AU2002240461A1

AU2002240461A1 - Comparing audio using characterizations based on auditory events

Info

Publication number: AU2002240461A1
Application number: AU2002240461A
Authority: AU
Inventors: Brett G. Crockett; Michael J. Smithers
Original assignee: Dolby Laboratories Licensing Corp
Current assignee: Dolby Laboratories Licensing Corp
Priority date: 2001-05-25
Filing date: 2002-02-22
Publication date: 2003-05-08
Anticipated expiration: 2022-02-22

Description

DESCRIPTION

Comparing Audio Using Characterizations Based on Auditory Events

TECHNICAL FIELD The invention relates to audio signals. More particularly, the invention relates to characterizing audio signals and using characterizations to determine if one audio signal is derived from another audio signal or if two audio signals are derived from the same audio signal.

BACKGROUND ART

The division of sounds into units perceived as separate is sometimes referred to as "auditoiy event analysis" or "auditoiy scene analysis" ("ASA"). An extensive discussion of auditoiy scene analysis is set forth by Albert S. Bregman in his book Auditory Scene Analysis - The Perceptual Organization of Sound, Massachusetts Institute of Technology, 1991, Fourth printing, 2001, Second MIT Press paperback edition. In addition, United States Patent 6,002,776 to Bhadkamkar, et al, December 14, 1999 cites publications dating back to 1976 as "prior ait work related to sound separation by auditoiy scene analysis." However, the Bhadkamkar, et al patent discourages the practical use of auditoiy scene analysis, concluding that "[tjechniques involving auditoiy scene analysis, although interesting from a scientific point of view as models of human auditoiy processing, are currently far too computationally demanding and specialized to be considered practical techniques for sound separation until fundamental progress is made."

Bregman notes in one passage that "[w]e hear discrete units when the sound changes abruptly in timbre, pitch, loudness, or (to a lesser extent) location in space." (Auditory Scene Analysis - The Perceptual Organization of Sound, supra at page 469). Bregman also discusses the perception of multiple simultaneous sound streams when, for example, they are separated in frequency.

There are many different methods for extracting characteristics or features from audio. Provided the features or characteristics are suitably defined, their extraction can be performed using automated processes. For example "ISO/IEC JTC 1/SC 29/WG 11" (MPEG) is currently standardizing a variety of audio descriptors as part of the MPEG-7 standard. A common shortcoming of such methods is that they ignore ASA. Such methods seek to measure, periodically, certain "classical" signal processing parameters such as pitch, amplitude, power, harmonic structure and spectral flatness. Such parameters, while providing useful information, do not analyze and characterize audio signals into elements perceived as separate according to human cognition.

Auditory scene analysis attempts to characterize audio signals in a manner similar to human perception by identifying elements that are separate according to human cognition. By developing such methods, one can implement automated processes that accurately perform tasks that heretofore would have required human assistance.

The identification of separately perceived elements would allow the unique identification of an audio signal using substantially less information than the full signal itself. Compact and unique identifications based on auditoiy events may be employed, for example, to identify a signal that is copied from another signal (or is copied from the same original signal as another signal).

DISCLOSURE OF THE INVENTION A method is described that generates a unique reduced-information characterization of an audio signal that may be used to identify the audio signal. The characterization may be considered a "signature" or "fingerprint" of the audio signal. According to the present invention, an auditoiy scene analysis (ASA) is performed to identify auditory events as the basis for characterizing an audio signal. Ideally, the auditory scene analysis identifies auditory events that are most likely to be perceived by a human listener even after the audio has undergone processing, such as low bit rate coding or acoustic transmission through a loudspeaker. The audio signal may be characterized by the boundary locations of auditoiy events and, optionally, by the dominant frequency subband of each auditoiy event. The resulting information pattern constitutes a compact audio fingerprint or signature that may be compared to one or more other such audio fingerprints or signatures. A detennination that at least a portion of the respective signatures are the same (to a desired degree of confidence) indicates that the related portions of the audio signals from which the respective signatures were derived are the same or were derived from the same audio signal. The auditory scene analysis method according to the present invention provides a fast and accurate method of comparing two audio signals, particularly music, by comparing signatures based on auditory event information. ASA extracts information or features underlying the perception of similarity, in contrast to traditional methods of feature extraction that extract features less fundamental to perceiving similaiities between audio signals (such as pitch amplitude, power, and harmonic structure). The use of ASA improves the chance of finding similarity in material that has undergone significant processing, such as low bit coding or acoustic transmission through a loudspeaker.

