Estimating signal power in compressed audio

Abstract

Estimating a signal power in a compressed audio signal [A] is provided, the audio signal comprising blocks of quantized samples, a given block being provided with a set of scale factors. The estimating is performed by extracting the set of scale factors from the compressed audio signal, and estimating the signal power in the given block based on a combination of the scale factors. Advantageously, the extracting step and estimating step are performed on only a sub-set of the set of scale factors. The signal power estimation may be used in a silence detector (11) for use in a receiver (1).

Description

[0001] The invention relates to estimating signal power in a compressed audio signal. The invention further relates to silence detection and to a receiver using such a silence detection.

[0002] WO 96/3271 A1 discloses a system for compression and decompression of audio signals for digital transmission, wherein ancillary data may be multiplexed and encoded with audio data and transmitted in such a way that it may be decoded. This document discloses on page 159 the calculation of a minimum scale factor value to look for in another channel to see if audio is present.

[0003] It is an object of the invention to provide an advantageous signal power estimation in compressed audio signals. To this end, the invention provides a method and a device for estimating a signal power, a silence detector and a receiver as defined in the independent claims. Advantageous embodiments are defined in the dependent claims.

[0004] According to a first aspect of the invention, a signal power is estimated in a compressed audio signal comprising blocks of quantized samples, wherein a given block is provided with a set of scale factors. The set of scale factors is extracted from the compressed audio signal, and the signal power is estimated in the given block based on a combination of the scale factors. The given block may be one or more audio frames or part of an audio frame. Scale factors can easily be extracted from the compressed audio signal. The invention is based on the insight that a scale factor represents the maximum possible value of the samples it relates to. A combination of the scale factors, e.g. a sum of the squared scale factors, therefore gives a rough estimation of the signal power, only requiring limited computational load. The rough estimation is quite sufficient for some applications such as e.g. silence detection in commercial detectors.

[0005] In a preferred embodiment, only a sub-set of the scale factors is used. By using only a sub-set of the total set of scale factors, the computational load is further reduced. This may result in a lower accuracy, but this is acceptable for some applications like silence detection in commercial detectors etc.

[0006] Forming a sub-set of scale factors may be performed by omitting scale factors in time direction and/or in frequency direction. For example, the sub-set may only include a sub-set of a plurality of narrow band sub-signals available in the compressed audio signal, the sub-set preferably including the scale factors of a number of lower frequency sub-signals.

[0007] In the case the compressed audio signal is a stereo or multi-channel signal, only a subset of the available channels may be used.

[0008] These and other aspects of the invention will be apparent from and elucidated with reference to the accompanying drawings.

[0009] In the drawings:

[0010] FIG. 1 shows a receiver according to an embodiment of the invention;

[0011] FIG. 2 shows an exemplary audio frame including 32 sub-bands, each sub-band being sub-divided in 3 blocks, each block being including 12 quantized samples and being provided with a scale factor;

[0012] FIG. 3 shows the exemplary audio frame of FIG. 2 wherein for each sub-band a maximum scale factor is selected, a possible selection is highlighted in gray;

[0013] FIG. 4 shows an exemplary diagram wherein circles represent local signal powers of detected silences and wherein crosses represent an average of these local signal powers; and

[0014] FIG. 5 shows an exemplary likelihood function related to FIG. 4.

[0015] The drawings only show those elements that are helpful to understand the embodiments of the invention.

[0016] FIG. 1 shows a receiver 1 according to an embodiment of the invention for receiving a compressed audio signal [A]. The receiver 1 comprises an input 10 for obtaining the compressed audio signal [A]. The input 10 may be an antenna, a network connection, a reading device, etc. The receiver 1 further comprises a silence detector 11 for detecting silences in the compressed audio signal, and an influencing block 12 for influencing the audio signal depending on the detection of the silences. The block 12 may e.g. be a decoder for decoding the compressed audio signal, wherein the decoding depends on the detected silences. The block 12 may also be a skipping block for skipping parts of the compressed audio depending on the detected silences. The silence detector 11 may be enhanced to form a commercial detector. Detected commercials may be skipped during decoding. The influenced audio signal A, decoded or still compressed, can be outputted to output 13. The output 13 may be a network connection, a reproduction device or a recording device. The compressed audio signal [A] may be included in a program stream, which program stream further includes a video signal. In that case, the program signal may be influenced in block 12 at least partly depending on the silences detected in the compressed audio signal. An advantageous application is a storage device, which stores only non-commercial content.

