This application is a 35 U.S.C. §371 national stage application of PCT International Application No. PCT/EP2011/056761, filed on 28 Apr. 2011, the disclosure and content of which is incorporated by reference herein in its entirety. The above-referenced PCT International Application was published in the English language as International Publication No. WO 2012/146290 A1 on 1 Nov. 2012.
The present technology relates to frame based audio signal classification.
Audio signal classification methods are designed under different assumptions: real-time or off-line approach, different memory and complexity requirements, etc.
For a classifier used in audio coding the decision typically has to be taken on a frame-by-frame basis, based entirely on the past signal statistics. Many audio coding applications, such as real-time coding, also pose heavy constraints on the computational complexity of the classifier.
Reference [1] describes a complex speech/music discriminator (classifier) based on a multidimensional Gaussian maximum a posteriori estimator, a Gaussian mixture model classification, a spatial partitioning scheme based on k-d trees or a nearest neighbor classifier. In order to obtain an acceptable decision error rate it is also necessary to include audio signal features that require a large latency.
Reference [2] describes a speech/music discriminator partially based on Line Spectral Frequencies (LSFs). However, determining LSFs is a rather complex procedure.
Reference [5] describes voice activity detection based on the Amplitude-Modulated (AM) envelope of a signal segment.
An object of the present technology is low complexity frame based audio signal classification.
This object is achieved in accordance with the attached claims.
A first aspect of the present technology involves a frame based audio signal classification method including the following steps:
A second aspect of the present technology involves an audio classifier for frame based audio signal classification including:
A third aspect of the present technology involves an audio encoder arrangement including an audio classifier in accordance with the second aspect to classify audio frames into speech/non-speech and thereby select a corresponding encoding method.
A fourth aspect of the present technology involves an audio codec arrangement including an audio classifier in accordance with the second aspect to classify audio frames into speech/non-speech for selecting a corresponding post filtering method.
A fifth aspect of the present technology involves an audio communication device including an audio encoder arrangement in accordance with the third or fourth aspect.
Advantages of the present technology are low complexity and simple decision logic. These features make it especially suitable for real-time audio coding.
The technology, together with further objects and advantages thereof, may best be understood by making reference to the following description taken together with the accompanying drawings, in which:
In the following description m denotes the audio sample index in a frame and n denotes the frame index. A frame is defined as a short block of the audio signal, e.g. 20-40 ms, containing M samples.
The present technology is based on a set of feature measures that can be calculated directly from the signal waveform (or its representation in a frequency domain, as will be described below) at a very low computational complexity.
The following feature measures are extracted from the audio signal on a frame by frame basis:
The feature measures Tn, En, ΔEn are calculated for each frame and used to derive certain signal statistics. First, Tn, En, ΔEn are compared to respective predefined criteria (see first two columns in Table 1 below), and the binary decisions for a number of past frames, for example N=40 past frames, are kept in a buffer. Note that some feature measures (for example Tn, En in Table 1) may be associated with several criteria. Next, signal statistics (fractions) are obtained from the buffered values. Finally, a classification procedure is based on the signal statistics.
Column 2 of Table 1 describes examples of the different criteria for each feature measure Tn, En, ΔEn. Although these criteria seem very different at first sight, they are actually equivalent to the feature intervals illustrated in column 3 in Table 1. Thus, in a practical implementation the criteria may be implemented by testing whether the feature measures fall within their respective feature intervals. Example feature intervals are given in column 4 in Table 1.
In Table 1 it is also noted that, in this example, the first feature interval for the feature measure En is defined by an auxiliary parameter EnMAX. This auxiliary parameter represents signal maximum and is preferably tracked in accordance with:
As can be seen from
An alternative to the described tracking method is to use a large buffer for storing past frame energy values. The length of the buffer should be sufficient to store frame energy values for a time period that is longer than the longest expected pause, e.g. 400 ms. For each new frame the oldest frame energy value is removed and the latest frame energy value is added. Thereafter the maximum value in the buffer is determined.
The signal is classified as speech if all signal statistics (the fractions Φi in column 5 in Table 1) belong to a pre-defined fraction interval (column 6 in Table 1), i.e. ∀Φiε{T1i,T2i}. An example of fraction intervals is given in column 7 in Table 1. If one or more of the fractions Φi is outside of the corresponding fraction interval {T1i,T2i}, the signal is classified as non-speech.
The selected signal statistics or fractions Φi are motivated by observations indicating that a speech signal consists of a certain amount of alternating voiced and un-voiced segments. A speech signal can typically also be active only for a limited period of time and is then followed by a silent segment. Energy dynamics or variations are generally larger in a speech signal than in non-speech, such as music, see
In the examples given above, the feature measures given in (1)-(4) are determined in the time domain. However, it is also possible to determine them in the frequency domain, as illustrated by the block diagram in
Equations (2) and (3) can be replaced by summation over frequency bins Xk(n) instead of input samples xm(n), which gives:
respectively.
