysis (ICA) can be stated asfollows: given an instantaneous linear mixture of signals produced by aset of sources, devise an algorithm that exploits a statistical discriminant todifferentiate these sources so as to provide for the separation of the sourcesignals in a blind (i.e., unsupervised) manner. From this statement, it is ap-parent that ICA theory and the task of blind source separation (BSS) areintrinsically related.
The earliest reference to this signal processing problem is the article by
Jutten and H´erault (1991), which was motivated by Hebb’s postulate oflearning (1949). This was followed by Comon’s article (1994) and that of Belland Sejnowski (1995). Comon used some signal processing and information-theoretic ideas to formulate a mathematical framework for instantaneouslinear mixing of independent source signals, in the course of which the no-tion of nongaussian ICA was clearly defined. Bell and Sejnowski developeda simple algorithm (Infomax) for BSS, which is inspired by Hebb’s postulateof learning and the maximum entropy principle.
It is well known that if we are to achieve the blind separation of an instan-
taneous linear mixture of independent source signals, then there must be a
9 It is noteworthy that our overview is by no means exclusive. In addition to the
three approaches being reviewed here, several other approaches, some of them quitepromising, have been discussed in the literature: Bayesian approaches (e.g., Knuth, 1999;Mohammad-Djafari, 1999; Rowe, 2002; Attias, 1999; Chan, Lee, & Sejnowski, 2003), time-frequency analysis approaches (e.g., Belouchrani & Amin, 1998; Rickard, Balan, & Rosca,2001; Rickard & Yilmaz, 2002; Yilmaz & Rickard, 2004), and neural network approaches(e.g., Amari & Cichocki, 1998; Grossberg, Govindarajan, Wyse, & Cohen, 2004). Due tospace limitation, we have not included these approaches in this article; the interestedreader is referred to Chen (2003) for a more detailed overview.
characteristic departure from the simplest possible source model: an inde-pendently and identically distributed (i.i.d.) gaussian model. The departurecan arise in three different ways, depending on which of the characteristicassumptions embodied in this simple source model is broken, as summa-rized here (Cardoso, 2001):
r Nongaussian i.i.d. model. In this route to BSS, the i.i.d. assumption
for the source signals is retained but the gaussian assumption is aban-doned for all the sources, except possibly for one of them. The In-fomax algorithm due to Bell and Sejnowski (1995), the natural gra-dient algorithm due to Amari, Cichocki, and Yang (1996), Cardoso’sJADE algorithm (Cardoso & Souloumiac, 1993; Cardoso, 1998), and theFastICA algorithm due to Hyv¨arinen and Oja (1997) are all based on thenongaussian i.i.d. model. These algorithms differ from each other inthe way in which source information residing in higher-order statisticsis exploited.
r Gaussian nonstationary model. In this second route to BSS, the gaus-
sian assumption is retained for all the sources, which means thatsecond-order statistics (i.e., mean and variance) are sufficient for char-acterizing each source signal. Blind source separation is achieved byexploiting the property of nonstationarity, provided that the sourcesignals differ from each other in the ways in which their statisticsvary with time. This approach to BSS was first described by Parra andSpence (2000) and Pham and Cardoso (2001). Whereas the algorithmsfocusing on the nongaussian i.i.d. model operate in the time domain,the algorithms that belong to the gaussian nonstationary model oper-ate in the frequency domain, a feature that also makes it possible forthese latter ICA algorithms to work with convolutive mixtures.
r Gaussian, stationary correlated-in-time model. In this third and final
route to BSS, the blind separation of gaussian stationary source signalsis achieved on the proviso that their power spectra are not proportionalto each other. Recognizing that the power spectrum of a wide-sensestationary random process is related to the autocorrelation functionvia the Wiener-Khintchine theorem, spectral differences among thesource signals translate to corresponding differences in correlated-in-time behavior of the source signals. It is this latter property that isavailable for exploitation. ICA algorithms that belong to this third classinclude those due to Tong, Soon, Huang, and Liu (1990), Belouchrani,Abed-Meraim, Cardoso, and Moulines (1997), Amari (2000), and Pham(2002).
In Cardoso (2001, 2003), Amari’s information geometry is used to explorea unified framework for the objective functions that pertain to these threeroutes to BSS.
It is right and proper to say that in their own individual ways, Comon’s
1994 article and the 1995 article by Bell and Sejnowski, have been thecatalysts for the literature in ICA theory, algorithms, and novel applica-tions.10 Indeed, the literature is so extensive and diverse that in the courseof ten years, ICA has established itself as an indispensable part of the ever-expanding discipline of statistical signal processing, and has had a greatimpact on neuroscience (Brown, Yamada, & Sejnowski, 2001). However, in-sofar as auditory phenomena are concerned, ICA algorithms do not exploitthe merit of spatial hearing; this limitation may be alleviated by adding aspatial filter in the form of an adaptive beamformer (using an array of mi-crophones) as the front-end processor to an ICA algorithm (e.g., Parra &Alvino, 2002). Most important, in the context of the cocktail party problemthat is of specific interest to this review, we may pose the following ques-tion: Given the ability of the ICA algorithm to solve the BSS problem, can italso solve the cocktail party problem? Our short answer to this fundamentalquestion is no; the rationale is discussed in section 6.
