Search

The findings are reported of an investigation into rhythmic differences between infant-directed speech (IDS) and adult-directed speech (ADS) in a corpus of utterances from Australian English mothers speaking to their infants and to another adult. Given the importance of rhythmic cues to stress and word-segmentation in English, the investigation focused on the extent to which IDS makes such cues salient. Two methods of analysis were used: one focused on segmental durational properties, using a variety of durational measures; the other focused on the prominence of vocalic/sonorant segments, as determined by their duration, intensity, pitch, and spectral balance, using individual measures as well as composite measures of prominence derived from auditory-model analyses. There were few IDS/ADS differences/trends on the individual measures, though mean pitch and pitch variability were higher in IDS than ADS, while IDS vowels showed more negative spectral tilt. However, the model-based analyses suggested that differences in the prominence of vowels/sonorant segments were reduced in IDS, with further analysis suggesting that pitch contributed little to prominence. The reduction in prominence contrasts may be due to the importance of mood-regulation in speech to young infants, and may suggest that infants rely on segmental cues to stress and word-segmentation.

In this paper data are presented from an experiment which provides evidence for the existence of a short latency, acoustically evoked potential of probable vestibular origin. The experiment was conducted in two phases using bone-conducted acoustic stimulation. In the first phase subjects were stimulated with 6-ms, 500-Hz tone bursts in order to obtain the threshold V(T) for vestibular evoked myogenic potentials (VEMP). It was confirmed that the difference between bone-conducted auditory and acoustic vestibular thresholds was slightly over 30 dB. The estimated threshold was then used as a reference value in the second part of the experiment to stimulate subjects over a range of intensities from -6 to +18 dB (re: V(T)). Averaged EEG recordings were made with eight Ag/AgCl electrodes placed on the scalp at Fpz, F3, F4, F7, F8, Cz, T3, and T4 according to the 10-20 system. Below V(T) auditory midlatency responses (MLRs) were observed. Above V(T) two additional potentials appeared: a positivity at about 10 ms (P10) which was maximal at Cz, and a negativity at about 15 ms (N15) which was maximal at Fpz. Extrapolation of the growth functions for the P10 and N15 indicated a threshold close to V(T), consistent with a vestibular origin of these potentials. Given the low threshold of vestibular acoustic sensitivity it is possible that this mode may make a contribution to the detection of and affective responses to loud low frequency sounds. The evoked potentials may also have application as a noninvasive and nontraumatic test of vestibular projections to the cortex.

In this paper new evidence is provided to indicate that vestibular responses may be obtained from loud dance music for intensities above 90 dB(A) SPL (Impulse-weighted). In a sample of ten subjects acoustically evoked EMG were obtained from the sternocleidomastoid muscle in response to a sample of techno music typical of that which may be experienced in a dance club. Previous research has shown that this response is vestibularly mediated since it can be obtained in subjects with loss of cochlear function, but is absent in subjects with loss of vestibular function (Colebatch et al. [J. Neurol. Neurosurg. Psychiatr. 57, 190–197 (1994)]. Given that pleasurable sensations of self-motion are widely sought after by more normal means of vestibular stimulation, it is suggested that acoustically evoked sensations of self-motion may account for the compulsion to exposure to loud music. Given further the similarity between the thresholds found, and the intensities and frequency distributions that are typical in rock concerts and dance clubs, it is also suggested that this response may be a physiological basis for the minimum loudness necessary for rock and dance music to work—the “rock and roll threshold” [Dibble, J. Audio Eng. Soc. 43(4), 251–266 (1995)].

A theory of the principal monaural pathway is described according to which temporal information is spatially coded on three dimensions roughly corresponding to the levels of, respectively, the cochlea, the inferior colliculus, and the auditory cortex [C. Schreiner and G. Langner, ‘‘Coding of temporal patterns in the central auditory nervous system,’’ Auditory Function: Neurobiological Bases of Hearing, edited by G. M. Edelman et al. (Wiley, New York, 1988), pp. 337–361]. This theory is expressed in the form of a computational model which simulates peripheral and central processing and which has the following main components: (1) a one‐dimensional linear bandpass filter‐bank to simulate the cochlea; (2) a two‐dimensional amplitude‐modulation (AM) bandpass filter‐bank to simulate the colliculus; (3) a three‐dimensional AM bandpass filter‐bank to simulate layer IV receptive cells in the cortex; (4) a cortical cross‐channel correlation mechanism; and (5) a central pattern recognition mechanism. Overall, periodicity pitch (i.e., periods of up to about 1000 Hz) is primarily associated with subcortical processing whereas time and rhythm (i.e., periods of up to about 20 Hz) are primarily associated with cortical processing. The model is applied to the following phenomena: virtual pitch shift, musical chords, the psychophysical law for interval discrimination, and the filled interval illusion.

