EP1388145A1

EP1388145A1 - Device and method for analysing an audio signal in view of obtaining rhythm information

Info

Publication number: EP1388145A1
Application number: EP02745267A
Authority: EP
Inventors: Jürgen HERRE; Jan Rohden; Christian Uhle; Markus Cremer
Original assignee: Fraunhofer Gesellschaft zur Forderung der Angewandten Forschung eV
Current assignee: Fraunhofer Gesellschaft zur Forderung der Angewandten Forschung eV
Priority date: 2001-05-14
Filing date: 2002-04-25
Publication date: 2004-02-11
Anticipated expiration: 2022-04-25
Also published as: DE10123366C1; US20040068401A1; WO2002093557A1; JP2004528596A; HK1059959A1; EP1388145B1; JP3914878B2; ATE279769T1; DE50201311D1

Abstract

The invention relates to a device for analysing an audio signal in view of obtaining rhythm information of said audio signal. Said device comprises a filter bank for decomposing (102) the audio signal into at least two partial band signals. Each partial band signal is analysed (106a, 106b) in terms of its periodicity, in order to obtain raw rhythm information for each partial band signal. Raw rhythm information is subjected to a quality evaluation (110a, 110b), in order to obtain a significance measure for each partial band signal. The rhythm information of the audio signal is determined, taking into account the significance measure of the partial band signal and the raw rhythm information (114). This enables a more robust analysis of the audio signal, as partial band signals containing clear rhythm information are preferred to partial band signals containing less clear rhythm information, for the determination of rhythm information.

Description

Apparatus and method for analyzing an audio signal for rhythm information

description

The present invention relates to signal processing concepts and, in particular, to the analysis of audio signals with regard to rhythm information.

In recent years, the availability of multimedia data such as B. audio or video data, increased significantly. This is due to a number of technical factors, which are particularly due to the wide availability of the Internet, powerful computer hardware and software, as well as powerful methods for data compression, i. H. Support source coding, audio and video processes.

The huge amounts of audiovisual data that are available worldwide, for example on the Internet, call for concepts that enable this data to be assessed, cataloged, etc. based on content criteria. There is a desire to be able to search for and locate multimedia data by specifying meaningful criteria.

This requires so-called “content-based” techniques, which extract so-called features from the audiovisual data, which are also referred to as “features” in specialist circles, which represent important characteristic properties of the signal. Based on such features or combinations of these features, similarity relationships or similarities between audio or video signals can be derived. This is done by comparison or in- Relation-setting of the extracted feature values from the different signals, which are also simply referred to as "pieces".

Of particular interest is the determination or extraction of features that are not only signal-theoretical, but also have the most direct semantic meaning possible. H. represent properties directly perceived by the listener.

This enables the user to formulate search queries in a simple and intuitive manner in order to find pieces from the entire existing database of an audio signal database. Likewise, semantically relevant features make it possible to model similarity relationships between pieces that come close to human perception. The use of features that have semantic meaning also makes it possible, for example, to automatically propose pieces that are of interest to a particular user if his preferences are known.

In the field of music analysis, the tempo is an important musical parameter that has semantic meaning. The tempo is usually measured in "beats per minute" (BPM). The automatic extraction of the tempo as well as the beats of the "beat" or generally speaking the automatic extraction of rhythm information is an example of obtaining a semantically important feature of a piece of music.

Furthermore, there is a wish that the feature extraction, ie the extraction of rhythm information from an audio signal, can take place robustly and computationally efficiently. Robustness means that it must not matter whether the piece has been source-coded and decoded again, whether the piece has been played over a loudspeaker and Microphone has been received, whether it is played loudly or quietly, or whether it is played by one or a plurality of instruments.

For the determination of the center of gravity and thus also the tempo, d. H. For the determination of rhythm information, the term "beat tracking" has also become established in the specialist circles. It is already known from the prior art to use beat tracking on the basis of a note-like or transcribed signal representation, for example in the midi The aim, however, is not to require such a meta representation, but to carry out an analysis directly with, for example, a PCM-coded or, generally speaking, digitally available audio signal.