4.2 Temporal Binding and Oscillatory Correlation.
theory was most elegantly illustrated by von der Malsburg (1981) in hisseminal technical report, Correlation Theory of Brain Function, in whichhe suggested that the binding mechanism is accomplished by the correla-tion correspondence between presynaptic and postsynaptic activities, andthe strengths of synapses follow the Hebbian postulate of learning. Whenthe synchrony between the presynaptic and postsynaptic neurons is strong(weak), the strength would correspondingly increase (decrease) temporally.
Such a synapse was referred to as the “Malsburg synapse” by Crick (1984).
The synchronized mechanism allows the neurons to be linked in multipleactive groups simultaneously and form a topological network. Moreover,von der Malsburg (1981) suggested a dynamic link architecture to solvethe temporal binding problem by letting neural signals fluctuate in timeand by synchronizing those sets of neurons that are to be bound togetherinto a higher-level symbol/concept. Using the same idea, von der Malsburgand Schneider (1986) proposed a solution to the cocktail party problem. Inparticular, they developed a neural cocktail party processor that uses syn-chronization and desynchronization to segment the sensory inputs. Corre-lations are generated by an autonomous pattern formation process via neu-ron coupling and a new synaptic modulation rule. Though based on simpleexperiments (where von der Malsburg and Schneider used amplitude mod-ulation and stimulus onset synchrony as the main features of the sound, in
10 Several special issues have been devoted to ICA: Journal of Machine Learning Re-
search (Dec. 2003), Neurocomputing (Nov. 1998, Dec. 2002), Signal Processing (Feb. 1999,Jan. 2004), Proceedings of the IEEE (Oct. 1998), and IEEE Transactions on Neural Networks(March 2004). For textbook treatments of ICA theory, see Hyv¨arinen, Karhunen, and Oja(2001), and Cichocki and Amari (2002).
Figure 3: Neural correlated oscillator model from Wang and Brown (1999;adapted with permission).
line with Helmholtz’s suggestion), the underlying idea is illuminating. Justas important, the model is consistent with anatomical and physiologicalobservations. Mathematical details of the coupled neural oscillator modelwere later explored in von der Malsburg and Buhmann (1992). Note thatcorrelation theory is also applicable to the feature binding problem in vi-sual or sensor-motor systems (K ¨onig & Engel, 1995; Treisman, 1996; von derMalsburg, 1995, 1999).
The idea of oscillatory correlation, as a possible basis for CASA, was mo-
tivated by the early work of von der Malsburg (von der Malsburg, 1981;von der Malsburg & Schneider, 1986), and it was extended to different sen-sory domains whereby phases of neural oscillators are used to encode thebinding of sensory components (Wang, Bhumann, & von der Malsburg,1990; Wang, 1996). Subsequently, Brown and Wang (1997) and Wang andBrown (1999) developed a two-layer oscillator network (see Figure 3) thatperforms stream segregation based on oscillatory correlation. In the oscil-latory correlation-based model, a stream is represented by a population ofsynchronized relaxation oscillators, each of which corresponds to an au-ditory feature, and different streams are represented by desynchronizedoscillator populations. Lateral connections between oscillators encode theharmonicity and proximity in time and frequency. The aim of the model is toachieve “searchlight attention” by examining the temporal cross-correlationbetween the activities of pairs (or populations) of neurons:
) and y
) are assumed to be two zero-mean observable time series.
The neural oscillator model depicted in Figure 3 comprises two layers:
a segmentation layer and a grouping layer. The first layer acts as a locallyexcitatory, globally inhibitory oscillator, and the second layer essentiallyperforms auditory scene analysis. Preceding the oscillator network, there isan auditory periphery model (cochlear and hair cells) as well as a middle-level auditory representation stage (correlogram). As reported by Wang andBrown (1999), the model is capable of segregating a mixture of voiced speechand different interfering sounds, thereby improving the signal-to-noise
ratio (SNR) of the attended speech signal. The correlated neural oscilla-tor is arguably biologically plausible (Wang & Brown, 1999). In specificterms, the neural oscillator acts as a source extraction functional block bytreating the attended signal as a “foreground” stream and putting the re-maining segments into a “background” stream. However, the performanceof the neural oscillator appears to deteriorate significantly in the presence ofmultiple competitive sources. Recently, van der Kouwe, Wang, and Brown(2001) compared their neural oscillator model with representative BSS tech-niques for speech segregation in different scenarios; they reported that theperformance of the oscillator model varied from one test to another, de-pending on the time-frequency characteristics of the sources. Under mostof the noise conditions, the BSS technique more or less outperformed theoscillator model; however, the BSS techniques worked quite poorly whenapplied to sources in motion or gaussian sources due to the violation ofbasic ICA assumptions.
4.3 Cortronic Network.
The idea of a so-called cortronic network was
motivated by the fact that the human brain employs an efficient sparsecoding scheme to extract the features of sensory inputs and accesses themthrough associative memory. Using a cortronic neural network architectureproposed by Hecht-Nielsen (1998), a biologically motivated connectionistmodel has been recently developed to solve the machine CPP (Sagi et al.,2001). In particular, Sagi et al. view the CPP as an aspect of the human speechrecognition problem in a cocktail party environment, and thereby regard thesolution as an attended source identification problem. Only one microphoneis used to record the auditory scene; however, the listener is assumed to befamiliar with the language of conversation of interest and ignoring otherongoing conversations. All the subjects were chosen to speak the same lan-guage and have the same voice qualities. The goal of the cortronic networkis to identify one attended speech of interest.