[0017] Embodiments of the invention are described in the context of silence detection for use in e.g. commercial detection. It is noted that EP 1 006 685 A2 discloses a method and apparatus for processing a television signal, and for detecting the presence of commercials in the television signal. A commercial candidate section detector detects a commercial candidate section on the basis of a quiet section and a scene change point. A commercial characteristic quantity detector judges whether the commercial candidate section has various characteristics of commercials, and adds a predetermined value to a commercial characteristic value on the basis of the judgment result. The commercial characteristics quantity detector compares the final commercial characteristic value with a predetermined threshold value, and judges on the basis of the comparison result whether the commercial candidate section is a commercial section. A quiet section detector compares the level of a digitized audio signal with a threshold value to detect quiet sections, and outputs the comparison result to a scene change detector. Further reference is made to EP 1 087 557 A2.

[0018] A commercial detector according to an embodiment of the invention automatically detects commercial blocks in audiovisual streams. This allows to skip commercials during any kind of processing such as key-frame extraction, editing or playback. For several audio features, local statistics are measured on a sliding window and compared to a statistical model of commercials. By this comparison a normalized likelihood function is derived which tells how the audio signal is locally similar to commercials. The likelihood function can be properly triggered for the commercial detection. The statistical window is chosen in order to be both detailed in the local analysis and robust against local irregularities and fluctuations, which do not affect the detection. The algorithm is adaptive to some conditions, which can vary along a single stream or between one stream and another. The algorithm is video independent. Video analysis can nevertheless be included to enhance or extend the classification. The algorithm can be applied to several kinds of storage systems.

[0019] Many audio coders (e.g. MPEG-1 Layer 1/2/3, MPEG-2 Layer 1/2/3, MPEG-2 AAC, MPEG-4 AAC, AC-3) are frequency domain coders. They split the source spectrum into a number of narrow band sub-signals and quantize each frequency component or sample separately. Frequency components or samples are quantized according to a scale factor and according to a bit allocation. These scale factors can be regarded as indicators of the maximum value of frequency components or samples.

[0020] In AC-3 the frequency components are represented by: mantissa.2(−exponent). Here the exponent acts as a scale factor for each mantissa, equal to 2(−exponent).

[0021] In MPEG-1 layer 2 the narrow band sub-signals are divided in groups of 12 quantized samples, where each group has a corresponding scale factor. This scale factor corresponds to the maximum value of the samples it relates to.

[0022] The detection algorithm preferably uses a subset of the scale factors. In all or a subset of the narrow band sub-signals an upper bound of the signal power is calculated by squaring the scale factors.

[0023] An embodiment using MPEG audio compression is described in more detail now. In MPEG-1 layer 2 the audio signal is divided in time intervals of 24 msec, 26.1 msec or 36 msec for a sampling rate of 48 kHz, 44.1 kHz or 32 kHz respectively. In each of these intervals the signal is encoded in a frame. Referring to FIG. 1, each frame interval is divided in three parts and the signal is decomposed in 32 subband components. For each subband component and each third of a frame (one rectangle in FIG. 1) 12 samples are quantized according to a scale factor and according to a number of bits properly chosen. The scale factor gives an upper bound estimate of the absolute value of the 12 samples. This estimate may not be very accurate, but this is not required for the commercial detection. The scale factors can be extracted from each audio frame with negligible computational load, as they are directly available in the frame as pseudo logarithmic indexes. Only some limited frame header decoding is required. No decompression is necessary.

[0024] In stereo mode each channel has its own 96 scale factors per frame. The detection algorithm selects only the maximum scale factor in each subband of the left or right channel (see FIG. 2): 32 values are buffered and converted to the linear (not logarithmic) format. For instance, for a 48 kHz audio sampling rate, only subbands 0 . . . 26 are used according to the standard: this gives 27 samples every 24 msec that is 1125 samples/sec, a very modest input data rate for the commercial detector. The squares of the buffered scale factors are calculated to obtain an upper bound on the subband signal powers. These are then used as follows:

[0025] (1) their sum gives an upper bound on the total short time power;

[0026] (2) they can be used to calculate a short time bandwidth estimation;

[0027] The following table gives a few of the pseudo logarithmic indexes for the scale factors in MPEG-1 layer 2 (see Table B.1 in ISO/IEC 11172-3: 1993):

index scalefactor
0     2.0000
1     1.5874
2     1.2599
3     1.0000
4     0.7937
5     0.6299

[0028] An estimate of the short time power for an audio frame j is given below:

Frame_powerj≈Σscalefactorj,i2≈Σ100.6-0.2·index

[0029] It is alternatively possible to use a look-up-table to find the scale factor. The summation is to be performed over the number of sub-bands at a given time instance. When a sub-set of sub-bands is used, the summation has to be performed over the total number of sub-bands or the number of used sub-bands depending on the application.

