US 20090265174 A9
The present invention provides an innovative technique for rapidly and accurately determining whether two audio samples match, as well as being immune to various kinds of transformations, such as playback speed variation. The relationship between the two audio samples is characterized by first matching certain fingerprint objects derived from the respective samples. A set (230) of fingerprint objects (231,232), each occurring at a particular location (242), is generated for each audio sample (210). Each location (242) is determined in dependence upon the content of the respective audio sample (210) and each fingerprint object (232) characterizes one or more local features (222) at or near the respective particular location (242). A relative value is next determined for each pair of matched fingerprint objects. A histogram of the relative values is then generated. If a statistically significant peak is found, the two audio samples can be characterized as substantially matching.
1. A method of characterizing a relationship between a first and a second audio samples, comprising the steps of:
generating a first set of fingerprint objects for the first audio sample, each fingerprint object occurring at a respective location within the first audio sample, the respective location being determined in dependence upon the content of the first audio sample, and each fingerprint object characterising one or more features of the first audio sample at or near each respective location;
generating a second set of fingerprint objects for the second audio sample, each fingerprint object occurring at a respective location within the second audio sample, the respective location being determined in dependence upon the content of the second audio sample, and each fingerprint object characterising one or more features of the second audio sample at or near each respective location;
pairing fingerprint objects by matching a first fingerprint object from the first audio sample with a second fingerprint object from the second audio sample that is substantially similar to the first fingerprint object;
generating, based on the pairing step, a list of pairs of matched fingerprint objects;
determining a relative value for each pair of matched fingerprint objects;
generating a histogram of the relative values; and
searching for a statistically significant peak in the histogram, the peak characterizing the relationship between the first and second audio samples.
2. The method according to
3. The method according to
4. The method according to
selecting a neighbourhood around the peak, and
calculating an average of the relative values in the neighbourhood.
5. The method according to
6. The method according to
(i) a ratio between a first and a second frequency values, each frequency value being respectively determined from a first and a second local features near the respective location of each fingerprint object;
(ii) a product between a frequency value and a delta time value, the frequency value being determined from a first local feature, and the delta time value being determined between the first local feature and a second local feature near the respective location of each fingerprint object; and
(iii) a ratio between a first and a second delta time values, the first delta time value being determined from a first and a second local features, the second delta time value being determined from the first and a third local features, each local feature being near the respective location of each fingerprint object.
7. The method according to
8. The method according to
9. The method according to
10. The method according to
11. The method according to
12. The method according to
13. The method according to
14. The method according to
15. The method according to
determining a relative pitch for the first and second audio samples using the respective variant components, wherein each variant component is a frequency value determined from a local feature near the respective location of each fingerprint object;
determining a relative playback speed for the first and second audio samples using the respective variant components, wherein each variant component is a delta time value determined from a first and a second local features near the respective location of each fingerprint object; and
detecting if the relative pitch and a reciprocal of the relative playback speed are substantially different, in which case the relationship between the first and second audio samples is characterized as nonlinear.
16. The method according to
for each pair of matched fingerprint objects in the list, determining a compensated relative time offset value, t−R*t′, where t and t′ are locations in time with respect to the first and second fingerprint objects;
generating a second histogram of the compensated relative time offset values; and
searching for a statistically significant peak in the second histogram of the compensated relative time offset values, the peak further characterizing the relationship between the first and second audio samples.
17. A computer program product for performing a method according to any preceding claim.
18. A computer system for performing a method according to any one of
This application is cross-referenced to and claims priority from PCT International Application PCT/US03/12126 filed on Apr. 18, 2003, and U.S. Provisional Application No. 60/376,055 filed on Apr. 25, 2002, which are hereby incorporated by reference.
This invention relates generally to audio signal processing over a large database of audio files. More particularly, it relates to an inventive technique for rapidly and accurately determining whether two audio samples match, as well as being immune to various transformations including playback speed variation. The inventive technique further enables accurate estimation of the transformations.
The need for fast and accurate automatic recognition of music and other audio signals continues to grow. Previously available audio recognition technology often traded off speed against accuracy, or noise immunity. In some applications, calculating a regression is necessary to estimate the slope of a time-time scatter-plot in the presence of extreme noise, which introduced a number of difficulties and lowered performance in both speed and accuracy. Previously existing audio recognition techniques were therefore incapable of performing fast and accurate recognition in the presence of significant playback speed variation, for example, in recognizing a recording that is played at a speed faster than normal.