The learning machine described in Sagi et al. (2001) is essentially an
associative memory neural network model. It consists of three distinct layers(regions): sound-input representation region, sound processing region, andword processing region. For its operation, the cortronic network rests ontwo assumptions:
r The network has knowledge of the speech signals (e.g., language con-
work of associative memory and pattern identification.
In attacking the machine CPP, the cortronic network undertakes threelevels of association (Sagi et al., 2001): (1) sound and subsequent sound,(2) sequence of sounds and the token (i.e., information unit) that is sparselycoded, and (3) a certain word and the word that follows it in the language.
In terms of performance described in Sagi et al., it appears that the cortronicnetwork is quite robust with respect to the variations of speech, speaker,and noise, even under a −8 dB SNR. From one microphone, it can extractone attended speech source in the presence of four additive speech inter-ferences (R. Hecht-Nielsen, personal communication, Sept. 2003). Differentfrom the other computational approaches proposed to solve the CPP, thecortronic network exploits the knowledge context of speech and language;it is claimed to address a theory of how the brain thinks rather than howthe brain listens.
5 Active Audition
The three approaches to computational auditory scene analysis describedin section 4 share a common feature: the observer merely listens to theenvironment but does not interact with it (i.e., the observer is passive).
In this section, we briefly discuss the idea of active audition in which theobserver (human or machine) interacts with the environment. This ideais motivated by the fact that human perception is not passive but active.
Moreover, there are many analogies between the mechanisms that go onin auditory perception and their counterparts in visual perception.11 In asimilar vein, we may look to active vision (on which much research has beendone for over a decade) for novel ideas in active audition as the frameworkfor an intelligent machine to solve the cocktail party problem.
According to Varela, Thompson, and Rosch (1991) and Sporns (2003),
embodied cognitive models rely on cognitive processes that emerge frominteractions between neural, bodily, and environment factors. A distinctivefeature of these models is that they use the world as their own model. Forexample, in active vision (also referred to as animated vision), proposed byBajcsy (1988) and Ballard (1991), among others, it is argued that vision isbest understood in the context of visual behaviors. The key point here is thatthe task of vision is not to build the model of a surrounding real world asoriginally postulated in Marr’s theory, but rather to use visual informationin the service of the real world in real time, and do so efficiently and in-expensively (Clark & Eliasmith, 2003). In effect, the active vision paradigmgives “action” a starring role (Sporns, 2003).
With this brief background on active vision, we may now propose a
framework for active audition, which may embody four specific functions:
1. Localization, the purpose of which is to infer the directions of incom-
ing sound signals. This function may be implemented by using an
11 The analogy between auditory and visual perceptions is further substantiated in
Shamma (2001), where it is argued that they share certain processing principles: lateralinhibition for edge and peak enhancement, multiscale analysis, and detection mechanismsfor temporal coincidence and spatial coincidence.
adaptive array of microphones, whose design is based on directionof arrival (DOA) estimation algorithms developed in the signal-processing literature (e.g., Van Veen & Buckley, 1997; Doclo & Moonen,2002).
2. Segregation and focal attention, where the attended sound stream of
interest (i.e., target sound) is segregated and the sources of interferenceare ignored, thereby focusing attention on the target sound source.
This function may be implemented by using several acoustic cues(e.g., ITD, IID, onset, and pitch) and then combining them in a fusionalgorithm.12
3. Tracking, the theoretical development of which builds on a state-space
model of the auditory environment. This model consists of a processequation that describes the evolution of the state with time and ameasurement equation that describes the dependence of the observ-ables on the state. More specifically, the state is a vector defined bythe acoustic cues (features) characterizing the target sound streamand its direction.13 By virtue of its very design, the tracker providesa one-step prediction of the underlying features of the target sound.
We may therefore view tracking as a mechanism for dynamic featurebinding.
4. Learning, the necessity of which in active audition is the key function
that differentiates an intelligent machine from a human brain. Au-dition is a sophisticated, dynamic information processing task per-formed in the brain, which inevitably invokes other tasks simultane-ously (such as vision and action). It is this unique feature that enablesthe human to survive in a dynamic environment. For the same rea-son, it is our belief that an intelligent machine that aims at solving acocktail party problem must embody a learning capability to adaptitself to an ever-changing dynamic environment. The learning abilitymust also be of a kind that empowers the machine to take “action”whenever changes in the environment call for it.
Viewed together, these four functions provide the basis for building an em-bodied cognitive machine that is capable of human-like hearing in an activefashion. The central tenet of active audition embodying such a machine isthat an observer may be able to understand an auditory environment more
12 This approach to segregation and focal attention is currently being pursued by the
first author, working with his research colleague Rong Dong at McMaster University.
13 In the context of tracking, Nix, Kleinschmidt, and Hohmann (2003) used a particle
filter as a statistical method for integrating temporal and frequency-specific features of atarget speech signal.
effectively and efficiently if the observer interacts with the environmentrather than is a passive observer.14
We conclude the article by doing two things. First, we present a philosoph-ical discussion that in the context of the cocktail party phenomenon, ICAalgorithm addresses an entirely different signal processing problem. Sec-ond, in an attempt to explain how the brain solves the CPP, we postulatea biologically motivated correlated firing framework for single-stream ex-traction based on our own recent work.