[0030] Silence detection is based on nested thresholds on:

[0031] 1) local signal power level, by using e.g. Frame_power as indicated above

[0032] 2) silence duration;

[0033] and at least one of the following parameters:

[0034] 3) local power linear deviation during silence; and

[0035] 4) local power fall rate before silence start; and

[0036] 5) local power rise rate at silence end;

[0037] Because signal power characteristics are very much dependent on the environment in which the silence detector operates the silence detector is preferably adaptive. Therefore, in order to be adaptive, local power level related parameters (i.e. 1), 3) and/or 4) ) are compared with their average values in time. A typical threshold for the local signal power is 0.01, i.e. the local signal power should be less than one percent of the time average of the signal power. The time average is calculated by using an adaptation window with length w frames. A practical solution is the following: 1$\mathrm{average\_frame}{\mathrm{\_power}}_{-1}=0$$\mathrm{average\_frame}{\mathrm{\_power}}_j=\{\begin{array}{c}\mathrm{average\_frame}{\mathrm{\_power}}_{j-1}\frac{j-1}{j}+\frac{{\mathrm{frame\_power}}_j}{j}\mathrm{if}\left(j<w\right)\\ \mathrm{average\_frame}{\mathrm{\_power}}_{j-1}\frac{w-1}{w}+\frac{{\mathrm{frame\_power}}_j}{w}\mathrm{else}\end{array}$

[0038] wherein j is the frame index.

[0039] The silence duration is the duration that the local signal power level is below a given fixed or adaptive threshold power level. The linear deviation is a summation of (frame power minus mean frame power) over at least part of the silence duration. The linear deviation and fall/rise rate are used to filter part of the silences, which may be perceptual but are not relevant for the commercial detection. The local signal power level is preferably determined by using the scale factors as described above, for example per audio frame or part of an audio frame.

[0040] A practical range for silence duration breaks between commercials in a commercial block is {fraction (3/25)} sec to {fraction (20/25)} sec.

[0041] The values of silence beginning time, silence duration and silence local power level are buffered for the statistical calculations mentioned below. The commercials are characterized with a local statistical model of the following features:

[0042] 1) time distance between two consecutive detected silences;

[0043] 2) local signal power level of the detected silences (absolute and/or relative);

[0044] 3) silence duration; and

[0045] 4) local bandwidth of the audio signal;

[0046] The local bandwidth of an audio frame j may be calculated from the scale factors in the following manner: 2${\mathrm{local\_bandwidth}}_j=\mathrm{const}.\left[\frac{\sum _{i=0}^{\mathrm{used\_subbands}-1}{i}^2·{\mathrm{scalefactor}}_{j,i}^2}{{\mathrm{frame\_power}}_j}-{\left(\frac{\sum _{i=0}^{\mathrm{used\_subbands}-1}i·{\mathrm{scalefactor}}_{j,i}^2}{{\mathrm{frame\_power}}_j}\right)}^2\right]$

[0047] For each feature a 0.5-normalised likelihood function is obtained, with values between 0 and 1. It represents how much the local statistics of this feature are similar to those of commercials. The different likelihood functions are then combined with different weights to obtain a global likelihood function, still 0.5 normalized, which exploits all the information at a time. The global likelihood function is calculated in each point of the time axis, which was buffered as a silence beginning instant. The value 0.5 means basically “total uncertainty” or “0.5 probability of being inside a commercial block”. The likelihood function can be used in different ways. It can be properly triggered to detect commercial boundaries. It can be used (as a normalized soft classification between commercials and non-commercials) by algorithms that make further analysis and classifications, exploiting optionally also video features. Video features of different levels (like mono-luminance, mono-chrominance frame detection, scene change detection) can be statistically analyzed together with audio features applying the same likelihood method or other methods. The triggered commercial detection with refilling has been developed and tested, based on the previously described audio analysis. The 0.5 normalization likelihood function L(t) can be used to decide whether a detected silence belongs to a commercial block. This can be done by means of a function Q(L(t)), which is defined as follows:

Q (L(t))=1 if L(t)>0.5

Q (L(t))=0 if L(t)<=0.5,

[0048] where a value of 0 and 1 mean that the detected silence belong to a non-commercial block and commercial block respectively.