Adding to the complexity of the problem is an increasingly popular kind of speed variation, pitch-corrected tempo variation, used by DJ's at radio stations, clubs, and elsewhere. Currently, there is no robust and reliable technique that can perform fast and accurate audio recognition in spite of the playback speed variations and/or pitch-corrected tempo variations.
The present invention fulfills the need in the audio recognition art by providing a fast and invariant method for characterizing the relationship between two audio files. The inventive method is accurate even in the presence of extreme noise, overcoming the aforementioned drawbacks of existing technology.
According to an aspect of the invention, the relationship between two audio samples can be characterized by first matching certain fingerprint objects derived from the respective samples. A set of fingerprint objects is generated for each audio sample. Each fingerprint object occurs at a particular location within the respective audio sample. Each location is determined in dependence upon the content of the respective audio sample and each fingerprint object characterizes one or more local features of the respective audio sample at or near the respective particular location. In one embodiment, each fingerprint object is further characterized by a variant component and an invariant component. A relative value is next determined for each pair of matched fingerprint objects. A histogram of the relative values is then generated. If a statistically significant peak is found in the histogram, then the two audio samples can be characterized as, for example, substantially matching.
According to another aspect of the invention, the above-described technique can be further enhanced by providing an estimate of a global relative value with a location of the peak on an axis of the histogram. The global relative value, in turn, can be refined by selecting a neighborhood around the peak of interest and calculating an average of the relative values in the selected neighborhood.
In yet another embodiment, in which a relative playback speed value is determined from the peak of the histogram, a compensated relative time offset value is calculated for each pair of matched fingerprint objects. Another histogram is generated based on the compensated relative time offset values. If a statistically significant peak is found in the second histogram, then the relationship between the two audio samples can be further characterized by the peak, providing further enhancement to the accuracy of the invention.
FIGS. 4A-B show exemplary histograms with and without a statistically significant peak.
FIGS. 5A-B illustrate the motion of time-frequency points as the playback speed varies.
FIGS. 6A-B show corresponding times in a first audio sample (sample sound) and a second audio sample (database sound) of matching hash tokens. The slope equals one when the playback speed of the sample sound is the same as the database sound.
FIGS. 7A-D illustrate fast and efficient slope finding and histogramming techniques of the present invention.
The present invention enables fast, robust, invariant, and scalable indexing and searching over a large database of audio files and is particularly useful for audio pattern recognition applications. In some embodiments, the techniques disclosed herein improve and enhance the audio recognition system and methods disclosed in the above-referenced U.S. patent application.
A very fast and efficient comparison operation between two audio sample files is essential in building a commercially viable audio recognition system. According to an aspect of the invention, the relationship between two audio samples can be characterized by first matching certain fingerprint objects derived from a spectrogram, such as one shown in
After a spectrogram of each audio sample is generated, it is scanned for local features, e.g., local energy peaks, as shown in
In an exemplary embodiment, each fingerprint object is characterized by its location, a variant component, and an invariant component. Each local feature is a spectrogram peak and each frequency value is determined from a frequency coordinate of a corresponding spectrogram peak. The peaks are determined by searching in the vicinity of each time-frequency coordinate and selecting the points that have a greater magnitude value than its neighbors. More specifically, as shown in
The above-described technique can be further enhanced by providing an estimate of a global relative value R with a location of the peak on an axis of the histogram, as illustrated in step 361. In some embodiments, R can be refined by first selecting a neighborhood around the peak of interest. In
Other kinds of time-frequency analyses may be implemented for extracting fingerprint objects, e.g., the Wigner-Ville distribution or wavelets. Also, instead of spectrogram peaks, other features, e.g., cepstral coefficients, can be used. Further, super-resolution techniques could be used to obtain finer frequency and time estimates of the time-frequency coordinates provided by the spectrogram peaks. For example, parabolic interpolation on frequency bins could be used to increase the frequency resolution. Related exemplary teachings can be found in “PARSHL: An Analysis/Synthesis Program for Non-Harmonic Sounds Based on a Sinusoidal Representation”, Julius O. Smith III and Xavier Serra, Proceedings of the International Computer Music Conference (ICMC-87, Tokyo), Computer Music Association, 1987, and in “Modern Spectral Estimation: Theory and Application,” by Steven M. Kay (January 1988) Prentice Hall, both of which are incorporated herein by reference.