Although in principle the invention may be practiced either in the analog or digital domain (or some combination of the two), in practical embodiments of the invention, audio signals are represented by samples in blocks of data and processing is done in the digital domain.

Referring to FIG. 1 A, auditoiy scene analysis 2 is applied to an audio signal in order to produce a "signature" or "fingerprint," related to that signal. In this case, there are two audio signals of interest. They may be similar in that one may be derived from the other or both may have been previously derived from the same original signal, but this is not known in advance. Thus, auditoiy scene analysis is applied to both signals. For simplicity, FIG. 1 A shows only the application of ASA to one signal. As shown in FIG. IB, the signatures for the two audio signals, Signature 1 and Signature 2, are applied to a conelator or conelation function 4 that generates a conelation score. A user may set a minimum conelation score as providing a desired degree of confidence that at least a portion of the two signatures are the same. In practice, the two signatures may be stored data. In one practical application, one of the signatures may be derived, for example, from an unauthorized copy of a musical work and the other signature may be one of a large number of signatures in a database (each signature being derived from a copyright owner's musical work) against which the unauthorized copy signature is compared until a match, to a desired degree of confidence, if any, is obtained. This may be conducted automatically by a machine, the details of which are beyond the scope of the present invention.

Because the signatures are representative of the audio signals but are substantially shorter (i.e., they are more compact or have fewer bits) than the audio signals from which they were derived, the similarity of the two signatures (or lack thereof) can be detemiined much faster than it would take to detemiine the similaiity between the audio signals.

Further details of FIGS. 1 A and IB are set forth below. In accordance with aspects of the present invention, a computationally efficient process for dividing audio into temporal segments or "auditory events" that tend to be perceived as separate is provided. A powerful indicator of the beginning or end of a perceived auditoiy event is believed to be a change in spectral content. In order to detect changes in timbre and pitch (spectral content) and, as an ancillary result, certain changes in amplitude, the audio event detection process according to an aspect of the present invention detects changes in spectral composition with respect to time. Optionally, according to a further aspect of the present invention, the process may also detect changes in amplitude with respect to time that would not be detected by detecting changes in spectral composition with respect to time.

In its least computationally demanding implementation, the process divides audio into time segments by analyzing the entire frequency band of the audio signal (full bandwidth audio) or substantially the entire frequency band (in practical implementations, band limiting filtering at the ends of the spectrum are often employed) and giving the greatest weight to the loudest audio signal components. This approach takes advantage of a psychoacoustic phenomenon in which at smaller time scales (20 msec and less) the ear may tend to focus on a single auditoiy event at a given time. This implies that while multiple events may be occumng at the same time, one component tends to be perceptually most prominent and may be processed individually as though it were the only event taking place. Taking advantage of this effect also allows the auditoiy event detection to scale with the complexity of the audio being processed. For example, if the input audio signal being processed is a solo instrument, the audio events that are identified will likely be the individual notes being played. Similarly for an input voice signal, the individual components of speech, the vowels and consonants for example, will likely be identified as individual audio elements. As the complexity of the audio increases, such as music with a drumbeat or multiple instruments and voice, the auditoiy event detection identifies the most prominent (i.e., the loudest) audio element at any given moment.

Alternatively, the "most prominent" audio element may be detemiined by taking hearing threshold and frequency response into consideration.

Optionally, according to further aspects of the present invention, at the expense of greater computational complexity, the process may also take into consideration changes in spectial composition with respect to time in discrete frequency bands (fixed or dynamically detemiined or both fixed and dynamically dete iined bands) rather than the full bandwidth. This alternative approach would take into account more than one audio stream in different frequency bands rather than assuming that only a single stream is perceptible at a particular time. Even a simple and computationally efficient process according to an aspect of the present invention for segmenting audio has been found useful to identify auditoiy events.

An auditoiy event detecting process of the present invention may be implemented by dividing a time domain audio wavefonn into time intervals or blocks and then conveiting the data in each block to the frequency domain, using either a filter bank or a time-frequency transfonnation, such as a Discrete Fourier Transfoiin (DFT) (implemented as a Fast Fourier Transforai (FFT) for speed). The amplitude of the spectral content of each block may be nonnalized in order to eliminate or reduce the effect of amplitude changes. The resulting frequency domain representation provides an indication of the spectial content (amplitude as a function of frequency) of the audio in the particular block. The spectral content of successive blocks is compared and a change greater than a threshold may be taken to indicate the temporal start or temporal end of an auditoiy event.