The model [N. Todd, ‘‘A theory of the principal monaural pathway. I. Pitch and time perception,’’ J. Acoust. Soc. Am. (these proceedings)] proposes to account for auditory streaming by a cross‐correlation mechanism modeled as an array of cortical columns. A column contains excilatory and inhibitory interneurons and pyramidal cells which receive input from both the thalamus and from neighboring columns. For columns with coherent thalamic inputs, the outputs of the pyramids sum across frequency because local interneurons compute a correlation between the AM transform of the inputs of the local and remote columns, and selectively gate the input to the pyramidal cell. (1) Grouping by frequency proximity—for columns separated by less than a critical band, thalamic inputs are unresolved. (2) Grouping by temporal proximity—at repetition rates below the mean of the AM distribution, AM fundamentals are less well represented; at higher repetition rates the AM harmonics are more separated. (3) Temporal development—w...

The notion that there is an intimate relationship between musical motion and physical movement is an old one and can be traced back to antiquity. Recently this idea has again received some attention, particularly in relation to musical expression. To use a modern metaphor, one can consider expressive performance to be analogous to the problems of kinematics and trajectory planning in robotics. The trajectories referred to, however, are not those of the performers limbs in physical space, but those of an abstract movement relative to a metrical grid associated with a musical score. Recent studies have attempted to substantiate this idea by comparing a model of motion with timing measurements of the final ritardandi from actual performances. This study extends these earlier analyses to include the accelerandi as well as the ritardandi from complete performances. One conclusion is that the variation of tempo in music can be reasonably compared with velocity in the equations of elementary mechanics. Further, it is suggested that the origin of metrical space, upon which the motion concept rests, lies in the way the auditory system processes rhythm.

A computational model of musical dynamics is proposed that complements an earlier model of expressive timing. The model, implemented in the artificial intelligence language LISP, is based on the observation that a musical phrase is often indicated by a crescendo/decrescendo shape. The functional form of this shape is derived by making two main assumptions. First, that musical dynamics and tempo are coupled, that is, ‘‘the faster the louder, the slower the softer.’’ This tempo/dynamics coupling, it is suggested, may be a characteristic of some classical and romantic styles perhaps exemplified by performances of Chopin. Second, that the tempo change is governed by analogy to physical movement. The allusion of musical expression to physical motion is further extended by the introduction of the concepts of energy and mass. The utility of the model, in addition to giving an insight into the nature of musical expression, is that it provides a basis for a method of performance style analysis.

Over 70 years ago John Tait raised a question with regards to the certainty that the line of division between the hearing and vestibular portions of the VIIIth cranial nerve had been drawn at the proper place [J. Tait, Ann. Otol. Rhinol. Laryngol. 41, 681–704 (1932)]. His reason for raising this question was the evidence which existed at that time that the otolith organs of anamniotes (fish and amphibians), most commonly the sacculus, were demonstrably organs of hearing in many species. He went on to speculate about a possible auditory role for the otolith organs in amniotes (reptiles, birds, and mammals), including in humans, although at this time there was no direct evidence for this. Since this time, however, there has developed a significant literature documenting evidence, not only for anamniote otolithic hearing, but also for a conserved acoustic sensitivity for the amniote otoliths. In the case of humans, this sensitivity has been exploited for the development of new acoustic tests of vestibular function. The time is ripe 70 years on, therefore, for us to revisit John Tait’s question and explore the possibility that the sacculus may indeed contribute to human hearing. [Work supported by the Royal Society.]

A recent study concerning the production, transmission and reception of alligator vocalizations has indicated that there are significant gaps in our knowledge of the physiology and physics of low‐frequency sound communication in air and water [Todd, J. Acoust. Soc. Am. 122, 2906–2915 (2007)]. In this paper these issues are discussed, and some possible solutions are considered for aquatic communication by both amphibians and crocodilians in the light of some other recent data on frog vocal behavior in water.