In a matching operation, two audio samples are compared via their respective fingerprint objects. As discussed before with reference to
In an exemplary embodiment, each fingerprint object contains an invariant component and a variant component. The invariant component refers to the ratios of frequency values corresponding to spectral peaks, as well as ratios of delta time (i.e., time difference) values between spectral peaks are invariant under time stretch. For example, referring to
To make matching efficient, we use the invariant part to index the fingerprints and use approximate or exact values to search. Searching using approximate matches allows for some extra robustness against distortions and rounding error, but incurs more cost if the search over the invariant components becomes a multidimensional range search. In the preferred embodiment, the invariant component of respective fingerprint objects is required to match exactly, thus yielding a system that is very fast, with a minor tradeoff against sensitivity of recognition in the presence of noise. It is important to note that this method works well even if only a minority of fingerprint objects in corresponding audio samples match correctly. In the histogram peak detection step, a peak may be statistically significant even if as few as 1-2% of the fingerprint objects are correctly matched and survive.
The variant component can also be used to narrow down the number of matching fingerprint objects, in addition to, or instead of the invariant component. For example, we could require that a variant component V from the first audio sample match a corresponding V′ from the second audio sample within +/−20%. In that case, we can form a representation of the numeric tokens such that the upper portion (e.g., most significant bits) contains the invariant components, and the lower portion (e.g., least significant bits) contains the variant components. Then, searching for an approximate match becomes a range search over the tokens composed using the lowest and highest values of the variant component. The use of an invariant component in matching is thus not strictly necessary if searching is done using a variant component. However, using an invariant component in the matching process is recommended since it helps to reduce the number of spurious matches, thus streamlining the histogramming process and reducing the amount of processing overhead.
On the other hand, the novel variant component itself may or may not be a part of the matching criteria between two fingerprint objects. The variant component represents a value that may be distorted by some simple parametric transformation going from an original recording to a sampled recording. For example, frequency variant components, such as f1, f2, f3, and time variant components such as (t2−t1), (t3−t1), or (t3−t2) may be chosen as variant components for playback speed variation. Suppose a second audio sample, say a matching rendition from a database, had a spectrogram with coordinates (t1′,f1′), (t2′,f2′), and (t3′,f3′), corresponding to the same points listed above for the first audio sample. Then the frequency component f1′ could have a scaled value f1′=Rf*f1, where Rf is a linear stretch parameter describing how much faster or slower the first sample recording was compared to the second. The variant component from each of the two matching audio samples can be used to calculate an estimate of the global stretch value, which describes a macroscopic parameter, by calculating the ratio between the two frequency values, Rf=f1′/f1. This gives the relative pitch ratio of the two matched time-frequency points; for example, Rf=2 means that the first audio sample has half the pitch (frequency) of the second. Another possibility is to use Rt=(t2′−t1′)/(t2−t1). In this case, the relative value R is the relative playback speed ratio, i.e., Rt=2 means that the first audio sample is playing back twice as fast as the second audio sample.
If Rf=1/Rt, i.e., f′/f=(t2−t1)/(t2′−t1′), then the two audio samples are related by a linear time stretch, due to the reciprocal time-frequency relationship for such audio samples. In this case, we can first use the histogramming method disclosed herein to form an estimate Rf of the relative frequency ratio using corresponding variant frequency components, and again to form an estimate of Rt of the relative playback speed, then perform a comparison to detect whether the playback relationship is linear or nonlinear.
In general, a relative value is calculated from matched fingerprint objects using corresponding variant components from the first and second audio samples. The relative value could be a simple ratio of frequencies or delta times, or some other function that results in an estimate of a global parameter used to describe the mapping between the first and second audio sample. But generally, any 2-input function F( ) may be used, e.g. R=F(v1,v1′), where v1 and v1′ are respective variant quantities. It is best if F( ) is a continuous function so that small errors in measuring v1 and v1′ result in small errors in the output R.
As described herein, a histogram is generated over the set of relative values calculated from the list of matching pairs of fingerprint objects. The histogram is then searched for a peak. The presence of a statistically significant peak in the histogram indicates that a possible match has occurred. This method particularly searches for a cluster in the histogram of relative values instead of differences of time offsets, such as (t1′−t1). According to a principle of the present invention, a histogram serves to form bins of count values, each bin corresponding to a particular value along the independent axis of the histogram. For the purpose of this invention, generating a histogram may be accomplished by simply sorting the list of relative values. Therefore, a fast and efficient way of detecting the peak of a histogram of a list of values is to sort the list in ascending order, then scan for the largest clump of items having the same or similar values.