6.1 Brain Theory and ICA.
In the course of over ten years, the idea of
ICA has blossomed into a new field that has enriched the discipline of signalprocessing and neural computation. However, when the issue of interest isa viable solution to the cocktail party problem, the ICA/BSS framework hascertain weaknesses:
r Most ICA/BSS algorithms require that the number of sources not be
fewer than the number of independent signal sources.15 In contrast,the human auditory system requires merely two outer ears to solvethe cocktail party problem, and it can do so with relative ease.
r The matter of independent signal sources is assumed to remain con-
stant in the ICA/BSS framework. This is an unrealistic assumptionin a neurobiological context. For instance, it is possible for an audi-tory environment to experience a varying number of speakers (i.e.,sound sources) or a pulse-like form of noise (e.g., someone laughing orcoughing), yet the human capability to solve the cocktail party prob-lem remains essentially unaffected by the variations in the auditoryscene.
r Last but by no means least, the ICA/BSS framework usually requires
the separation of all source signals. With both outer ears of the humanauditory systems focused on a signal speaker of interest in a complexauditory scene, the cocktail party problem is solved by extracting thatspeaker’s speech signal and practically suppressing all other forms ofnoise or interference.
14 In this concluding statement on active audition, we have paraphrased the essence
of active vision (Blake & Yuille, 1992), or more generally, that of active perception (Bajcsy,1988).
15 In the literature, there are also some one-microphone BSS approaches that attempt
to exploit either the acoustic features and masking technique (Roweis, 2000), the splittingof time-domain disjoint/orthogonal subspaces (Hopgood & Rayner, 1999), or the priorknowledge of the source statistics (Jang & Lee, 2003).
Simply put, viewed in the context of the cocktail party problem, ICA/BSS
algorithms seem not to be biologically plausible. Rather, we say that for acomputational auditory scene analysis framework to be neurobiologicallyfeasible, it would have to accommodate the following ingredients:
r Ability to operate in a nonstationary convolutive environment, where
the speech signal of interest is corrupted by an unknown number ofcompeting speech signals or sources of noise.
r Ability to switch the focus of attention from one speech signal of in-
terest to another and do so with relative ease.
r Working with a pair of sensors so as to exploit the benefit of binaural
hearing. However, according to Sagi et al. (2001), it is claimed that a sin-gle sensor (reliance on monaural hearing) is sufficient to solve the cock-tail party problem. There is no contradiction here between these twostatements, as demonstrated by Cherry and coworkers over 50 yearsago: simply put, binaural hearing is more effective than monaural hear-ing by working with a significantly smaller SNR.
r Parallelism, which makes it possible for the incoming signal to be
worked on by a number of different paths, followed by a fusion oftheir individual outputs.
6.2 Correlative Neural Firing for Blind Single-Source Extraction.
relation theory has played an influential role on memory recall, coincidencedetection, novelty detection, perception, and learning, which cover most ofthe intelligent tasks of a human brain (Cook, 1991). Eggermont (1990), in hisinsightful book, has presented a comprehensive investigation of correlativeactivities in the brain; in that book, Eggermont argued that correlation, in oneform or another, is performed in 90 percent of the human brain. Eggermont(1993) also investigated the neural correlation mechanisms (such as coinci-dence detection and tuned-delay mechanisms) in the auditory system andshowed that synchrony is crucial in sound localization, pitch extractionof music, and speech coding (such as intensity-invariant representation ofsounds, suppressing less salient features of speech sound, and enhancingthe representation of speech formant).
Recently, we (Chen, 2005; Chen & Haykin, 2004) proposed a stochastic-
correlative firing mechanism and an associated learning rule for solving asimplified form of CPP. The idea of correlative firing mechanism is simi-lar to that of synchrony and correlation theory (von der Malsburg, 1981;Eggermont, 1990) and, interestingly enough, it is motivated by some earlywork on vision (Harth, Unnikrishnan, & Pandya, 1987; Mumford, 1995). Inour proposed correlative firing framework, the auditory cortex implementsa certain number of parallel circuits, each responsible for extracting the at-tended sound source of interest. By analogy with figure-ground segregation,the circuit extracts the “figure” from a complex auditory scene. To modelthe selective attention in the thalamus, a gating network is proposed for
deciding or switching the attention of the segregated sources, as depictedin the schematic diagram of Figure 2. Specifically, the proposed stochasticcorrelative learning rule (Chen, 2005; Chen & Haykin, 2004) can be viewedas a variant of the ALOPEX (ALgorithm Of Pattern EXtraction), an opti-mization procedure that was originally developed in vision research (Harth& Tzanakou, 1974; Tzanakou, Michalak, & Harth, 1979). The stochastic cor-relative learning rule is temporally asymmetric and Hebbian-like. The mostattractive aspect of this learning rule lies in the fact that it is both gradientfree and model independent, which make it potentially applicable to anyneural architecture that has hierarchical or feedback structures (Haykin,Chen, & Becker, 2004); it also allows us to develop biologically plausiblesynaptic rules based on the firing rate of the spiking neurons (Harth et al.,1987). Another appealing attribute of this learning rule is its parallelism inthat synaptic plasticity allows a form of synchronous firing, which lendsitself to ease of hardware implementation. The proposed correlative firingframework for the CPP assumes a time-invariant instantaneous linear mix-ing of the sources as in the ICA theory. However, unlike ICA, our methodis aimed at extracting the “figure” (i.e., one single stream), given two sen-sors (corresponding to two ears), from the “background” (auditory scene)containing more than two nongaussian sound sources; additionally, ourinterpretation of mixing coefficients in the mixing matrix as neurons’ fir-ing rates (instead of acoustic mixing effect) is motivated from the notionof the firing-rate stimulus correlation function in neuroscience (Dayan &Abbott, 2001), and the learning rule aimed at extracting a single source thatbears the highest synchrony in terms of neurons’ firing rate. In simulatedexperiments reported in Chen and Haykin, (2004), we have demonstrateda remarkable result: the algorithm is capable of extracting one single sourcestream given only two sensors and more than four sources (including onegaussian source). This is indeed an intractable task for ICA algorithms. Ourresults emphasize that the main goal of the solution to the CPP is not toseparate out the competing sources but rather to extract the source signal ofinterest (or “figure”) in a complex auditory scene that includes competingspeech signals and noise.
