As suggested in the Blind Watchmaker, human selection can be a remarkable source of information for guiding a genetic algorithm when objective cost functions are unknown. Properly harnessing such input, however, requires an understanding of the numbers humans produce as well as the limitations humans face when performing extensive judgment tasks. The Interactive Augmented Genetic Algorithm (IAGA) modifies both the procedural and algorithmic components of Interactive Genetic Algorithms to better match the human-selection process. experimental results show that cochlear implant recipients are successful in using the IAGA to select processing parameters to improve their perception of music.

Binaural displays for immersive listening must model realistic acoustic environments, multiple sound sources, and accommodate source and head motion. Many displays accomplish this by convolving collections of spatially distributed point sources with head-related transfer functions (HRTFs). The computational load of such a system scales linearly with the number of HRTFs modeled by the display. Realistic scenes often require a large number of HRTFs, and this framework is computationally burdensome. We propose a method that significantly eases this load by formulating the HRTF filter array as a state-space system. Three state-space architectures are explored. The performance of the most general architecture is found to suffer due to the interaural time delay (ITD). This problem may be circumvented with two alternative architectures, although the ideal choice depends on the specific display application. For each architecture, two order-reduction techniques are explored. Both techniques are based on the Hankel operator; one is ad hoc and simple to implement whereas the other is optimal in the Hankel norm. The two methods yield similar auditory performance, although the optimal method may be desirable for HRTF approximation.

Binaural displays for immersive listening must model reverberant acoustic environments, multiple sound sources, and compensate for head motion. Many displays accomplish this by convolving collections of spatially distributed sources with head-related transfer functions (HRTFs). In the absence of reflections, the computational load scales linearly with the number of sources; however, when reflections are present, the load scales exponentially such that the binaural display reaches the limit of available processing power for even relatively sparse acoustic scenes. We propose a method that significantly eases this exponential growth by formulating the HRTF filter array as a MIMO state-space system. Two MIMO architectures are explored; the relative merits of each are found to depend on the specific application. Hankel-optimal methods are found to be a good choice for model reduction, and yield displays with superior approximation quality relative to conventional FIR filter arrays of equal computational complexity.

Much research in music information retrieval has focused on query-by-humming systems, which search melodic databases using sung queries. The database retrieval aspect of such systems has received considerable attention, but query processing and the melodic representation have not been examined as carefully. Common methods for query processing are based on musical intuition and historical momentum rather than specific performance criteria; existing systems often employ rudimentary note segmentation or coarse quantization of note estimates. In this work, we examine several alternative query processing methods as well as quantized melodic representations. One common difficulty with designing query-by-humming systems is the coupling between system components. We address this issue by measuring the performance of the query processing system both in isolation and coupled with a retrieval system. We first measure the segmentation performance of several note estimators. We then compute the retrieval accuracy of an experimental query-by-humming system that uses the various note estimators along with varying degrees of pitch and duration quantization. The results show that more advanced query processing can improve both segmentation performance and retrieval performance, although the best segmentation performance does not necessarily yield the best retrieval performance. Further, coarsely quantizing the melodic representation generally degrades retrieval accuracy.

With the growing prevalence of large databases of multimedia content, methods for facilitating rapid browsing of such databases or the results of a database search are becoming increasingly important. However, these methods are necessarily media dependent. We present a system for producing short, representative samples (or "audio thumbnails") of selections of popular music. The system searches for structural redundancy within a given song with the aim of identifying something like a chorus or refrain. To isolate a useful class of features for performing such structure-based pattern recognition, we present a development of the chromagram, a variation on traditional time-frequency distributions that seeks to represent the cyclic attribute of pitch perception, known as chroma. The pattern recognition system itself employs a quantized chromagram that represents the spectral energy at each of the 12 pitch classes. We evaluate the system on a database of popular music and score its performance against a set of "ideal" thumbnail locations. Overall performance is found to be quite good, with the majority of errors resulting from songs that do not meet our structural assumptions.

In this paper, we present a spectrum-based system for singer identification that operates for the ideal case in which audio samples contain only the singer's voice. Our method begins with the computation of a robust estimate of the spectral envelope called the composite transfer function (CTF). The CTF is derived from the instantaneous amplitude and frequency of the sinusoidal partials which make up the vocal signal. Unlike traditional source-filter theory , the CTF does not explicitly separate the spectral characteristics of the vocal source and the vocal tract filter. The principal components of the CTFs are used as features for a quadratic classifier to identify singers. The approach is validated on a database containing samples from twelve classically trained singers. In cross validation experiments, test set accuracies of approximately 95% are found for a baseline case. The classifier's performance is not degraded when different vowels are included in classifier training and evaluation. Restricting the frequency range of the CTFs and using a test set containing samples extracted from solo performances of Italian arias reduces the test set accuracy to 70-80%.