[0049] In a practical embodiment, a sequence of commercials is only detected if it lasts at least 60 sec. If only for a short interval inferior to 45 sec the likelihood function goes below 0.5, Q(t) is set to 1. This procedure has been called “internal refilling”. The internal refilling eliminates sporadic internal missing detections. An “external refilling” is applied at the beginning and end of the commercials. For instance if:

t i , t i+1 , . . . , t i+N , . . . is a sequence of instants in which detected silences start and

L (ti)=0.2

L (ti+1)=0.4

L (ti+2)=0.6

L (tj)>0.5 for each j=i+3, . . . , i+N

L (tj)<0.5 for j>i+N and if

t i+2 −t i+1 <45.0 sec

t i+N+1 −t i+N <45.0 sec

[0050] then

Q (L(ti))=0

Q (L(ti+1))=1

Q (L(ti+2))=1

[0051] . . .

Q (L(ti+N+1))=1

Q (L(tj))=0 for j>i+N+1

[0052] The external refilling is effective in avoiding the systematic miss of the first and last spots. This fact is related to windowing details. The external and internal refilling can be considered as a special non-linear filtering, upper driven. A general-purpose statistical model of commercial blocks may be used. It is possible to refine the statistical detail, using different commercial block models for the different times of the day and/or the different kind of programs (soaps, talk shows, football matches, etc.) and/or the different channels. Although this is not necessary to obtain satisfying performances, it may of course improve them. It is a matter of trade off between the complexity of the target system and its performance. Adaptability of the detection is preferred as the conditions change in time for a single channel. Moreover adaptability to channel switching is preferred. In particular the local minimum noise level may change in time for a single channel and can change a lot from one channel to another: this is critical for silence detection. Besides, adaptability in the statistical model of commercial blocks is not critical but useful. The system may be implemented as a fully self-training (adaptive) on the local minimum noise level. The only constraint is applying a reset of the algorithm every time the channel is switched. This is because the adaptability is fast in the initial period and slower in the following, for matters of trade off between adaptability and precision. If the algorithm is made fast adaptive at any time, the precision of the detection will decrease because inside the commercial blocks a relatively fast adaptation will decrease the precision. In a practical embodiment, the switch-adaptability is valid only in the first minutes (i.e. reset for any successive switch) while the along-a-single-channel adaptability always holds. Stability of adaptability is ensured by an asymmetric scheme. When the minimum noise level is decreasing adaptability is faster than when it is increasing. This means for instance that the local power energy threshold for the silence detection decreases relatively fast when a silence is detected with a local power energy lower than the ones detected before. There are two kinds of errors which can occur: either missing commercial detection or false commercial detection. Both are relatively low and confined to the beginning or ending part of commercial blocks. The algorithm is anyhow flexible: decision parameters can vary the trade off between the two error rates, depending on which is more relevant. For instance, if commercial block detection is a preprocessing for automatic key-frame extraction, then a low missing detection rate is more important. Low false detection is more relevant in the case of a simple playback. Referring to the features chosen (but others may be added), it is possible to evaluate separately local power energy and bandwidth without subband analysis. The value of the bandwidth is required with a low sampling rate on a two minutes (other values may be chosen) symmetric sliding window. Therefore it can be estimated for instance by the average of successive short interval FFT's with a low number of points. It is possible to implement different kinds of normalizations and combinations of one or several likelihood functions, either term by term or globally. A practical implementation is based on product combination term by term or globally with renormalization. The product is basically a Boolean AND extended from the Boolean set {0, 1} to the continuous interval [0, 1]. It ensures good selectivity. Roughly speaking, different conditions are softly imposed all at a time. They do not need to be all perfectly fulfilled but they need to be all mostly fulfilled. An addition combination instead would have been a sort of extension of the Boolean OR, which does not ensure sufficient selectivity. Further selectivity and robustness is ensured by the hard decision on the likelihood with a duration threshold. Likelihood-noise tolerance is ensured by the internal refilling as well.

[0053] In the following example a recording of 36 minutes is considered. The recording starts with the last part of a movie. Seconds [646, 866] contain commercials. In second 866 a TV show starts. Other commercials are in the interval [1409, 1735] seconds. FIG. 4 plots with circles the local signal powers calculated during each detected silence. The crosses represent the backward average of these values. It is evident that commercial silences (in the intervals [646, 866] and [1409, 1735]) are mainly cut silences, with lower power. It can be roughly noticed the different distribution of the silences inside the commercials. For instance most of them are distant 10-30 sec. Statistical details like the ones shown are used in the likelihood function estimation. FIG. 5 plots the obtained likelihood function. The filled triggering detects [648, 866] and [1408, 1735].