To conclude, our proposed stochastic correlative neural firing mecha-
nism, embodying its own learning rule, is indeed aimed at blind signalsource extraction. Most important, it provides a primitive yet arguably con-vincing basis for how the human auditory system solves the cocktail partyproblem; that is, it addresses question 2 of the introductory section. How-ever, in its present form, the learning rule is not equipped to deal withquestion 3 pertaining to computational auditory scene analysis.
The work reported in this review is supported by the Natural Sciences andEngineering Research Council of Canada. It was carried out under a joint
Canada-Europe BLISS project. We thank the editor and two anonymousreviewers for constructive suggestions and valuable comments that havehelped us reshape the review into its present form. We also thank P. Divenyi,S. Grossberg, S. Shamma, S. Becker, and R. Dong for helpful comments onvarious early drafts of this review and J. J. Eggermont, R. Hecht-Nielsen,and W. A. Yost for kindly supplying some references pertinent to the cocktailparty problem. We also greatly appreciate certain authors and publishersfor permission to reproduce or adapt Figures 1 and 3 in the review.
Amari, S. (2000). Estimating functions of independent component analysis for tem-
porally correlated signals. Neural Computation, 12
Amari, S., & Cichocki, A. (1998). Adaptive blind signal processing—neural network
approaches. Proceedings of the IEEE, 86
Amari, S., Cichocki, A., & Yang, H. H. (1996). A new learning algorithm for blind
signal separation. In D. Touretzky, M. Mozer, & M. Hasselmo (Eds.), Advances inneural information processing systems, 8
(pp. 757–763). Cambridge, MA: MIT Press.
Arons, B. (1992). A review of the cocktail party effect. Journal of the American Voice
Attias, H. (1999). Independent factor analysis. Neural Computation, 11
Bajcsy, R. (1988). Active perception. Proceedings of the IEEE, 76
Ballard, D. H. (1991). Animate vision. Artificial Intelligence, 48
Bell, A., & Sejnowski, T. J. (1995). An information-maximization approach to blind
separation and blind deconvolution. Neural Computation, 7
Belouchrani, A., Abed-Meraim, K., Cardoso, J.-F., & Moulines, E. (1997). A blind
source separation technique based on second order statistics. IEEE Transactionson Signal Processing, 45
Belouchrani, A., & Amin, M. G. (1998). Blind source separation technique based
on time-frequency representations. IEEE Transactions on Signal Processing, 46
Blake, A., & Yuille, A. (Eds.). (1992). Active vision
. Cambridge, MA: MIT Press.
Blauert, J. (1983). Spatial hearing: The psychophysics of human sound localization
Bregman, A. S. (1990). Auditory scene analysis: The perceptual organization of sound
Bregman, A. S. (1998). Psychological data and computational ASA. In D. F. Rosenthal
& H. G. Okuno (Eds.), Computational auditory scene analysis
. Mahwah, NJ: Erlbaum.
Bronkhorst, A. (2000). The cocktail party phenomenon: A review of research on
speech intelligibility in multiple-talker condition. Acoustica, 86
Brown, G. J. (1992). Computational auditory scene analysis: A representational approach
Unpublished doctoral dissertation, University of Sheffield.
Brown, G. J., & Cooke, M. P. (1994). Computational auditory scene analysis. Computer
Speech and Language, 8
Brown, G. J., & Wang, D. L. (1997). Modelling the perceptual segregation of concurrent
vowels with a network of neural oscillation. Neural Networks, 10
Brown, G. D., Yamada, S., & Sejnowski, T. J. (2001). Independent component analysis
at the neural cocktail party. Trends in Neuroscience, 24
Cardoso, J.-F. (1998). Blind signal separation: Statistical principles. Proceedings of the
Cardoso, J.-F. (2001). The three easy routes to independent component analysis: Con-
trasts and geometry. In Proc. ICA2001
. San Diego.
Cardoso, J.-F. (2003). Dependence, correlation and gaussianity in independent com-
ponent analysis. Journal of Machine Learning Research, 4
Cardoso, J.-F., & Souloumiac, A. (1993). Blind beamforming for non-Gaussian signals.