In this article the results are reported of a study to measure the intensity of the vocal displays of a population of American alligators (Alligator mississippiensis). It was found that the dominant frequencies in air range between 20 and 250 Hz with a source sound pressure level (SPL) of 91-94 dB at 1 m. The active space for the air-borne component is defined by the background and was estimated to be in a range up to 159 m in the 125-200 Hz band. For the water-borne component the dominant frequency range was 20-100 Hz with a source SPL of 121-125 dB at 1 m. The active space in water is defined by hearing thresholds and was estimated to range up to 1.5 km in the 63-100 Hz band. In the lowest frequency bands, i.e., 16-50 Hz, the estimated active space for otolith detection of near-field particle motion in water ranged to 80 m, which compared significantly with far-field detection for these frequencies. It is suggested that alligator vocal communication may involve two distinct sensory mechanisms which may subserve the functions of scene analysis and reproduction, respectively.

The aim of this study was to carry out an acoustic analysis of alligator vocal displays recorded simultaneously in air and water in order to estimate the role of the sacculus in crocodilian acoustic communication. Crocodilians are interesting animals to study as they are aquatic, highly vocal, and morphologically have a large sacculus. Alligators, in particular, are known to be one of the most vocal species. Previous studies of alligator hearing in air and water have indicated a maximum sensitivity between 800 and 1000 Hz [Higgs et al., J. Comp. Physiol. 188, 217–223 (2002)]. Analysis of alligator vocalizations, however, indicates that most of the power in adult vocalizations in water is typically between 30–50 Hz. These results leave the very real possibility that the basilar papilla may not be the primary receptor during crocodilian water‐borne communication. At 1‐m distance peak SPLs of 140 dB can be obtained in water and, in conjunction with existing threshold data, the effective radius of saccular acoustic sensing of the above source can be estimated to be approximately 30 m, which is easily greater than the typical distance between performer and receiver animals during a display.

It has long been noted that the world’s languages vary considerably in their rhythmic organization. Different languages seem to privilege different phonological units as their basic rhythmic unit, and there is now a large body of evidence that such differences have important consequences for crucial aspects of language acquisition and processing. The most fundamental finding is that the rhythmic structure of a language strongly influences the process of spoken‐word recognition. This finding, together with evidence that infants are sensitive from birth to rhythmic differences between languages, and exploit rhythmic cues to segmentation at an earlier developmental stage than other cues prompted the claim that rhythm is the key which allows infants to begin building a lexicon and then go on to acquire syntax. It is therefore of interest to determine how differences in rhythmic organization arise at the acoustic/auditory level. In this paper, it is shown how an auditory model of the primitive representation of sound provides just such an account of rhythmic differences. Its performance is evaluated on a data set of French and English sentences and compared with the results yielded by the phonetic accounts of Frank Ramus and his colleagues and Esther Grabe and her colleagues.

In the paper a second algorithm is presented for the computation of a primitive representation of rhythmic and temporal structure [N. P. Todd, C. S. Lee, and D. J. O’Boyle, J. New Mus. Res. 28, 5–28 (1999)]. The algorithm consists of the following stages: (1) cochlear filtering, by means of an ear model based on the gammatone filter‐bank, (2) modulation filtering, by means of a constant‐Q band‐pass modulation filter‐bank; (3) peak detection and sensorimotor filtering; (4) feedforward control of a simple plant to simulate tapping. The output of the modulation filter‐bank produces a representation in the form of a modulation spectrogram whereby any periodic input gives rise to a harmonic series. Metrical structure may be inferred from the harmonic structure of the spectrum with the addition of a learning and recognition mechanism. This algorithm may account for a number of additional phenomena including the approximate Weber law property of tempo sensitivity, adaptation to perturbations in the period or pha...

In the paper an algorithm is presented for the computation of a primitive representation of rhythmic structure [N. P. Todd, J. New Mus. Res. 23, 25–70 (1994)]. The algorithm consists of the following stages: (1) cochlear filtering, by means of an ear model based on the gammatone filter‐bank [M. Cooke, Modeling Auditory Processing and Organisation (Cambridge University Press, Cambridge, MA, 1992)]; (2) modulation filtering, by means of a Gaussian low‐pass modulation filter‐bank; (3) peak detection and integration of the output of the modulation filter‐bank. When the peak outputs of the modulation filters are plotted in a time‐constant/time diagram they form a representation, referred to as a rhythmogram, which resembles a more conventional tree diagram of rhythmic structure. A single acoustic event gives rise to a single branch in the representation so that integration along a branch yields a single number corresponding to the perceptual prominence or stress of the event. The integration mechanism can be thought of as an acoustic sensory memory process and accounts for a number of important phenomena in rhythm perception, including persistence, partial backward masking, interval produced accents and grouping by temporal proximity and accent. This representation may form the input to higher level learning and recognition mechanisms.