Similarly, equation (4) may be replaced by:
The description above has focused on the three feature measures Tn, En, ΔEn to classify audio signals. However, further feature measures handled in the same way may be added. One example is a pitch measure (fundamental frequency) {circumflex over (P)}n, which can be calculated by maximizing the autocorrelation function:
It is also possible to perform the pitch estimation in the cepstral domain. Cepstral coefficients cm(n) are obtained through inverse Discrete Fourier Transform (DFT) of log magnitude spectrum. This can be expressed in the following steps: perform a DFT on the waveform vector; on the resulting frequency vector take the absolute value and then the logarithm; finally the Inverse Discrete Fourier Transform (IDFT) gives the vector of cepstral coefficients. The location of the peak in this vector is a frequency domain estimate of the pitch period. In mathematical notation:
The steps, functions, procedures and/or blocks described herein may be implemented in hardware using any conventional technology, such as discrete circuit or integrated circuit technology, including both general-purpose electronic circuitry and application-specific circuitry.
Alternatively, at least some of the steps, functions, procedures and/or blocks described herein may be implemented in software for execution by a suitable processing device, such as a micro processor, Digital Signal Processor (DSP) and/or any suitable programmable logic device, such as a Field Programmable Gate Array (FPGA) device.
It should also be understood that it may be possible to reuse the general processing capabilities of the encoder. This may, for example, be done by reprogramming of the existing software or by adding new software components.
Although most of the example embodiments above have been illustrated in the time domain, it is appreciated that they may also be implemented in the frequency domain, for example for transform coders. In this case the feature extractor 14 will be based on, for example, some of the equations (6)-(10). However, once the feature measures have been determined, the same elements as in the time domain implementations may be used.
With an embodiment based on equations (1), (2), (4), (5) and Table 1, the following performance was obtained for audio signal classification:
The audio classification described above is particularly suited for systems that transmit encoded audio signals in real-time. The information provided by the classifier can be used to switch between types of coders (e.g., a Code-Excited Linear Prediction (CELP) coder when a speech signal is detected and a transform coder, such as a Modified Discrete Cosine Transform (MDCT) coder when a music signal is detected), or coder parameters. Furthermore, classification decisions can also be used to control active signal specific processing modules, such as speech enhancing post filters.
However, the described audio classification can also be used in off-line applications, as a part of a data mining algorithm, or to control specific speech/music processing modules, such as frequency equalizers, loudness control, etc.
It will be understood by those skilled in the art that various modifications and changes may be made to the present technology without departure from the scope thereof, which is defined by the appended claims.
CELP Code-Excited Linear Prediction
DFT Discrete Fourier Transform
DSP Digital Signal Processor
FPGA Field Programmable Gate Array
IDFT Inverse Discrete Fourier Transform
LSFs Line Spectral Frequencies
MDCT Modified Discrete Cosine Transform
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/EP2011/056761 | 4/28/2011 | WO | 00 | 10/24/2013 |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO2012/146290 | 11/1/2012 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
5579435 | Jansson | Nov 1996 | A |
5712953 | Langs | Jan 1998 | A |
6640208 | Zhang | Oct 2003 | B1 |
7127392 | Smith | Oct 2006 | B1 |
20020165713 | Skoglund | Nov 2002 | A1 |
Number | Date | Country |
---|---|---|
2 096 629 | Dec 2008 | EP |
WO 9839768 | Sep 1998 | WO |
WO 0217299 | Feb 2002 | WO |
Entry |
---|
Written Opinion of the International Searching Authority, PCT/EP2011/056761, Jan. 12, 2012. |
J-H. Chen, A. Gersho, “Adaptive Postfiltering for Quality Enhancement of Coded Speech”, IEEE Transactions on Speech and Audio Processing, vol. 3, No. 1, Jan. 1993, pp. 59-71. |
International Search Report, PCT/EP2011/056761, Jan. 12, 2012. |
E. Scheirer and M. Slaney, “Construction and Evaluation of a Robust Multifeature Speech/Music Discriminator”, ICASSP '97 Proceedings of the 1997 IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 2, p. 1331-1334, 1997. |
K. El-Maleh, M. Klein, G. Petrucci, P. Kabal, “Speech/music discrimination for multimedia applications”, available at: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.93.3453&rep=rep1&type=pdf. |
Number | Date | Country | |
---|---|---|---|
20140046658 A1 | Feb 2014 | US |