The professional publication "Tempo and Beat Analysis of Acoustic Musical Signals" by Eric D. Scheirer, J. Acoust. Soc. Am. 103: 1, (Jan 1998), pages 588-601, discloses a method for the automatic extraction of a rhythmic pulse Musical excerpts: The input signal is split into a series of subbands by means of a filter bank, for example into 6 subbands with crossover frequencies of 200 Hz> 400 Hz, 800 Hz, 1600 Hz and 3200 Hz. Low pass filtering is carried out for the first subband. For the last one Sub-band is high-pass filtered, band-pass filtering is described for the remaining sub-bands in between. Each sub-band is processed as follows. The sub-band signal is first rectified. In other words, the absolute value of the samples is determined. The resulting n values are then smoothed, for example with an averaging over a suitable window to egg n receive envelope signal. The envelope signal can be subsampled to reduce the computational complexity. The envelope signals are differentiated, ie Sudden changes in the signal amplitude are preferably passed on through the differentiation filter. The result is then limited to non-negative values. Each envelope signal is then placed in a bank of resonant filters, ie oscillators, each containing a filter for each tempo range, so that the filter that matches the musical tempo is most strongly stimulated. For each filter, the energy of the output signal is calculated as a measure of the correspondence between the tempo of the input signal and the tempo associated with the filter. The energies for each tempo are finally summed up over all subbands, the largest energy sum identifying the tempo supplied as the result, ie the rhythm information.

A major disadvantage of this method is the great computation and storage complexity, in particular for realizing the large number of parallel-oscillating “oscillators”, of which only one is ultimately selected. This makes efficient implementation, for example for real-time applications, almost impossible.

The professional publication "Pulse Tracking with a Pitch Tracker" by Eric D. Scheirer> Proc. 1997 Workshop on Applications of- Signal Processing to Audio and Acoustics, Mohonk, NY, Oct 1997, describes a comparison of the "oscillator concept" described above with an alternative concept based on the use of autocorrelation functions to extract periodicity from an audio signal, i.e. H. the rhythm information of a signal. An algorithm for modulating human pitch perception, i.e. H. the pitch, is used for "beat tracking".

The known algorithm is shown in Fig. 3 as a block diagram. The audio signal is input via an audio input 300 ner analysis filter bank 302 supplied. The analysis filter bank generates a number n of channels, ie individual subband signals, from the audio input. Each subband signal contains a certain range of frequencies of the audio signal. The filters of the analysis filter bank are selected so that they approximate the selection characteristics of the human inner ear. Such an analysis filter bank is also referred to as a gamma-tone filter bank.

The rhythm information of each subband signal is evaluated in the devices 304a to 304c. For each input signal, an envelope-like output signal is first calculated (corresponding to a so-called "inner hair cell" processing in the ear) and subsampled. An autocorrelation function (AKF) is calculated from this result in order to determine the periodicity of the signal as a function of the delay, ie the " Lag ".

An autocorrelation function, which represents aspects of the rhythm information of each subband signal, is then present at the output of the devices 304a to 304c for each subband signal.

The individual autocorrelation functions of the subband signals are then combined in a device 306 by summation in order to obtain a sum autocorrelation function (SAKF) which reproduces the rhythm information of the signal at the audio input 300. This information can be output at a tempo output 308. Large values in the total autocorrelation indicate that there is a high periodicity of the beginning of notes for a delay (lag) assigned to a peak of the SAKF. Therefore, for example, the greatest value of the sum auto-correlation function is sought within the musically sensible delays. Musically sensible delays include the tempo range between 60 bpm and 200 bpm. Means 306 may also be arranged to convert a delay time into tempo information. For example, a peak of a one second delay corresponds to a rate of 60 beats per minute. Smaller delays indicate higher speeds, while larger delays indicate lower speeds than 60 bpm.