Evidence accumulated over the last few decades supports the case that the sacculus has conserved an acoustic sensitivity throughout vertebrate phylogeny [N. P. Todd, J. Acoust. Soc. Am. 110, 380–390 (2001)]. In humans saccular acoustic sensitivity may be demonstrated by means of myogenic vestibular evoked potentials (MVEP) but the question remains whether saccular acoustic sensitivity in humans has any perceptual or behavioral significance. It has been shown that MVEP have the following properties: (1) MVEP can be obtained to natural acoustic stimuli above about 90 dB SPL; (2) MVEP have a frequency tuning property with a best frequency between 300 Hz–350 Hz; (3) MVEP exhibit adaptation, characteristic or normal sensory processing; and (4) there is a change of the quality of sensation above the MVEP threshold, indicative that saccular acoustic sensitivity may play a role in the perception of loud sound. In this paper new evidence is discussed which indicates that both myogenic and neurogenic vestibular evoked potentials (NVEP), particularly those obtainable from prefrontal cortex, may be linked to subjective and behavioral responses to loud sound. These responses may in turn be related to individual differences in the functioning of the mesolimbic dopamine system, including prefrontal cortex.

There has amassed considerable evidence that the sacculus, an organ of hearing in lower vertebrates, has retained some acoustic sensitivity throughout phylogeny. In humans, myogenic vestibular‐evoked potentials (MVEPs) may be obtained from motorneurones inervated by the vestibulo‐spinal tract, particularly from the cervical region of the spinal cord. The MVEP has been studied principally as a noninvasive clinical tool for evaluation of normal otolith vestibular function, since traditional nystagmographic methods only really assess canal function. However, we have been interested in what ‘‘auditory’’ function acoustic sensitivity sacculus may have in the perception of sound. The results of a series of experiments are reported which may be summarized as follows. (1) MVEP has a frequency tuning [Todd, Cody, and Banks, Hear. Res. (in press)], with a maximum response frequency between 300 and 350 Hz and a band‐width of about 3 octaves. (2) MVEP can be obtained to natural acoustic stimuli [Todd and Cody, J. Acoust. Soc. Am. 107, 496–500 (2000)], such as dance music, above about 90 dB SPL. (3) MVEP can be obtained to continuous sounds. Given the above then, there are a number of possibilities where acoustic sensitivity of the sacculus may play a role in human perception.

Discussed is a computational model which attempts to simulate the principal brain structures involved in sensory‐motor sequencing, including auditory cortex, posterior parietal cortex, cerebellum, motor cortex, and the motor output system. A particular focus in this presentation is on the auditory cortical component responsible for computing a 3‐D spatio‐temporal power spectrum of the auditory image flow [N. Todd, Proc. 16th ICA (1998)], which captures information topographically on three levels, roughly corresponding to timbre, pitch, and rhythm, and reflects the temporal resolution of the cochlea, inferior colliculus, and auditory cortex, respectively. By considering the spatio‐temporal orientations of cortical receptive fields (RFs), it is possible to delineate a set of stationary and moving primitive features upon which the higher‐order constructs of pitch and timbre are built. A separate model population of secondary auditory cortical neurons, which receive inputs from the population of primary RFs, provides a potential means of accounting for a number of important phonological and melodic processes. Implications of this model are discussed for the development of new algorithms for the representation and segregation of speech via self‐organization of primary and secondary RFs by means of the principle of information maximization.

In order to evaluate the cortical model proposed in Todd [these proceedings] it was applied to the phenomenon of modulation detection. It was assumed that a discrimination task involves the correlation of the spatio‐temporal power spectra of two image sequences, such as may occur in a typical 2A‐2IFC procedure. The correlation was computed by k(1−r), where r is the product moment correlation, for four different modulation frequencies, 4, 16, 64, and 256 Hz at four modulation depths for both AM and FM of a 1000‐Hz carrier of 1000‐ms duration against the image power spectrum of an unmodulated 1000‐Hz 1000‐ms tone. The obtained measures were then linearly regressed (with no intercept) as a function of the square of the modulation index. In all cases k(1−r) gave a good fit (R2>0.95) for both AM and FM. For each of the modulation frequencies the slope of the regression could be used to estimate the model thresholds. The model shows the correct behavior for AM and FM detection. This mechanism, it is suggested, ...