IEE Proceedings–F, 140
Chan, K., Lee, T.-W., & Sejnowski, T. J. (2003). Variational Bayesian learning of ICA
with missing data. Neural Computation, 15
Chen, Z. (2003). An odyssey of the cocktail party problem
(Tech. Rep.). Hamilton,
Ontario: Adaptive Systems Lab, McMaster University. Available online: http://soma.crl.mcmaster.ca/∼zhechen/download/cpp.ps
Chen, Z. (2005). Stochastic correlative firing for figure-ground segregation. Biological
Chen, Z., & Haykin, S. (2004). Figure-ground segregation in sensory perception using
a stochastic correlative learning rule
(Tech. Rep.). Hamilton, Ontario: AdaptiveSystems Lab, McMaster University. Available online: http://soma.crl.mcmaster.
Cherry, E. C. (1953). Some experiments on the recognition of speech, with one and
two ears. Journal of the Acoustical Society of America, 25
Cherry, E. C. (1957). On human communication: A review, survey, and a criticism
Cherry, E. C. (1961). Two ears—but one world. In W. A. Rosenblith (Ed.), Sensory
(pp. 99–117). New York: Wiley.
Cherry, E. C., & Sayers, B. (1956). Human “cross-correlation”—A technique for mea-
suring certain parameters of speech perception. Journal of the Acoustical Society ofAmerica, 28
Cherry, E. C., & Sayers, B. (1959). On the mechanism of binaural fusion. Journal of the
Acoustical Society of America, 31
Cherry, E. C., & Taylor, W. K. (1954). Some further experiments upon the recognition
of speech, with one and, with two ears. Journal of the Acoustical Society of America,26
Cichocki, A., & Amari, S. (2002). Adaptive blind signal and image processing
. New York:
Clark, A., & Eliasmith, C. (2003). Philosophical issues in brain theory and connec-
tionism. In M. Arbib (Ed.), Handbook of brain theory and neural networks
(2nd ed.,pp. 886–888). Cambridge, MA: MIT Press.
Comon, P. (1994). Independent component analysis, a new concept? Signal Processing,
Cook, J. E. (1991). Correlated activity in the CNS: A role on every timescale? Trends
in Neuroscience, 14
Cooke, M. (1993). Modelling auditory processing and organization
Cooke, M. P. (2002, December). Computational auditory scene analysis in listeners and
. Tutorial at NIPS2002, Vancouver, Canada.
Cooke, M. P., & Brown, G. J. (1993). Computational auditory scene analysis: Exploit-
ing principles of perceived continuity. Speech Communication, 13
Cooke, M., & Ellis, D. (2001). The auditory organization of speech and other sources
in listeners and computational models. Speech Communication, 35
Crick, F. (1984). Function of the thalamic reticular complex: The searchlight hypoth-
esis. Proc. Natl. Acad. Sci. USA, 81
Dayan, P., & Abbott, L. F. (2001). Theoretical neuroscience: Computational and mathemat-
ical modeling of neural systems
. Cambridge, MA: MIT Press.
Doclo, S., & Moonen, M. (2002). Robust adaptive time-delay estimation for speaker
localization in noisy and reverberant acoustic environments. EURASIP Journal ofApplied Signal Processing, 5
Divenyi, P. (Ed.). (2004). Speech separation by humans and machines
. Berlin: Springer.
Durlach, N. I., & Colburn, H. S. (1978). Binaural phenomena. In E. C. Cartrette & M. P.
Friedman (Eds.), Handbook of perception.
New York: Academic Press.
Eggermont, J. J. (1990). The correlative brain: Theory and experiment in neural interaction
Eggermont, J. J. (1993). Function aspects of synchrony and correlation in the auditory
nervous system. Concepts in Neuroscience, 4
Ellis, D. (1996). Prediction-driven computational auditory scene analysis
Engel, A. K., K ¨onig, P., & Singer, W. (1991). Direct physiological evidence for scene
segmentation by temporal coding. Proc. Natl. Acad. Sci. USA, 88
Feng, A. S., & Ratnam, R. (2000). Neural basis of hearing in real-world situations.
Annual Review of Psychology, 51
Grossberg, S., Govindarajan, K., Wyse, L. L., & Cohen, M. A. (2004). ARTSTREAM: A
neural network model of auditory scene analysis and source segregation. NeuralNetworks, 17
Harth, E., & Tzanakou, E. (1974). Alopex: A stochastic method for determining visual
receptive fields. Vision Research, 14
Harth, E., Unnikrishnan, K. P., & Pandya, A. S. (1987). The inversion of sensory
processing by feedback pathways: A model of visual cognitive functions. Science,237
Hawley, M. L., Litovsky, R. Y., & Colburn, H. S. (1999). Speech intelligibility and lo-
calization in a multisource environment. Journal of the Acoustical Society of America,105
Haykin, S. (2003, June). The cocktail party phenomenon
. Presentation at the ICA Work-
Haykin, S., Chen, Z., & Becker, S. (2004). Stochastic correlative learning algorithms.