This method has an advantage over the first-mentioned method in that no oscillators need to be implemented with a large amount of computation and memory. On the other hand, the concept is disadvantageous in that the quality of the results depends very much on the type of audio signal. If, for example, a dominant rhythm instrument can be heard from an audio signal, the concept described in FIG. 3 will work well. If, on the other hand, the voice is dominant, which will not provide particularly clear rhythm information, the rhythm determination will be ambiguous. In the audio signal there could also be a tape that only contains rhythm information, ie.z. B. a higher frequency band in which, for example, a hi-hat of a drum kit is positioned, or a low frequency band in which the bass drum of a drum kit is positioned on the frequency scale. Due to the combination of the individual information, however, the somewhat unambiguous information of these special subbands is overlaid or "watered down" by the ambiguous information of the other subbands.

Another problem with using autocorrelation functions to extract the periodicity of a subband signal is that the sum autocorrelation function, obtained by means 306 is ambiguous. The sum autocorrelation function at output 306 is ambiguous in that an autocorrelation function peak is generated even when a delay is multiplied. This is understandable from the fact that a sine component with a period of tO, if it is subjected to an auto-correlation function processing, generates maxi a at multiples of the delays, ie at 2t0, 3t0, etc., in addition to the desired maximum at tO.

The specialist publication "A Computationally Efficient Multi-pitch Analysis Model" by Tolonen and Karjalainen, IEEE Transactions on Speech and Audio Processing, Volume 8, No. 6, Nov. 2000, discloses a computationally efficient model for a periodicity analysis of complexes The calculation model divides the signal into two channels, namely a channel below 1000 Hz and a channel above 1000 Hz. From this an autocorrelation of the lower channel and an autocorrelation of the envelope curve of the upper channel are calculated. Finally, the two autocorrelation functions are summed up In order to eliminate the ambiguities of the sum autocorrelation function, the sum autocorrelation function is further processed in order to obtain a so-called Enhanced Summary-Autocorrelation Function (ESACF) (further developed sum autocorrelation function) tors spread versions of the autocorrelation function from the sum autocorrelation function with subsequent limitation to non-negative values.

The object of the present invention is to create a computing time-efficient and robust device and a computing time-efficient and robust method for analyzing an audio signal with regard to rhythm information. This object is achieved by a device for analyzing an audio signal according to claim 1 or by a method for analyzing an audio signal according to claim 11.

The present invention is based on the knowledge that in the individual frequency bands, i. H. the sub-bands, there are often different favorable conditions for finding rhythmic periodicities. While in pop music, for example, the signal is often dominated in the middle range, for example around 1 kHz, by vocals that do not correspond to the beat, percussion sounds are often present in the higher frequency ranges. B. the hi-hat of the drums, which allow a very good extraction of rhythmic regularities. In other words, depending on the audio signal, different frequency bands contain a different amount of rhythmic information or have a different quality or significance for the rhythm information of the audio signal.

According to the invention, the audio signal is therefore first broken down into subband signals. Any subband signal. is examined for its periodicity to obtain raw rhythm information for each subband signal. An evaluation of the quality of the periodicity of each subband signal is then carried out in accordance with the present invention in order to obtain a measure of significance for each subband signal. A high level of significance indicates that there is clear rhythm information in this subband signal, while a low level of significance indicates that there is less clear rhythm information in this subband signal. According to a preferred exemplary embodiment of the present invention, when examining a subband signal with regard to its periodicities, a modified envelope curve of the subband signal is first calculated and then an autocorrelation function of the envelope curve is calculated. The autocorrelation function of the envelope represents the raw rhythm information. Clear rhythm information is available if the autocorrelation function has clear maxima, while less clear rhythm information is available if the autocorrelation function of the envelope of the subband signal has fewer pronounced signal peaks or no signal peaks at all. An autocorrelation function that has significant signal peaks is therefore given a high level of significance, while an autocorrelation function that has a relatively flat profile is obtained a low level of significance.