[0054] Possible Variations

[0055] 1) It is possible to buffer a bigger portion of the scale factors. It is also possible to subsample them. The current selection of 32 out of the 96 left channel scale factor has proved to be effective.

[0056] 2) It is possible to choose a different set of audio features. Careful investigation is needed of course before introducing other features.

[0057] 3) As mentioned above, it is possible to implement different kinds of normalizations and combinations of one or several likelihood functions. The current implementation is based on product combination with renormalization. The product is basically a Boolean AND extended from the Boolean set {0, 1} to the continuous interval [0, 1]. It ensures good selectivity. A semi-sum is a sort of extension of the Boolean OR, which does however not ensure sufficient selectivity.

[0058] 4) The choice of triggering the global likelihood function with refilling can be modified, for instance if different windowing modalities and/or different audio features are used.

[0059] 5) Recognition of particular audio sequences, like those regularly put at the beginning and/or end of commercial blocks by many broadcasters, might also be achieved by processing the scale factors directly.

[0060] At the transmitter side it is possible to help the likelihood function by adapting the cut silences in such a way that they are better detected, for example by lowering their signal power, by adapting the silence duration, by increasing the signal power fall rate, and/or by decreasing the power deviation during the silence. On the contrary, it is also possible to lower the detectability of cut silences by increasing their signal power e.g. by introducing noise, by adapting the silence duration, by decreasing the signal power fall rate and/or by increasing the power deviation during the silence. Further, it is also possible to introduce fake cut silences in the signal. In practical embodiments, a fake cut silence of 0.15 seconds with low power similar to that of a cut silence and a separation of 30 seconds will probably spoil the commercial block detection. Fake cut silences are preferably inserted in already existing silences such as speech silences. In that case, they will hardly be noticeable by an average user.

[0061] The algorithm detects commercial blocks in audio-visual material and marks their boundaries. Commercial blocks can then be skipped during any kind of processing like browsing, automatic trailer creation, editing or simple playback. This functionality can be integrated in several kinds of storage systems, with very low additional cost. It can be applied either in real time during acquisition of the data or off line to stored material.

[0062] It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design many alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. This word ‘comprising’ does not exclude the presence of other elements or steps than those listed in a claim. The invention can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In a device claim enumerating several means, several of these means can be embodied by one and the same item of hardware. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Claims

1. A method of estimating a signal power in a compressed audio signal, the audio signal comprising blocks of quantized samples, a given block being provided with a set of scale factors, the method comprising: extracting the set of scale factors from the compressed audio signal, and estimating the signal power in the given block based on a combination of the scale factors.

2. A method as claimed in claim 1, wherein the extracting step and estimating step are performed on only a sub-set of the set of scale factors.

3. A method as claimed in claim 1 or 2, wherein the audio signal comprises a plurality of narrow band sub signals, each narrow band sub signal being subdivided in signal blocks, each block including quantized samples and each block being provided with one or more scale factors, wherein the extracting step is performed on one or more of the sub-signals.

4. A method as claimed in claim 3, wherein the extracting step is performed on only a sub-set of the plurality of sub-signals.

5. A method as claimed in claim 4, wherein the sub-set mainly includes the scale factors of a number of lower frequency sub-signals.

6. A method as claimed in claim 2, 3, 4 or 5, wherein the blocks are sub-divided into at least two subblocks, each subblock being provided with a scale factor, wherein per block only a maximum scale factor is extracted for use in the estimating step.

7. A method as claimed in any of the previous claims, wherein the compressed audio signal is a stereo or multi-channel signal, wherein the extracting step is performed on only a subset of the available channels.

8. A device for estimating a signal power in a compressed audio signal, the audio signal comprising blocks of quantized samples, a given block being provided with a set of scale factors, the device comprising: means for extracting the set of scale factors from the compressed audio signal, and means for estimating the signal power in the given block based on a combination of the scale factors.

9. A silence detector comprising: a device as claimed in claim 8 for obtaining an estimate of the signal power of a compressed audio signal, and means for evaluating the estimate of the signal power in order to detect silences in the audio signal.

10. A receiver for receiving a compressed audio signal, the receiver comprising: an input for obtaining a compressed audio signal, a silence detector as claimed in claim 9 for detecting silences in the compressed audio signal, and means for influencing the audio signal wherein the influencing at least partly depends on the detection of the silences.