IEEE Transactions on Signal Processing, 52
Hebb, D. O. (1949). The organization of behavior: A neuropsychological theory
. New York:
Hecht-Nielsen, R. (1998). A theory of cerebral cortex. In Proc. 1998 Int. Conf. Neural
Information Processing, ICONIP’98
(pp. 1459–1464). Burke, VA: IOS Press.
Hopgood, E., & Rayner, P. (1999). Single channel signal separation using linear time-
varying filters: Separability of non-stationary stochastic signals. In Proc. ICASSP
(Vol. 3, pp. 1449–1452). Piscataway, NJ: IEEE Press.
Hyv¨arinen, A., Karhunen, V., & Oja, E. (2001). Independent component analysis
Hyv¨arinen, A., & Oja, E. (1997). A fast fixed-point algorithm for independent com-
ponent analysis. Neural Computation, 9
James, W. (1890). Psychology (briefer course)
. New York: Holt.
Jang, G.-J., & Lee, T. W. (2003). A maximum likelihood approach to single-channel
source separation. Journal of Machine Learning Research, 4
Jones, M., & Yee, W. (1993). Attending to auditory events: The role of temporal
organization. In S. McAdams & E. Bigand (Eds.), Thinking in sound
Oxford: Clarendon Press.
Julesz, B., & Hirsh, I. J. (1972). Visual and auditory perception—an essay in compar-
ison. In E. David & P. B. Denes (Eds.), Human communication: A unified view
. NewYork: McGraw-Hill.
Jutten, C., & H´erault, J. (1991). Blind separation of sources, part I–III. Signal Processing,
Kandel, E. R., Schwartz, J. H., & Jessell, T. M. (Eds.). (2000). Hearing. In Principles of
(4th ed.). New York: McGraw-Hill.
Knuth, K. H. (1999). A Bayesian approach to source separation. In Proc. ICA’99
K ¨onig, P., & Engel, A. K. (1995). Correlated firing in sensory-motor systems. Current
Opinions in Neurobiology, 5
K ¨onig, P., Engel, A. K., & Singer, W. (1996). Integrator or coincidence detector? The
role of the cortical neuron revisited. Trends in Neuroscience, 19
MacLean, W. R. (1959). On the acoustics of cocktail parties. Journal of the Acoustical
Society of America, 31
Marr, D. (1982). Vision
. San Francisco: Freeman.
McGurk, H., & MacDonald, J. (1976). Hearing lips and seeing voices. Nature, 264
Mohammad-Djafari, A. (1999). A Bayesian approach to source separation. In Proc.
19th Int. Workshop on Maximum Entropy and Bayesian Methods (MaxEnt99)
Moore, B. C. J. (1997). An introduction to the psychology of hearing
(4th ed.). San Diego:
Moray, N. (1959). Attention in dichotic listening: Affective cues and the influence of
instructions. Quarterly Journal of Experimental Psychology, 27
Mumford, D. (1991). On the computational architecture of the neocortex. Part I: The
role of the thalamocortical loop. Biological Cybernetics, 65
Mumford, D. (1995). Thalamus. In M. Arbib (Ed.), The handbook of brain theory and
(pp. 981–984). Cambridge, MA: MIT Press.
Nix, J., Kleinschmidt, M., & Hohmann, V. (2003). Computational auditory scene
analysis of cocktail-party situations based on sequential Monte Carlo methods.
In Proc. 37th Asilomar Conference on Signals, Systems and Computers
Redondo Beach, CA: IEEE Computer Society Press.
Parra, L., & Alvino, C. (2002). Geometric source separation: Merging convolutive
source separation with geometric beamforming. IEEE Transactions on Speech andAudio Processing, 10
Parra, L., & Spence, C. (2000). Convolutive blind source separation of nonstationary
sources. IEEE Transactions on Speech and Audio Processing, 8
Pham, D. T. (2002). Mutual information approach to blind separation of stationary
sources. IEEE Transactions on Information Theory, 48
Pham, D. T., & Cardoso, J. F. (2001). Blind separation of instantaneous mixtures
of non-stationary sources. IEEE Transactions on Signal Processing, 49
Rickard, S., Balan, R., & Rosca, J. (2001). Real-time time-frequency based blind source
separation. In Proc. ICA2001
(pp. 651–656). San Diego, CA.
Rickard, S., & Yilmaz, O. (2002). On the approximate W
-disjoint orthogonality of
speech. In Proc. ICASSP2002
(pp. 529–532). Piscataway, NJ: IEEE Press.
Rosenthal, D. F., & Okuno, H. G. (Eds.). (1998). Computational auditory scene analysis
Rowe, D. B. (2002). A Bayesian approach to blind source separation. Journal of Inter-
disciplinary Mathematics, 5
Roweis, S. (2000). One microphone source separation. In T. K. Leen, T. G. Dietterich, &
V. Tresp (Eds.), Advances in neural information processing systems, 13
Cambridge, MA: MIT Press.
Sabine, H. J. (1953). Room acoustics. Transactions of IRE, 1
Sagi, S., Nemat-Nasser, S. C., Kerr, R., Hayek, R., Downing, C., & Hecht-Nielsen, R.
(2001). A biologically motivated solution to the cocktail party problem. NeuralComputation, 13
Sayers, B., & Cherry, E. C. (1957). Mechanism of binaural fusion in the hearing of
speech. Journal of the Acoustical Society of America, 31
Schultz, S. R., Golledge, H. D. R., & Panzeri, S. (2001). Synchronization, binding and
the role of correlated firing in fast information transmission. In S. Wermter, J.