According to the invention, the individual raw rhythm information of the individual subband signals are not simply combined “blindly”, but are used for each subband signal, taking into account the significance measure, in order to obtain the rhythm information of the audio signal. If a subband signal has a high significance measure, it becomes so when determining the rhythmin - - Formations preferred, while a ... sub-band signal that has a low degree of significance, ie that has a low quality with regard to the rhythm information, is hardly taken into account in the determination of the rhythm information of the audio signal or, in extreme cases, is not taken into account at all.

This can be implemented in a computationally efficient manner by means of a weighting factor which depends on the significance measure. While a sub-band signal that is of good quality for the rhythm information, that is, that has a high degree of significance, could be given a weighting factor of 1, another sub-band signal that has a smaller degree of significance is get a weighting factor less than 1. In extreme cases, a subband signal that has a completely flat autocorrelation function will have a weighting factor of 0. The weighted autocorrelation functions, ie the weighted raw rhythm information, are then simply added up. In extreme cases, if only one subband signal of all subband signals provides good rhythm information while the other subband signals have autocorrelation functions with a flat profile, this weighting can result in all subband signals except the one subband signal receiving a weighting factor of 0, i.e. not at all when determining the rhythm information are taken into account, so that the rhythm information of the audio signal is determined only from a single subband signal.

The concept according to the invention is advantageous in that it enables the rhythm information to be determined robustly, since subband signals with no clear or even deviating rhythm information, ie if the vocals have a different rhythm than the actual beat of the piece, do not ^contain the rhythm information of the audio signal. water down "or" falsify ". They also become very intoxicating. Subband signals,. welςhe__ .. provide a system autocorrelation function with a completely flat profile, do not deteriorate the signal / noise ratio when determining the rhythm information. Exactly this would occur, however, if, as in the prior art, all the autocorrelation functions of the subband signals were simply added up with the same weight.

Another advantage of the method according to the invention is that a significance measure can be determined with a small additional computing effort, and that the evaluation of the raw rhythm information with the significance measure and the subsequent summation can be carried out efficiently without a large amount of memory and computing time, which the concept according to the invention also recommends in particular for real-time applications.

Preferred embodiments of the present invention are explained in detail below with reference to the accompanying drawings. Show it:

1 shows a block diagram of a device for analyzing an audio signal with a quality evaluation of the raw rhythm information;

2 shows a block diagram of a device for analyzing an audio signal using weighting factors on the basis of the significance measures;

3 shows a block diagram of a known device for analyzing an audio signal with regard to rhythm information;

FIG. 4 shows a block diagram of a device for analyzing an audio signal with regard to rhythm information using an autocorrelation function with a subband-wise postprocessing of the rhythm raw information; FIG. and

FIG. 5 shows a detailed block diagram of the device for post-processing from FIG. 4.

1 shows a block diagram of a device for analyzing an audio signal with regard to rhythm information. The audio signal is transmitted via an input 100 to a device 102 for splitting the audio signal into at least two Subband signals 104a and 104b supplied. Each subband signal 104a, 104b is fed into means 106a and 106b for examining it for periodicities in the subband signal to obtain raw rhythm information 108a and 108b for each subband signal. The raw rhythm information is then fed to a device 110a or 110b for evaluating a quality of the periodicity of each of the at least two subband signals in order to obtain a significance measure 112a, 112b for each of the at least two subband signals. Both the raw rhythm information 108a, 108b and the significance measures 112a, 112b are fed to a device 114 for determining the rhythm information of the audio signal. When determining the rhythm information of the audio signal, the device 114 takes into account the significance measures 112a, 112b for the subband signals and the raw rhythm information 108a, 108b of at least one subband signal.

If the device 110a for quality assessment has determined, for example, that there is no particular periodicity in the subband signal 104a, the significance measure 112a will be very small or equal to 0. In this case, the device -114- for determining the rhythm information determines that the significance measure 112a is equal to zero, so that the raw rhythm information 108a of the subband signal 104a is no longer taken into account when determining the rhythm information of the audio signal Need to become. The rhythm information of the audio signal is then determined solely and exclusively on the basis of the raw rhythm information 108b of the subband signal 104b.