We describe a system for integrating a genetic algorithm (GA) with perceptual feedback to perform an efficient search in a perceptual space. The main system components are an efficient method for estimating perceptual rank order and genetic operators that take advantage of the types of parameters found in certain classes of audio processing systems. Preference judgments are used, resulting in a lightweight user interface. The application to subjectively fitting a portable hearing aid based solely on binary feedback is discussed. An experiment was conducted using eight normal and eight hearing impaired subjects. Three parameters were varied to control cancellation of acoustic feedback. The GA worked well for fitting this system, as indicated by both objective and subjective measures. In addition, users had greatly differing preferences for feedback cancellation parameters and these preferences did not change much when subjects were retested.

We demonstrate that a genetic algorithm (GA) can efficiently generate accurate low-order pole-zero approximations of head-related transfer functions (HRTFs) from measured impulse responses by minimizing a logarithmic error criterion. This approach is much simpler and comparable or superior in efficiency to competing search algorithms. We build on previous work in low-order HRTF approximation. By applying the GA, we converge to solutions of equal quality in about 30 s compared to over 20 min. This work develops a basic steady-state GA using a pole-zero filter design problem as an illustrative example. We propose a domain-appropriate error measure. We then apply the algorithm to designing filters to approximate measured HRTFs. Detailed performance measurements are presented. In the appendix, we propose a widely applicable population variation metric. A lower bound is developed for this metric and is used to detect convergence.

This paper presents an alternative visualization tool for head-related transfer functions (HRTFs) which represents HRTF data sets as magnitude spatial frequency response surfaces. Qualitative analysis of HRTF data is easier in the spatial domain than in the magnitude frequency domain and allows quick comparisons between different subjects' HRTF sets. In addition, these surfaces exhibit many well-known HRTF-related psychophysical phenomena due to head, torso, and pinna filtering. Finally, these surfaces suggest an interpolation algorithm by which directional transfer functions (DTFs) corresponding to arbitrary spatial locations can be computed from existing DTF measurements at known locations.

Pole-zero approximations of human free-field-to-eardrum transfer functions-i.e., head-related transfer functions (HRTFs)-can be used to synthesize binaural acoustic fields over headphones with fewer parameters, and therefore, lower computational complexity than is typically achieved by all-zero approximations. While standard pole-zero designs minimize the squared difference between the desired and designed frequency responses, the synthesis of binaural acoustic fields is more likely to be sensitive to relative, rather than absolute, differences in the HRTF approximations. The paper proposes a new design algorithm that minimizes an error criterion based on complex logarithms. As a result, both log-magnitude and phase errors are minimized. Slight modifications to the proposed algorithm result in one for which the phase error can be ignored. A comparison of this modified algorithm to others that have been developed to minimize the log-magnitude error shows implementation advantages without sacrificing performance. In achieving comparable errors, application of the proposed algorithm to the HRTF approximation results in pole-zero model orders smaller than those required for approximating the HRTFs using a least-squares error criterion.

The major time and frequency analysis methods that have been applied to music processing are traced and application areas described. Techniques are examined in the context of Cohen's class, facilitating comparison and the design of new approaches. A trumpet example illustrates most techniques. The impact of different analysis methods on pitch and timbre examination is shown. Analyses spanning Fourier series and transform, pitch synchronous analysis, heterodyne filter, short-time Fourier transform (STFT), phase vocoder, constant-Q and wavelet transforms, the Wigner (1932) distribution, and the modal distribution are all covered. The limitations of windowing methods and their reliance on steady-state assumptions and infinite duration sinusoids to define frequency and amplitude are detailed. The Wigner distribution, in contrast, uses the analytic signal to define instantaneous frequency and power parameters. The modal distribution is shown to be a linear transformation of the Wigner distribution optimized for estimating those parameters for a musical signal model. Application areas consider analysis, resynthesis, transcription, and visualization. The more stringent requirements for time-frequency (TF) distributions in these applications are compared with the weaker requirements found in speech analysis and highlight the need for further theoretical research.

The subjective tuning of multidimensional systems is considered. Listener preference for certain attributes of sound over others involves multiple options, costs, and payoffs, and may require complex strategies to arrive at an optimal solution. When the complexity of such strategies becomes excessive, listeners adopt simpler strategies that may lead to poor, but achievable, solutions. Active sensory tuning (AST) is aimed at providing the listener an efficient search strategy that yields good solutions within reasonable time bounds. Within an engineering context, AST provides a direct link between the human user and the engineer's design parameters whereby the user can tune the design to their desired goal. This contrasts with much simpler, techniques whereby the user can rank their preference but the engineer must interpret such rankings with respect to the selected parameters.