A model is presented here based on the theory that the auditory cortex computes the spatio‐temporal power spectrum of the auditory image flow [D. Heeger, ‘‘Model for the extraction of image flow,’’ J. Opt. Soc. Am. 4, 1450–1470 (1987)]. A computational implementation of the model has the following components: (1) a gammatone filter‐bank spaced on the ERB rate scale; (2) an orthogonal, ERB spaced modulation filter‐bank which generates a 2‐D modulation map which is then low‐pass filtered; and (3) a population of causal 3‐D spatio‐temporal Gabor filters which compute the spatio‐temporal power spectrum of the image sequence. The cortical filters are organized into families such that each member of the family occupies the same spatial‐frequency band but has a different spatio‐temporal orientation. The spatial frequencies range from 0.05 cycles erb−1 to 0.4 cycles erb−1. The temporal frequencies range from ±0.5 to ±64 Hz. The spatio‐temporal power spectrum may be visualized by replacing each image pixel with th...

Previously, a multiscale mechanism [N. Todd and G. Brown, J. Acoust. Soc. Am. 94(A) (1994)] has been shown to be effective in locating significant events in a speech signal and in estimating their relative stress. Two experiments were carried out in order to determine whether major stress points identified by the model correspond to so‐called ‘‘p centers’’ [Morton et al., Psychol. Rev. 83(5), 405–408 (1976)]. Recordings were made of (1) two subjects reciting the digits ‘‘one’’ to ‘‘eight’’ and (2) a number of subjects reciting metrical verse in synchrony with a metronome. The speech was recorded on one channel, the metronome on the other. The speech was then extracted without the metronome and fed into the model and the predicted syllabic stress points recovered and compared with the metronome. In the case of the digits the mean error of prediction was −9 ms (N=48). Given that the model has an internal delay of about 50 ms (20 Hz), this is good agreement. The variance was quite high, 1609 ms, but actually...

The model [N. Todd, ‘‘A theory of the principal monaural pathway I. Pitch and time perception,’’ J. Acoust. Soc. Am. 99, 2491(A) (1996)] proposes also to account for the psychophysics of time by the interaction of a ‘‘sensory memory,’’ in the form of the collective response of cortical bandpass cells, and a ‘‘long‐term memory,’’ which is the result of learning in a cortical neural network. Since the time constants of the cortical cells are relatively long, their response to an inferior collicular (ICC) input lasts long after the ICC input has ceased (and hence the form of the sensory memory above). The cortical amplitude modulation response is roughly a Laplace transform of the ICC response envelope, and so is highly dependent on envelope duration. The cortical response thus captures both the pitch and durational properties of the tone. It is shown how this model is able to provide an account of two fundamental aspects of auditory temporal processing: (1) the psychophysical law for time interval discrimin...

Over the last 2 decades much work has been done on the nature of expression in musical performance. Expression, which takes the form of subtle transformations of timing, dynamics, pitch and timbre, etc., may be measured fairly accurately when a score‐based representation is available. It is generally agreed that the major function of expression in these cases is the communication of the performer’s structural/motional interpretation of the piece being performed. Indeed, it is possible to demonstrate that the greater part of expressive variation may be accounted for by the two major components of rhythmic structure, namely, meter and grouping. However, little attention has been paid as to how the perception of rhythm may be influenced by expression. In this paper the perception of rhythm in expressive performance is reviewed from the point of view of a recent auditory‐motor theory of rhythm perception [N. P. Todd and G. J. Brown, ‘‘Vizualization of rhythm, time and metre,’’ Artificial Intelligence Rev. (19...

Previously, a method of performance analysis has been described which is based on a multitimescale decomposition of the acoustic signal [Todd, ‘‘Wavelet analysis of rhythm,’’ J. Acoust. Soc. Am. 93, 2290(A) (1993)]. This analysis is sensitive to a variety of expressive devices employed by musical performers including tempo, dynamics, and articulation. The output of the analysis is a ‘‘rhythmogram’’ which resembles the time‐span reduction of the theory of Lerdahl and Jackendoff. In this paper, the results of some analyses of real and synthetic performances are presented [Todd, ‘‘The dynamics of dynamics: a model of musical expression,’’ J. Acoust. Soc. Am. 91, 3540–3550 (1992)].