In the following, FIG. 2 is discussed with regard to a special embodiment of the device from FIG. 1. As a device 102 for decomposing the audio signal, a conventional before analysis filter bank are used, which delivers a user-selectable number of subband signals on the output side. Each subband signal is then subjected to the processing of the devices 106a, 106b and 106c, whereupon the devices 110a to 110c then determine significance measures of each raw rhythm information. In the preferred embodiment shown in FIG. 2, the device 114 comprises a device 114a for calculating weighting factors for each subband signal on the basis of the significance measure for this subband signal and optionally also for the other subband signals. A weighting of the raw rhythm information 108a to 108c then takes place in the device 114b with the weighting factor for this subband signal, whereupon, also in the device 114b, the weighted raw rhythm information is combined, e.g. B. summed up, in order to obtain the rhythm information of the audio signal at the tempo output 116.

^"The inventive concept ^turns' thus follows. After evaluating the rhythmic information of the individual bands which .stattfinden example, envelope shaping, smoothing, differentiating, limiting to positive values and making -the autocorrelation function., Can (Einrich- .. obligations 106a to 106c), an evaluation of the value or the quality of these intermediate results takes place in the devices 110a to 110c.This is achieved with the aid of an evaluation function which evaluates the reliability of the individual results with a significance measure. The significance measures of all subband signals become a weighting factor for each band for the extraction of the rhythm information The overall result of the rhythm extraction is then achieved in the device 114b by combining the band-wise individual results taking into account their respective weighting factors. As a result, an algorithm for rhythm analysis implemented in this way shows a good ability to reliably find rhythmic information in a signal even under unfavorable conditions. The concept according to the invention is therefore characterized by a high level of robustness.

In a preferred embodiment, the raw rhythm information 108a, 108b, 108c, which represent the periodicity of the respective subband signal, is determined by means of an autocorrelation function. In this case, it is preferred to determine the significance measure by dividing a maximum of the autocorrelation function by an average of the autocorrelation function and then subtracting the value 1. It should be noted that every autocorrelation function always delivers a local maximum at a delay of 0, which represents the energy of the signal. This maximum should be disregarded so that the quality determination is not falsified.

Furthermore, the autocorrelation function should only be considered in a special tempo range, ie from a maximum deceleration, .. which corresponds to the smallest, inter, e.ss., n, _> tempo, to a minimum deceleration, which corresponds to the highest of interest Pace corresponds. A typical tempo range is between 60 bpm and 200 bpm.

Alternatively, the ratio between the arithmetic mean value of the autocorrelation function in the tempo area of interest and the geometric mean value of the autocorrelation function in the tempo area of interest can be determined as a significance measure. It is known that if all values of the autocorrelation function are the same, ie if the autocorrelation function has a flat course, the geometric one The mean value of the autocorrelation function and the arithmetic mean value of the autocorrelation function are the same. In this case, the significance measure would have a value of 1, which means that the raw rhythm information is not significant.

In the case of a system auto-correlation function with strong peaks, the ratio of the arithmetic mean to the geometric mean would be greater than 1, which means that the auto-correlation function has good rhythm information. However, the smaller the ratio between the arithmetic mean and the geometric mean, the flatter the autocorrelation function and the fewer periodicities it contains, which in turn means that the rhythm information of this subband signal is less significant, i. H. have a lower quality, which will result in a low or a weighting factor of 0.

There are various options with regard to the weighting factors. A relative weighting is preferred, such that all weighting factors of all subband signals add up to 1, i. H. that the weighting factor of a band,. «- is determined-.-,. ,, as ,,, the significance value.« ... this, it., „. band divided by the sum of all significance values. In this case, a relative weighting is carried out before the summation of the weighted raw rhythm information in order to obtain the rhythm information of the audio signal.