An increasing number of topographical studies find that natural surfaces possess power-law roughness spectra. Power-law spectra introduce unique difficulties in the spectral estimation process. The authors describe how an improper window choice allows leakage that yields a spectral estimate that is insensitive to the spectral slope. In addition, the commonly used Fourier-based spectral estimates have higher variances than other available estimators. Higher variance is particularly problematic when data records are short, as is often the case in remote sensing studies. The authors show that Capon's spectral estimator has less variance than Fourier-based estimators and measures the spectral slope more accurately. The authors also show how estimates of a 2D roughness spectrum can be obtained from estimates of the 1D spectrum for the isotropic power-law case.>

A time-sequenced adaptive filter has been developed to predict beats of the intracardiac electrogram (i.e. signals recorded within the heart) for purposes of morphological comparison by implantable devices. After adapting to electrogram morphology during normal sinus rhythm, the filter exhibits an abrupt increase in mean-squared error at the onset of ventricular tachycardia, an abnormal and often life-threatening cardiac rhythm. This marked increase in error can be used to identify the abnormal rhythm and may lead to faster medical treatment than that provided by many existing schemes for arrhythmia detection.>

An algorithm for estimating the frequencies of two sinusoids in white noise is presented. The method uses the eigenvectors of the estimated autocorrelation matrix (ACM) to yield the signal vectors. Since the signal and the eigenvectors span the same space, one can be obtained from the other through a linear transformation. The coefficients of the linear combination are derived as the eigenvectors of a 2*2 matrix whose elements are a function of the signal power, ACM dimension, and the angle between the two signal vectors. A method for estimating the unknown angle between the signal vectors is described. Knowledge of this angle completely determines the 2*2 matrix whose eigenvectors then yield the signal vectors. A periodogram can now be used to estimate the unknown frequencies from each signal vector. The method has the attractive property of separating a two-sinusoid problem into two single-sinusoid problems. Simulation results that describe the performance of this algorithm at low signal-to-noise ratios are presented.>

A nonasymptotic theoretical analysis for the performance of eigenstructure spectral methods for direction-of-arrival (DOA) estimation is proposed. The performance analysis is based on perturbation analysis of covariance eigenstructure and directions-of-arrival. The performance is expressed in terms of signal-to-noise ratio (SNR), the number of snapshots, and other parameters including source locations, sensor locations, and sensor gains without utilizing asymptotic conditions. From the expression, it is proved that the absolute bias and the variance of MUSIC direction-of-arrival estimation is inversely proportional to signal-to-noise ratio and to the number of snapshots. Simulations confirm the inverse proportional relationship under conditions including a small-number of snapshots.>

Multiple-sensor processing is considered, and a unified method for representing multiple-sensor data is developed. When resolution varies between sensors, such a multiple-sensor system can be viewed as samples of a scale-space signal representation. It is demonstrated that if the spatial transfer function of the sensors are Gaussian, then scale-space filtering can be used to recover small-scale (fine-resolution) information through extrapolation in scale. As an example of multiple-sensor processing, multispectral processing of remote sensing, in which images of surface scenes are simultaneously generated at different (center) frequencies, is considered. The fingerprints of extrapolated signals approximate the actual multispectral fingerprints at small scales and can be used when the multispectral fingerprints are not available.>

Major components of scale-space theory are Gaussian filtering, and the use of zero-crossing thresholders and Laplacian operators. Properties of scale-space filtering may be useful for data analysis in multiresolution machine-sensing systems. However, these systems typically violate the Gaussian filter assumption, and often the data analyses used (e.g. trend analysis and classification) are not consistent with zero-crossing thresholders and Laplacian operators. The authors extend the results of scale-space theory to include these more general conditions. In particular, it is shown that relaxing the requirement of linear scaling allows filters to have non-Gaussian spatial characteristics, and that relaxing of the scale requirements (s to 0) of the impulse response allows the use of scale-space filters with limited frequency support (i.e. bandlimited filters). Bandlimited scale-space filters represent an important extension of scale-space analysis for machine sensing.>

The problem of compact representations of planar curves using autoregressive (AR) models is discussed. The circular AR representation has been generalized to include the set of open curves. It is shown that the well-known covariance and autocorrelation methods used in speech processing fit naturally into the framework of estimating the AR coefficients for closed contours and for open planar curves, respectively. Application to error-free data compression of handwritten characters has yielded compression ratios as high as 2 to 3. Experiments have been carried out to evaluate the applicability of this method to the problem of interpolation of missing points of a curve.>

The Toeplitz form for an observed covariance matrix is investigated. A least-squares solution is found with respect to a given inner product space. The use of this form as a means for enhancing the Pisarenko Harmonic Decomposition method of estimating the location of line components in a spectrum is reported. The more general problem of finding a rank-p Toeplitz form for an observed covariance matrix is reviewed and an iterative solution is proposed. MUSIC-based estimates of line components are improved at lower SNRs by using the rank-p Toeplitz form instead of the original rank-p covariance matrix.

Spectral estimation for a class of smooth, broad spectra generated by the Turbulent Boundary Layer is discussed in terms of spectral fidelity and sensitivity to statistical error. When the number of sensors is limited, which is often the case in array problems, MLM-based estimators exhibit better performance than all-pole models. These performance differences underscore the fact that for certain classes of spectra high-resolution and high-fidelity are competing criteria.