In order to minimize the computational complexity, only a single band of frequencies of the time domain audio wavefonn may be processed, preferably either the entire frequency band of the spectrum (which may be about 50 Hz to 15 kHz in the case of an average quality music system) or substantially the entire frequency band (for example, a band defining filter may exclude the high and low frequency extremes). Preferably, the frequency domain data is noiirialized, as is described below.

The degree to which the frequency domain data needs to be noπnalized gives an indication of amplitude. Hence, if a change in this degree exceeds a predeteπnined threshold, that too may be taken to indicate an event boundary. Event start and end points resulting from spectral changes and from amplitude changes may be ORed together so that event boundaiies resulting from both types of change are identified. In practical embodiments in which the audio is represented by samples divided into blocks, each auditoiy event temporal start and stop point boundary necessarily coincides with a boundary of the block into which the time domain audio waveform is divided. There is a trade off between real-time processing requirements (as larger blocks require less processing overhead) and resolution of event location (smaller blocks provide more detailed infoimation on the location of auditoiy events).

As a further option, as suggested above, but at the expense of greater computational complexity, instead of processing the spectial content of the time domain wavefomi in a single band of frequencies, the spectrum of the time domain wavefomi prior to frequency domain conversion may be divided into two or more frequency bands. Each of the frequency bands may then be converted to the frequency domain and processed as though it were an independent channel. The resulting event boundaries may then be ORed together to define the event boundaiies for that channel. The multiple frequency bands may be fixed, adaptive, or a combination of fixed and adaptive. Tracking filter techniques employed in audio noise reduction and other aits, for example, may be employed to define adaptive frequency bands (e.g., dominant simultaneous sine waves at 800 Hz and 2 kHz could result in two adaptively-dete ined bands centered on those two frequencies).

Other techniques for providing auditoiy scene analysis may be employed to identify auditoiy events in the present invention.

DESCRD7TION OF THE DRAWINGS

FIG. 1A is a flow chart showing the extraction of a signature from an audio signal in accordance with the present invention. The audio signal may, for example, represent music (e.g., a musical composition or "song").

FIG. IB is a flow chart illustrating the conelation of two signatures in accordance with the present invention.

FIG. 2 is a flow chart showing the extraction of audio event locations and the optional extraction of dominant subbands from an audio signal in accordance with the present invention.

FIG. 3 is a conceptual schematic representation depicting the step of spectral analysis in accordance with the present invention.

FIGS. 4 A and 4B are idealized audio wavefoims showing a plurality of audio event locations or event borders in accordance with the present invention. FIG. 5 is a flow chart showing in more detail the conelation of two signatures in accordance with the conelation 4 of FIG. 2 of the present invention.

FIGS. 6A-D are conceptual schematic representations of signals illustrating examples of signature alignment in accordance with the present invention. The figures are not to scale. In the case of a digital audio signal represented by samples, the horizontal axis denotes the sequential order of discrete data stored in each signature array.

BEST MODE FOR CARRYING OUT THE INVENTION

In a practical embodiment of the invention, the audio signal is represented by samples that are processed in blocks of 512 samples, which coιτesρonds to about 11.6 msec of input audio at a sampling rate of 44.1 kHz. A block length having a time less than the duration of the shortest perceivable auditoiy event (about 20 msec) is desirable. It will be understood that the aspects of the invention are not limited to such a practical embodiment. The principles of the invention do not require arranging the audio into sample blocks prior to detennining auditoiy events, nor, if they are, of providing blocks of constant length. However, to minimize complexity, a fixed block length of 512 samples (or some other power of two number of samples) is useful for three primary reasons. First, it provides low enough latency to be acceptable for real-time processing applications. Second, it is a power-of-two number of samples, which is useful for fast Fourier tiansfonn (FFT) analysis. Third, it provides a suitably large window size to peifonn useful auditoiy scene analysis.

In the following discussions, the input signals are assumed to be data with amplitude values in the range [-1,+1].

Auditory Scene Analysis 2 (FIG. 1A) Following audio input data blocking (not shown), the input audio signal is divided into auditoiy events, each of which tends to be perceived as separate, in process 2 ("Auditoiy Scene Analysis") of FIG. 1 A. Auditoiy scene analysis may be accomplished by an auditoiy scene analysis (ASA) process discussed above. Although one suitable process for perfbπning auditoiy scene analysis is described in further detail below, the invention contemplates that other useful techniques for performing ASA may be employed.