As has already been explained, it is preferred to carry out the evaluation of the rhythm information using an autocorrelation function. This case is shown in Fig. 4. The audio signal is fed via the audio signal input 100 into the device 102 for splitting the audio signal into subband signals 104a and 104b. Each subband signal is then examined in the device 106a or 106b, as has been carried out, using an autocorrelation function in order to determine the periodicity of the subband signal. The raw rhythm information 108a, 108b is then available at the output of the device 106a or 106b. These are fed into a device 118a or 118b in order to postprocess the raw rhythm information output by the device 116a by means of the autocorrelation function. This ensures, among other things, that the ambiguities of the autocorrelation function, that is to say that signal peaks also occur in the case of integer multiples of the delays, are eliminated on a sub-band basis in order to obtain reworked raw rhythm information 120a or 120b.

This has the advantage that the ambiguities of the autocorrelation functions, i. H. of the raw rhythm information 108a, 108b are already eliminated in sub-bands, and not only, as in the prior art, after the summation of the individual autocorrelation functions. In addition, the elimination of the ambiguities in the autocorrelation functions on a band-by-band basis by means 118a, 118b enables the raw rhythm information of the subband signals to be handled independently of one another. -become -you can ... you can, for example, be subjected to a quality assessment using the device 110a for the raw rhythm information 108a or using the device 110b for the raw rhythm information 108b.

As shown by the dashed lines in FIG. 4, the quality assessment can, however, also take place on the basis of the post-processed raw rhythm information, this latter possibility being preferred since the quality assessment on the basis of the post-processed rhythm Raw information ensures that the quality of information that is no longer ambiguous is assessed.

The determination of the rhythm information by means 114 then takes place on the basis of postprocessed rhythm information of a channel and preferably also on the basis of the significance measure for this channel.

If a quality assessment is carried out on the basis of the raw rhythm information, i.e. the signal in front of the device 118a, this is advantageous in that if it is determined that the significance measure is equal to 0, i. H. that the autocorrelation function has a flat course, the postprocessing by means of the device 118a can be dispensed with entirely in order to save computing time resources.

In the following, FIG. 5 is discussed in order to show a more detailed structure of a device 118a or 118b for postprocessing the rhythm raw information. First, the subband signal, for example 104a, is fed into the device 106a for examining the periodicity of the subband signal ,, _^ .mitt.els .. ,, “- - _{= =} ₅ .Autokorrelationsfunkti.oι» «, .., _^ , ,around. Get raw rhythm information 108a. In order to eliminate the ambiguities on a sub-band basis, just as in the prior art, a spread auto-correlation function can be calculated by means of a device 121, the device 121 being arranged to calculate the spread auto-correlation function so that it is spread by an integral multiple of a delay , Means 122 is arranged in this case to subtract the spread autocorrelation function from the original autocorrelation function, ie the raw rhythm information 108a. In particular, it is preferred to first double the spread Calculate autocorrelation function in the device 121 and then subtract it from the raw rhythm information 108a. Then, in the next step, an autocorrelation function spread by a factor of 3 is calculated in the device 121 and subtracted from the result of the previous subtraction, so that all ambiguities are gradually eliminated from the raw rhythm information.

Alternatively or additionally, the device 121 can be arranged to calculate an autocorrelation function compressed by an integer factor, which is then added by the device 122 to the rhythm raw information in order to also include components for delays t0 / 2, tO / 3, etc . to create.

In addition, the spread or compressed versions of the raw rhythm information 108a can be weighted before adding or subtracting, in order to achieve flexibility in the sense of a high level of robustness.

To examine a sub-band signal on the basis of an autocorrelation function by the method, the periodicity "-.kan -So _a more Improvement .. he elt z. ₁ _J -W. Are “,. ,, if. the. Properties of the autocorrelation function are included and the post-processing is carried out using the device 118a or 118b. A periodic sequence of note starts with a distance tO not only generates an AKF peak with a delay tO but also at 2t0, 3t0, etc. This will lead to ambiguity in the tempo detection, i.e. the search for significant maxima in the autocorrelation function. The ambiguities can be cleared if versions of the AKF that are spread by integer factors are subtracted from the initial value on a sub-band basis (weighted). Furthermore, there is the problem with the autocorrelation function that it does not provide any information at tO / 2, t0 / 3 ... etc., that is to say at twice, three times, etc. of the “basic tempo”, which can lead to incorrect results especially if Two instruments that are in different subbands together define the rhythm of the signal, which is taken into account by calculating versions of the auto-correlation function that are compressed by integer factors, and then adding or unweighting these to the raw rhythm information.