As discussed herein, with the present invention, two audio samples can be correctly matched even if only as few as 2% of the fingerprint objects survive all the distortions and are correctly matched. This is possible by scoring the comparison between the two audio samples. Specifically, a neighborhood is chosen around the peak of the histogram and all the matching pairs falling into the neighborhood are counted, giving the score. Additionally, a weighted score may be computed, discounting the contribution of pairs that are farther from the center of the peak.
One way to estimate the cutoff criterion is to assume that the probability distribution of the score of a non-matching track falls off with an exponential tail. The model is applied to the actual measured distribution of scores of non-matching tracks. Next the cumulative probability distribution of the highest score over a database of N tracks (e.g., taken as the Nth power of the cumulative probability distribution of a single non-matching score) is calculated. Once the probability curve is known and a maximum level of false positives is chosen (e.g., 0.5%), then a numeric threshold can be chosen and used to determine whether the histogram peak has a statistically significant number of matching pairs.
Once a statistically significant histogram peak is found, a high-resolution “hyperfine” estimate of the global relative value (such as relative playback speed) may be computed. This is accomplished by choosing a neighborhood around the peak, e.g., including an interval about 3 or 5 bins wide centered on the peak histogram bin, and calculating an average of the relative values in the neighborhood. Using this technique, we can find relative playback speed accurate to within 0.05%. With offset derivation disclosed herein, the global time offset may be estimated with better than 1 millisecond accuracy, which is finer than the time resolution of the spectrogram frames discussed above.
As discussed in the above-referenced U.S. patent application, in the case that the samples actually matched, a diagonal line could be seen in a scatterplot where matching samples have the corresponding time coordinates (t′,t) of matching fingerprint objects plotted against each other, as shown in
The present invention provides an inventive method of estimating the relative playback speed (or, equivalently, the reciprocal of the relative pitch, in the case of a linear playback relationship) that solves the problem of finding a regression line in the time-time scatterplot even if the slope of the match does not equal to one, e.g.,
To find the offset, assume that the corresponding time points have the relation
In summary, the first histogramming stage provides a way to estimate the relative playback speed, as well determining whether a match exists. The second histogramming stage ensures that the candidate matching audio samples have a significant number of fingerprint objects that are also temporally aligned. The second histogramming stage also serves as a second independent screening criterion and helps to lower the probability of false positives, thus providing a stronger criterion for deciding whether two audio samples match. The second histogramming stage may be optionally performed only if there is a statistically significant peak in the first histogram, thus saving computational resource and effort. A further optimization may be optionally performed, e.g., to reduce computational clutter, instead of computing the second histogram over all the pairs of matched fingerprint objects in the list, the second histogram can be generated using only the matching pairs corresponding to the first histogram peak.
Synchronization of Multiple Recordings
The present invention may be implemented for cueing and time alignment of unsynchronized audio recordings. For example, suppose a DAT recorder and a cassette recorder were operated independently with different microphones at slightly different locations or environments. If it is later desired to combine the two recordings from respective recorders into one mix, the two tracks may be synchronized using the robust regression technique described herein to obtain the time offset. As such, even if the unsynchronized recorders operate at slightly different speeds, the relative speed can be determined with a high degree of accuracy, allowing one recording be compensated with reference to another. This is especially useful if it is found that one of the recordings has become corrupted and needs to be supplemented from another source. The time alignment and synchronization as described herein thus allow for transparent mixing.
Since the comparison method is extremely fast, it is possible to pre-process a large database of audio samples into respective lists of fingerprint objects. As one skilled in the art would appreciate, an unknown audio sample may therefore be pre-processed into its own respective list of fingerprint objects using currently available data processing techniques. The above described matching, histogramming, and peak detection techniques can then be carried out using the pre-processed fingerprint objects in the database to find a match.
Although the present invention and its advantages have been described in detail, it should be understood that the present invention is not limited to or defined by what is shown or discussed herein. In particular, drawings and description disclosed herein illustrate technologies related to the invention, show examples of the invention, and provide examples of using the invention and are not to be construed as limiting the present invention. Known methods, techniques, or systems may be discussed without giving details, so to avoid obscuring the principles of the invention. As it will be appreciated by one of ordinary skill in the art, the present invention can be implemented, modified, or otherwise altered without departing from the principles and spirit of the present invention. For example, methods, techniques, and steps described herein can be implemented or otherwise realized in a form of computer-executable instructions embodied in a computer readable medium. Alternatively, the present invention can be implemented in a computer system having a client and a server. The client sends information, e.g., fingerprint objects, necessary for the characterization of the relationship between the first and second audio samples to the server where the characterization is performed. Accordingly, the scope of the invention should be determined by the following claims and their legal equivalents.