FIG. 2 outlines a process in accordance with techniques of the present invention that may be used as the auditoiy scene analysis process of FIG. 1 A. The ASA step or process 2 is composed of three general processing substeps. The first substep 2-1 ("Perfoim Spectral Analysis") takes the audio signal, divides it into blocks and calculates a spectial profile or spectral content for each of the blocks. Spectral analysis transfonns the audio signal into the short-tenn frequency domain. This can be perfonned using any filterbank; either based on transforms or banks of band-pass filters, and in either linear or warped frequency space (such as the Bark scale or critical band, which better approximate the chaiacteristics of the human ear). ith any filterbank there exists a tradeoff between time and frequency. Greater time resolution, and hence shorter time intervals, leads to lower frequency resolution. Greater frequency resolution, and hence naιτower subbands, leads to longer time intervals. The first substep 2-1 calculates the spectial content of successive time segments of the audio signal. In a practical embodiment, described below, the ASA block size is 512 samples of the input audio signal (FIG.3). In the second substep 2- 2, the differences in spectial content from block to block are detennined ("Perform spectral profile difference measurements"). Thus, the second substep calculates the difference in spectial content between successive time segments of the audio signal. In the third substep 2-3 ("Identify location of auditoiy event boundaiies"), when the spectral difference between one spectral-profile block and the next is greater than a threshold, the block boundary is taken to be an auditoiy event boundary. Thus, the third substep sets an auditory event boundary between successive time segments when the difference in the spectial profile content between such successive time segments exceeds a threshold. As discussed above, a powerful indicator of the beginning or end of a perceived auditoiy event is believed to be a change in spectial content. The locations of event boundaiies are stored as a signature. An optional process step 2-4 ("Identify dominant subband") uses the spectral analysis to identify a dominant frequency subband that may also be stored as part of the signature.

In this embodiment, auditoiy event boundaiies define auditoiy events having a length that is an integral multiple of spectial profile blocks with a minimum length of one spectral profile block (512 samples in this example). In principle, event boundaiies need not be so limited. Either overlapping or non-overlapping segments of the audio may be windowed and used to compute spectial profiles of the input audio. Overlap results in finer resolution as to the location of auditoiy events and, also, makes it less likely to miss an event, such as a transient. However, as time resolution increases, frequency resolution decreases. Overlap also increases computational complexity. Thus, overlap may be omitted. FIG. 3 shows a conceptual representation of non- overlapping 512 sample blocks being windowed and transfom ed into the frequency domain by the Discrete Fourier Tiansfonn (DFT). Each block may be windowed and transfomied into the frequency domain, such as by using the DFT, preferably implemented as a Fast Fourier Tiansfonn (FFT) for speed. The following variables may be used to compute the spectral profile of the input block:

N = number of samples in the input signal

M = number of windowed samples used to compute spectral profile P = number of samples of spectial computation overlap Q = number of spectial windows/regions computed

In general, any integer numbers may be used for the variables above. However, the implementation will be more efficient if M is set equal to a power of 2 so that standard FFTs may be used for the spectral profile calculations. In a practical embodiment of the auditoiy scene analysis process, the parameters listed may be set to:

M = 512 samples (or 11.6 msec at 44.1 kHz) P = 0 samples (no overlap) The above-listed values were deteπnined experimentally and were found generally to identify with sufficient accuracy the location and duration of auditoiy events. However, setting the value of P to 256 samples (50% overlap) has been found to be useful in identifying some hard-to-find events. While many different types of windows may be used to minimize spectral artifacts due to windowing, the window used in the spectral profile calculations is an M-point Hanning, Kaiser- Bessel or other suitable, preferably non-rectangular, window. The above-indicated values and a Hanning window type were selected after extensive experimental analysis as they have shown to provide excellent results across a wide range of audio material. Non-rectangular windowing is prefened for the processing of audio signals with predominantly low frequency content. Rectangular windowing produces spectral artifacts that may cause incorrect detection of events. Unlike certain codec applications where an overall overlap/add process must provide a constant level, such a constraint does not apply here and the window may be chosen for characteristics such as its time/frequency resolution and stop-band rejection.

In substep 2-1 (FIG. 2), the spectrum of each M-sample block may be computed by windowing the data by an M-point Hanning, Kaiser-Bessel or other suitable window, converting to the fiequency domain using an M-point Fast Fourier Transform, and calculating the magnitude of the FFT coefficients. The resultant data is nonnalized so that the largest magnitude is set to unity, and the norøialized aιτay of M numbers is converted to the log domain. The array need not be converted to the log domain, but the conversion simplifies the calculation of the difference measure in substep 2-2. Furthemiore the log domain more closely matches the log domain amplitude nature of the human auditoiy system. The resulting log domain values have a range of minus infinity to zero. In a practical embodiment, a lower limit can be imposed on the range of values; the limit may be fixed, for example -60 dB, or be frequency-dependent to reflect the lower audibility of quiet sounds at low and veiy high frequencies. (Note that it would be possible to reduce the size of the array to M/2 in that the FFT represents negative as well as positive frequencies).