Austin, & D. Willshaw (Eds.), Emergent neural computational architectures based onneuroscience
. Berlin: Springer-Verlag.
Shamma, S. (2001). On the role of space and time in auditory processing. Trends in
Cognitive Sciences, 5
Singer, W. (1993). Synchronization of cortical activity and its putative role in infor-
mation processing and learning. Annual Review of Physiology, 55
Singer, W. (1995). Synchronization of neural responses as putative binding mecha-
nism. In M. Arbib (Ed.), Handbook of brain theory and neural networks
Cambridge, MA: MIT Press.
Sporns, O. (2003). Embodied cognition. In M. Arbib (Ed.), Handbook of brain theory
and neural networks
(2nd edition, pp. 395–398). Cambridge, MA: MIT Press.
Tong, L., Soon, V., Huang, Y., & Liu, R. (1990). AMUSE: A new blind identification
problem. In Proc. ICASSP
(pp. 1784–1787). Piscataway, NJ: IEEE Press.
Treisman, A. M. (1996). The binding problem. Current Opinion in Neurobiology, 6
Treisman, A. M., & Gelade, G. (1980). A feature integration theory of attention. Cog-
nitive Psychology, 12
Tzanakou, E., Michalak, R., & Harth, E. (1979). The Alopex process: Visual receptive
fields by response feedback. Biological Cybernetics, 35
van der Kouwe, A. J. W., Wang, D. L., & Brown, G. J. (2001). A comparison of audi-
tory and blind separation techniques for speech segregation. IEEE Transactions onSpeech and Audio Processing, 9
Van Veen, B., & Buckley, K. (1997). Beamforming techniques for spatial filtering. In
V. K. Madisetti & D. B. Williams (Eds.), Digital signal processing handbook
. BocaRaton, FL: CRC Press.
Varela, F., Thompson, E., & Rosch, E. (1991). The embodied mind: Cognitive science and
. Cambridge, MA: MIT Press.
von der Malsburg, C. (1981). The correlation theory of brain function
. (Internal Rep. 81-
2). G ¨ottingen: Department of Neurobiology, Max-Plank-Institute for BiophysicalChemistry.
von der Malsburg, C. (1995). Binding in models of perception and brain function.
Current Opinion in Neurobiology, 5
von der Malsburg, C. (1999). The what and why of binding: The modeler’s perspec-
von der Malsburg, C., & Buhmann, J. (1992). Sensory segmentation with coupled
neural oscillators. Biological Cybernetics, 67
von der Malsburg, C., & Schneider, W. (1986). A neural cocktail-party processor.
Biological Cybernetics, 54
Wang, D. L. (1996). Primitive auditory segregation based on oscillatory correlation.
Cognitive Science, 20
Wang, D. L., & Brown, G. J. (1999). Separation of speech from interfering sounds based
on oscillatory correlation. IEEE Transactions on Neural Networks, 10
Wang, D. L., Buhmann, J., & von der Malsburg, C. (1990). Pattern segmentation in
associative memory. Neural Computation, 2
Warren, R. M. (1970). Perceptual restoration of missing speech sounds. Science, 167
Warren, R. M., Obusek, C. J., & Ackroff, J. M. (1972). Auditory induction: Perception
synthesis of absent sounds. Science, 176
Wood, N. L., & Cowan, N. (1995). The cocktail party phenomenon revisited: Attention
and memory in the classic listening selective procedure of Cherry (1953). Journalof Experimental Psychology, General, 124
Yilmaz, O., & Rickard, S. (2004). Blind separation of speech mixtures via time-
frequency masking. IEEE Transactions on Signal Processing, 52
Yost, W. A. (1991). Auditory image perception and analysis. Hearing Research, 46
Yost, W. A. (1997). The cocktail party problem: Forty years later. In R. Gilkey &
T. Anderson (Eds.), Binaural and spatial hearing in real and virtual environments
(pp. 329–348). Ahwah, NJ: Erlbaum.
Yost, W. A. (2000). Fundamentals of hearing: An introduction
(4th ed.). San Diego: Aca-
Yost, W. A., & Gourevitch, G. Eds. (1987). Directional hearing
. New York: Springer-
Received April 1, 2004; accepted February 23, 2005.
EL CONOCIMIENTO AL SERVICIO DE LA PERSONAALGUNAS VERDADES SOBRE PRINCIPIOS ACTIVOS Descripción General Este documento ha sido elaborado con la finalidad de poder presentar información respecto al mecanismo de acción de las pastil as anticonceptivas. Por supuesto, la prudencia indica que es necesario contrastar esta información con profesionales de la salud, ojalá que sean PROVIDA;
INT J TUBERC LUNG DIS 5(3):272–277 © 2001 IUATLD Pulmonary resection in the treatment of patients with pulmonary multidrug-resistant tuberculosis in Taiwan C-Y. Chiang,* M-C. Yu,* K-J. Bai,* J. Suo,* T-P. Lin,* Y-C. Lee† * Chronic Disease Control Bureau, Department of Health, Shen-Keng Shiang, Taipei County, † Department of Surgery, National Taiwan University Hospital, Taipei, Tai