Recently a computational model of prosody perception based on a multi‐time‐scale decomposition of the output from a cochlear model has been demonstrated [Todd and Brown, ‘‘A multi‐scale auditory model of prosodic perception,’’ Proceedings of the International Conference on Spoken Language Processing (1994)]. This model determines the temporal grouping and prominence of syllables from a speech signal. In this paper we present evidence to show that the model is able to carry out a complete segmentation of a speech signal, from the level of individual phonemes and phoneme clusters up to the phrase and utterance level. Implications for speech recognition are discussed.

There is now considerable evidence that auditory processing involves two parallel pathways. The first, a phasic pathway, which specializes in temporal processing, involves a hierarchy of temporal periodicities ranging from pitch periods to rhythmic periods. The second is a tonal pathway, which specializes in the spectral processing. Here a computational model of the phasic process based on a multi‐timescale decomposition of the output from a cochlear model is presented. This model determines (1) the locatation of the syllables of a speech signal, (2) the hierarchical grouping of the syllables, and (3) the stress of the syllables. In order to test this model its output was compared with that of conventional metrical phonology for a variety of utterances.

In this paper a new method of performance analysis is described based on a multi‐timescale decomposition of the acoustic signal. This analysis is sensitive to a variety of expressive devices employed by musical performers including tempo, dynamics, and articulation. The output of the analysis is a hierarchical tree structure that resembles the time‐span reduction of the theory of Lerdahl and Jackendoff. Analyses of percussive performance are demonstrated for a variety of percussive signals. The model that forms the basis of the analytic method also provides an account of the phenomena of rhythmic streaming whereby a sequence of events differentiated only by loudness can form coherent streams.

The application of wavelet analysis to rhythm is described. Rather than carry out the multiscale decomposition on the sound signal itself this method decomposes the sound energy flux [N. P. McAngus Todd, J. Acoust. Soc. Am. 92 (A), 2380 (1992)], This enables the analysis of rhythmic phenomena that have frequencies several orders of magnitude below pitch and are not represented in the sound signal. The decomposition is carried out by a logarithmically spaced Laguerre series approximation Gaussian filter bank. The analog design was discretized to a cascaded IIR structure using a bilinear transformation. This has the following advantages over the usual FIR approach: (a) plausibility of interpretation as a perceptual model since the ideal Gaussian is not physically realizable; (b) a smaller number of parameters; (c) smaller delay times than the Gaussian ideal; (d) avoidance of the need for down/up sampling. Two complementary structural components are obtained from the projection in the frequency‐time plane of...

The rhythm of a piece of music or a poem can be described by hierarchical tree structure. A new application of wavelet analyses to rhythm in music and speech is described whose result is a hierarchical structure [N. P. McAngus Todd, J. Acoust. Soc. Am. 92, 2380 (A) (1992)]. This structure has two complementary components that are obtained from the projection of loci, in the frequency‐time plane, of zero crossings of the slope of the energy flux surface in the energy‐time plane: (a) a segmentation structure, corresponding to positive second derivatives at the zero crossings; (b) a stress structure, corresponding to negative second derivatives at the zero crossings. The value of the stress is given by the second derivative, which has a ‘‘Mexican hat’’ impulse response. These structures resemble the grouping/metrical structures of the theory of F. Lerdahl and R. Jackendoff [A Generative Theory of Tonal Music (MIT, Cambridge, MA, 1983)]. Example analyses of musical and speech performance are demonstrated for ...

Research on musical performance has shown that expressive variations in tempo, dynamics, etc. not explioit in a score can be acounted for by structural factors such as phrasing and rhythm [N. P. McAngus Todd, J. Acoust. Soc. Am. 91, 3540–3550 (1992)]. There now exists a number of models that attempt to recover rhythmic structure from a performed sequence [H. C. Longuet‐Higgins, Nature 263, 646–653 (1976); P. Desain, Music Percept. 9(4), 439–454 (1992)]. However, they do not address the question of how the human auditory system might detect temporal events before such higher level processing. A low‐level representation is proposed, analogous to the primal sketch in vision [D. Marr and E. Hildreth, Proc. R. Soc. London Ser. B 207, 187–217 (1980)], enocoding intensity changes at different time scales via an array of filters tuned at very low infrasonic frequencies. This representation could act as a front end to perceptual models. Various outputs from an analog implementation of the filter array are demonstr...