Substep 2-2 calculates a measure of the difference between the spectra of adjacent blocks. For each block, each of the M (log) spectral coefficients from substep 2-1 is subtracted from the coπesponding coefficient for the preceding block, and the magnitude of the difference calculated (the sign is ignored). These M differences are then summed to one number. Hence, for the whole audio signal, the result is an array of Q positive numbers; the greater the number the more a block differs in spectrum from the preceding block. This difference measure could also be expressed as an average difference per spectial coefficient by dividing the difference measure by the number of spectral coefficients used in the sum (in this case M coefficients).

Substep 2-3 identifies the locations of auditoiy event boundaiies by applying a threshold to the array of difference measures from substep 2-2 with a threshold value. When a difference measure exceeds a threshold, the change in spectrum is deemed sufficient to signal a new event and the block number of the change is recorded as an event boundary. For the values of M and P given above and for log domain values (in substep 2-1) expressed in units of dB, the threshold may be set equal to 2500 if the whole magnitude FFT (including the mirrored part) is compared or 1250 if half the FFT is compared (as noted above, the FFT represents negative as well as positive frequencies — for the magnitude of the FFT, one is the mmor image of the other). This value was chosen experimentally and it provides good auditory event boundary detection. This paiameter value may be changed to reduce (increase the threshold) or increase (decrease the threshold) the detection of events. The details of this practical embodiment are not critical. Other ways to calculate the spectral content of successive time segments of the audio signal, calculate the differences between successive tune segments, and set auditoiy event boundaiies at the respective boundaiies between successive time segments when the difference in the spectial profile content between such successive time segments exceeds a threshold may be employed. For an audio signal consisting of Q blocks (of size M samples), the output of the auditoiy scene analysis process of function 2 of FIG. 1 A is an array B(q) of information representing the location of auditoiy event boundaries where q - 0, 1, . . . , Q-l. For a block size of M = 512 samples, overlap of P = 0 samples and a signal- sampling rate of 44.1kHz, the auditoiy scene analysis function 2 outputs approximately 86 values a second. Preferably, the anay B(q) is stored as the signature, such that, in its basic fom , without the optional dominant subband frequency infonnation, the audio signal's signature is an array B(q) representing a string of auditoiy event boundaiies.

An example of the results of auditoiy scene analysis for two different signals is shown in FIGS. 4 A and 4B. The top plot, FIG. 4 A, shows the results of auditoiy scene processing where auditoiy event boundaiies have been identified at samples 1024 and 1536. The bottom plot, FIG. 4B, shows the identification of event boundaiies at samples 1024, 2048 and 3072. Identify dominant subband (optional) For each block, an optional additional step in the ASA processing (shown in FIG. 2) is to extract infonnation from the audio signal denoting the dominant frequency "subband" of the block (conversion of the data in each block to the frequency domain results in infonnation divided into frequency subbands). This block-based infonnation may be converted to auditory-event based infonnation, so that the dominant frequency subband is identified for every auditoiy event. This infonnation for every auditoiy event provides the conelation processing (described below) with further infonnation in addition to the auditoiy event boundary information.

The dominant (largest amplitude) subband may be chosen from a plurality of subbands, three or four, for example, that are within the range or band of frequencies where the human ear is most sensitive. Alternatively, other criteria may be used to select the subbands. The spectrum may be divided, for example, into three subbands. The prefened fiequency range of the subbands is: Subband 1 301 Hz to 560Hz

Subband 2 560Hz to 1938Hz

Subband 3 1938Hz to 9948Hz

To detemiine the dominant subband, the square of the magnitude spectrum (or the power magnitude spectrum) is summed for each subband. This resulting sum for each subband is calculated and the largest is chosen. The subbands may also be weighted prior to selecting the largest. The weighting may take the fonn of dividing the sum for each subband by the number of spectral values in the subband, or alternatively may take the fonn of an addition or multiplication to emphasize the importance of a band over another. This can be useful where some subbands have more energy on average than other subbands but are less perceptually important.