The AKF postprocessing thus takes place sub-band, whereby an autocorrelation function is calculated for at least one sub-band signal and this is combined with stretched or spread versions of this function.

Claims

claims

1. Device for analyzing an audio signal with regard to rhythm information of the audio signal, with the following features:

means (102) for splitting the audio signal into at least two subband signals (104a, 104b);

means for examining (106a, 106b) a subband signal for periodicity in the subband signal to obtain raw rhythm information (108a, 108b) for the subband signal;

means for evaluating (110a, 110b) a quality of the periodicity of the raw rhythm information (108a) of the subband signal (104a) in order to obtain a measure of significance (112a) for the subband signal; and

a device (114) for determining the rhythm information _r . of the _r audio signal. under consideration of the _{_^.} Significance measure (112a) of the subband signal and the rhythm raw information (108a, 108b) of at least one subband signal.

2. The apparatus of claim 1, wherein the means for examining (106a, 106b) is designed to calculate an autocorrelation function for each of the at least two subband signals.

3. Device according to claim 1 or 2, wherein the device for examining (106a, 106b) has the following features: means for forming an envelope of a subband signal;

means for smoothing the envelope of the subband signal to obtain a smoothed envelope;

means for differentiating the smoothed envelope to obtain a differentiated envelope;

means for limiting the differentiated envelope to positive values to obtain a limited envelope; and

means for forming an autocorrelation function of the limited envelope to obtain the raw rhythm information (108a, 108b).

4. The device _' - ^' according to claim 2 or 3, in which the device for evaluating (110a, 110b) the quality is designed to use a ratio of a maximum of the autocorrelation function to a mean value of the autocorrelation function as a significance measure.

5. Apparatus according to claim 2 or 3, in which the device for evaluating (110a, 110b) the quality is designed to use as a significance measure a ratio of an arithmetic mean of the raw rhythm information to a geometric mean of the raw rhythm information.

6. The device according to claim 4 or 5,

in which the device for evaluating (110a, 110b) the quality is designed to only include the autocorrelation function within a tempo range that extends from a minimum delay to get a maximum pace to a maximum delay to get a minimum pace.

7. Device according to one of the preceding claims, in which the device for determining (114) has the following features:

means (114a) for deriving a weighting factor for a subband using the measure of significance for the subband;

means (114b) for weighting the sub-band's rhythm raw information using the sub-band weighting factor to obtain weighted sub-band rhythm raw information and summarizing the sub-band's weighted rhythm raw information with raw weighted or unweighted rhythm information other sub-band to get the rhythm information of the audio signal.

8. The device according to claim _. 7 ,. in which the means (114a) for deriving a weighting factor is arranged to derive a relative weighting factor for each subband signal, a sum of the weighting factors giving 1 for all subband signals.

9. The apparatus of claim 8, wherein the means (114a) for deriving a weighting factor is arranged to derive a weighting factor as a ratio of the significance measure of a subband signal to the sum of the significance measures of all subband signals.

10. The apparatus of claim 9, wherein the means (106a, 106b) for examining a subband signal is arranged to examine a subband signal whose length is greater than 10 seconds.

11. A method for analyzing an audio signal for rhythm information of the audio signal, comprising the following steps:

Decomposing the audio signal into at least two subband signals (104a, 104b);

Examining (106a, 106b) a subband signal for periodicity in the subband signal to obtain raw rhythm information (108a, 108b) for the subband signal;

Evaluating (110a, 110b) a quality of the periodicity of the raw rhythm information (108a) of the subband signal (104a) in order to obtain a measure of significance (112a) for the subband signal; and

Determining the rhythm information of the audio signal taking into account the significance measure (112a) of the subband signal and the raw rhythm information (108a, 108b) of at least one subband signal.