Considering an audio signal consisting of Q blocks, the output of the dominant subband processing is an anay DS(q) of infonnation representing the dominant subband in each block (q = 0, 1, . . . , Q-l). Preferably, the anay DS(q) is stored in fhe signature along with the anay B(q). Thus, with the optional dominant subband information, the audio signal's signature is two arrays B(q) and DS(q), representing, respectively, a stiing of auditoiy event boundaries and a dominant frequency subband within each block. Thus, in an idealized example, the two anays could have the following values (for a case in which there are three possible dominant subbands).

1 0 1 0 0 0 1 0 0 1 0 0 0 0 0 1 0 (Event Boundaries) 1 1 2 2 2 2 1 1 1 3 3 3 3 3 3 1 1 (Dominant Subbands)

In most cases, the dominant subband remains the same within each auditory event, as shown in this example, or has an average value if it is not unifonn for all blocks within the event. Thus, a dominant subband may be detemiined for each auditory event and the array DS(q) may be modified to provide that the same dominant subband is assigned to each block within an event.

Correlation The detennination of whether one signature is the same or similar to another stored signature may be accomplished by a conelation function or process. The conelation function or process compares two signatures to detennine their similarity. This may be done in two steps as shown in FIG. 5: a step 5-1 that removes or minimizes the effect of temporal shift or delay on the signatures, followed by a step 5-2 that calculates a measure of similaiity between the signatures.

The first-mentioned step 5-1 minimizes the effect of any delay between two signatures. Such delay may have been deliberately added to the audio signal or could be the result of signal processing and/or low bit rate audio coding. The output of this step is two modified signatures in a fonn suitable for calculation of a measure of their similarity.

The second-mentioned step 5-2 compares the two modified signatures to find a quantitative measure of their similaiity (a correlation score). This measure of similarity can then be compared against a threshold to detennine if the signatures are the same or different to a desired level of confidence. Two suitable conelation processes or functions are described. Either one of them or some other suitable conelation process or function may be employed as part of the present invention. First Conelation Process or Function Removal of Temporal Delay Effects This conelation function or process isolates a single region or portion from each of the signatures such that these two regions are the most similar portions in the respective signatures and have the same length. The isolated region could be the total overlapping regions between the two signatures, as shown in the examples in FIGS. 6A-D, or the isolated region could be smaller than the overlapping regions.

The prefened method uses the whole overlapping region from the two signatures. Some examples are shown in FIG. 6. The overlapping region for the two signatures could be a portion fiom the end of one signature and the beginning of the other signature (FIGS. 6B and 6C). If one of the signatures is smaller that the other, then the overlapping region between the two signatures could be all of the smaller signature and a portion of the larger signature (FIG. 6 A and 6D).

There are a number of different ways to isolate a common region from two arrays of data. A standard mathematical technique involves using the cross conelation to find a lag or delay measure between the arrays of data. When the beginning of each of two arrays of data is aligned, the lag or delay is said to be zero. When the beginning of each of two arrays of data is not aligned, the lag or delay is non-zero. The cross conelation calculates a measure for each possible lag or delay between the two airays of data: this measure is stored as an anay (the output of the cross conelation function). The lag or delay that represents the peak in the cross correlation array is considered to be the lag or delay of one anay of data with respect to the other. The following paragraphs expresses such a conelation method in a mathematical fonn. Let S, (length N, ) be an anay from Signature 1 and S, (length N₂) an array from Signature 2. First calculate the cross-conelation anay R_EιE (see, for example,

John G. Proakis, Dimitiis G. Manolakis, Digital Signal Processing: Principles, Algorithms, and Applications, Macmillan Publishing Company, 1992, ISBN 0-02- 396815-X). •W)= ∑S(n) "-l) 1 = 0+1+2,.... (1)

Preferably, tle cross-conelation is perfonned using standard FFT based techniques to reduce execution time.

Since both S, and S₂ are bounded, R_E>Ei has length N, +N₂ -1. Assuming S, and S₂ are similar, the lag / conesponding to the maximum element in R_EΆ represents the delay of S₂ relative to S, . l_peak=l for MAX(R_EIE2(1)) (2)

Since this lag represents the delay, the common spatial regions or spatially overlapping parts of signatures S, and S₂ are retained as S,^' and S₂ ; each having the same length, N₁₂.

Expressed as equations, the overlapping parts S[ and S₂ of the signatures S and S₂ are defined as:

₇ . _ „ ₍ Jl_peak ≤ n < l_pmk +MIN(N - l_peak ,N₂) m = n- l_peak l_peak ≥ 0 ^{λm) lW}i 0<n<MIN(N„N₂+l_peak) m = n l_peak<0

n l_peak≥0

^ΛM) l_peak l_peak<0

(3)

The length of S,^' and S₂ is:

_N

First Conelation Process or Function Similarity Measure This step compares the two signatures to find a quantitative measure of their sήnilaiity. The prefened method uses the coefficient of conelation (Eqn.5). This is a standard textbook method (William Mendenhall, Dennis D. Wackeiiy, Richard L. Scheaffer, Mathematical Statistics with Applications: Fourth Edition, Duxbuiy Press, 1990, ISBN 0-534-92026-8).

where σ, and σ₂ are the standard deviations of S,' and S₂ respectively.

The co variance of Sj and S₂ is defined as:

where /, and μ₂ are the means of Sj and S₂ respectively.

The coefficient of conelation, p , is in the range -1 < p < 1 where -1 and 1 indicate perfect conelation. Preferably, a threshold is applied to the absolute value of this measure to indicate a co ect match.

f TRUE abs(ρ) > threshold Match = \ ^K (7)

[FALSE abs(p) ≤ threshold ^J

In practice, the value of the threshold may be tuned (on a large training set of signatures) to ensure acceptable false rejection and detection rates.

The first conelation process or function is prefened for signatures that have large misaligmnent or delay, and for signatures in which the length of one signature is significantly smaller than the length of the other signature. Second Conelation Process or Function Removal of Temporal Delay Effects The second conelation process or function transfonns the signatures from their cunent temporal domain into a domain that is independent of temporal delay effects. The method results in two modified signatures that have the same length, such that they can be directly conelated or compared.

There are a number of ways to tiansfonn data in such a manner. The preferred method uses the Discrete Fourier Transfoπn (DFT). The DFT of a signal can be separated into magnitude and phase. A spatial shift or time delay of the signal (input to the DFT) alters the phase of the DFT but not the magnitude. Thus the magnitude of the DFT of a signal can be considered as a time-invariant representation of the signal.

This property of the DFT allows each of the two signatures to be transformed into a time-invariant representation. If both signatures have the same length, the magnitude DFT can be directly computed for each of the signatures and the results stored as the modified signatures. If the length of each of the signatures is different, then prior to calculating the DFT, either the longer signature can be truncated to have the same length as the shorter signature, or the shorter signature can be zero padded or extended to have the same length as the longer signature. The following paragraphs express the method in a mathematical fonn.

Let S, (length N, ) be an array from Signature 1 and S₂ (length N₂) an array from Signature 2. Firstly, the longer signature is truncated or the shorter signature zero padded such that both signatures have the same length, N_n . The transfomied signature arrays, Sj and S₂ , are created by taking the magnitude DFT as follows:

.e --j^jlrβcnl N k = 0,l,2,....,N_l2 -l (8)

ΛΓ,, -I

S₂(k) = ∑S₂(n).e-^J2MN» k = 0,l,2,....,N_u -l (9)

In practice, for each signature it is beneficial to subtract its mean prior to calculating the DFT. Some windowing may also be applied to the S, and S₂ signatures prior to taking the Discrete Fourier Tiansfonn, however in practice no particular windowing has been found to produce the best results.

Second Conelation Process or Function Similarity Measure This similaiity measure step compares the two signatures to find a quantitative measure of their similarity. The prefened method uses the coefficient of conelation (Eqn. 9). This is a standard textbook method (William Mendenhall, Dennis D. Wackeiiy, Richard L. Scheaffer, Mathematical Statistics with Applications: Fourth Edition, Duxbury Press, 1990, ISBN 0-534-92026-8).

where σ, and σ₂ are the standard deviations of Sj and S₂ respectively.

The covaiiance of Sj and S₂ is defined as:

where μ and μ₂ are the means of Sj and S₂ respectively.

The coefficient of conelation, p , is in the range - 1 < p ≤ 1 where -1 and 1 indicate perfect conelation. Preferably, a threshold is applied to the absolute value of this measure to indicate a conect match.

, , [ TRUE abs(p) > threshold ,_{1 1 s}

Match = ^y' , (11)

[FALSE abs(p) ≤ threshold '

In practice, the value of the threshold may be tuned (on a large ti'aining set of signatures) to ensure acceptable false rejection and detection rates. In practical applications, many signatures may be stored together to fonn a library of signatures representing "known" audio content. In this situation, the ability to discriminate between signatures can be improved by calculating a mean signature and subtracting this mean signature from each of two signatures under comparison.

For example, given a database containing W signatures, Sj, to S^_, , the mean signature is calculated as follows.

1 w~\ W = π ∑S_w' (k) k = 0,1,2,....,N_I2 -l (12)

When comparing two signatures (even if one of the signatures is not in the library) the mean signature is subtracted from both signatures prior to calculating the covariance (subsequently used in the coefficient of conelation). The covaiiance becomes:

ΛΛ,-1

∑[(S_i'(k)-S_M ^, _EAN(k))-μ,}.[(S₂(k)-S_M ^, _EAN(k))-μ₂] ov(Sj,S ) = -^ (13)

12 where μ, and μ₂ are the means of Sj -S_m ^' _AN and Sj -S_M ^' _EAN respectively.

The second conelation process or function is prefened for signatures that have small misaligmnent or delay, and for signatures where the lengths of the signatures are similar. It is also significantly faster than the first conelation process or function. However since some infonnation is inherently lost (by discarding the phase of the DFTs), it results in a slightly less accurate measure of similaiity.

Applications As briefly mentioned earlier, an application of this invention is searchable audio databases; for example a record company's library of songs. Signatures could be created for all the songs from the library and the signatures stored in a database. This invention provides a means for taking a song of unknown origin, calculating its signature and comparing its signature veiy quickly against all the signatures in the database to detennine the identity of the unknown song.

In practice, the accuracy of (or confidence in) the similarity measure is proportional to the size of the signatures being compared. The greater the length of the signatures, the greater the amount of data being used in the compaiison and hence the greater the confidence or accuracy in the similaiity measure. It has been found that signatures generated from about 30 seconds of audio provide for good discrimination. However the larger the signatures, the longer the time required to perform a compaiison. Conclusion

It should be understood that implementation of other vaiiations and modifications of the invention and its various aspects will be apparent to those skilled in the art, and that the invention is not limited by these specific embodiments described. It is therefore contemplated to cover by the present invention any and all modifications, vaiiations, or equivalents that fall within the true spirit and scope of the basic underlying principles disclosed and claimed herein.

The present invention and its various aspects may be implemented as software functions peifonned in digital signal processors, programmed general-purpose digital computers, and/or special puipose digital computers. Interfaces between analog and digital signal streams may be peifonned in appropriate haidware and/or as functions in software and/or finnwaie.

Claims

1. A method for detennining if one audio signal is derived fiom another audio signal or if two audio signals are derived from the same audio signal, comprising comparing reduced-infonnation characterizations of said audio signals, wherein said reduced-infonnation characterizations are based on auditoiy scene analysis.

2. The method of claim 1 wherein said comparing includes removing from the chaiacterizations or minimizing in the characterizations the effect of temporal shift or delay on the audio signals, calculating a measure of similaiity, and comparing the measure of similaiity against a threshold.

3. The method of claim 2 wherein said removing identifies a portion in each of said characterizations, such that the respective portions are the most similar portions in the respective chaiacterizations and the respective portions have the same length.

4. The method of claim 3 wherein said removing identifies a portion in each of said characterizations by perforating a cross-conelation.

5. The method of claim 4 wherein said calculating calculates a measure of sήnilarity by calculating a coefficient of conelation of the identified portion in each of said characterizations.

6. The method of claim 2 wherein said removing transfomis the characterizations into a domain that is independent of temporal delay effects.

7. The method of claim 6 wherein said removing transfomis the characterizations into the frequency domain.

8. The method of claim 7 wherein said calculating calculates a measure of similarity by calculating a coefficient of conelation of the identified poition in each of said characterizations.

9. The method of any one of claims 1-8 wherein one of said characterizations is a characterization fiom a libraiy of chaiacterizations representing known audio content.

10. The method of claim 9 further comprising subtracting a mean of the characterizations in said libraiy from both characterizations after said removing and prior to said comparing.

11. The method of any one of claims 1-10 wherein said reduced-infonnation characterizations based on auditoiy scene analysis are arrays of infonnation representing at least the location of auditoiy event boundaiies.

12. The method of claim 11 wherein said auditoiy event boundaiies are detemiined by calculating the spectral content of successive time segments of said audio signal, calculating the difference in spectial content between successive time segments of said audio signal, and identifying an auditoiy event boundary as the boundary between successive time segments when the difference in the spectral content between such successive tune segments exceeds a threshold.

13. The method of claim 12 or claim 13 wherein said arrays of infonnation also represent the dominant frequency subband of each of said auditoiy events.