JP4219539B2 - Acoustic classification device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a sound classification apparatus for classifying sounds such as voice and music.
[0002]
[Prior art]
The following documents have been published in the past as documents disclosing a technology for automatically identifying sounds such as voice and music.
[0003]
Reference 1: J.A. Audio Eng. Soc. , Vol. 47, No9, 1999 Sep. 720-725
A method is disclosed in which an acoustic signal is frequency-analyzed and the acoustic signal is identified using a difference in frequency characteristics between music and speech. This method has a disadvantage that the detection accuracy of features is low because the main frequency components of music and speech cannot generally be separated.
[0004]
Reference 2: Proc. ICASSP 96,993-996, author John
Saunders
Disclosed is a method of describing speech and music using a multivariate discriminant function by describing five types of feature vectors based on four types of statistics and energy extracted from energy based on the zero-cross distribution of acoustic signals. Has been. This method has the disadvantage that the statistics extracted from the zero cross account for 80% of the total, the identification result depends on the zero cross distribution, and the classification system drops depending on the type of sound.
[0005]
Reference 3: Proc. ICASSP 97, 1331-1334, Authors Eric Scheider, Malcom Slaney
A method for classifying speech and music using a multivariate discriminant function based on 13 types of feature vectors such as the number of zero crosses obtained by analyzing an acoustic signal within a certain window, frequency centroid, time ratio of low power level, etc. Is disclosed. This method has as many as 13 types of acoustic features, and takes a long time to obtain a classification result, and is not suitable for high-speed classification processing.
[0006]
[Problems to be solved by the invention]
As described above, the methods described in the documents 1 to 3 cannot satisfy both the processing speed and the classification accuracy.
[0007]
Therefore, an object of the present invention is to provide an acoustic classification device that can perform classification at high speed and with high classification accuracy.
[0008]
[Means for Solving the Problems]
In order to achieve such an object, the invention of claim 1 is an acoustic classification device for classifying an acoustic signal, wherein the acoustic characteristic is extracted from the acoustic signal as a feature of the sound for each predetermined time length. The acoustic features of the learning acoustic signal whose classification contents are known are acquired by the extraction means and the feature extraction means, and the neural network is trained so that the acquired acoustic features are input and the corresponding classification contents are output. The acoustic features of the acoustic signal to be classified are acquired by the learning means and the feature extracting means, the acquired acoustic features are input to the neural network, and the output of the neural network is received to obtain the classification target acoustic signal. comprising an identification means for classifying acoustic signal, the statistical properties of the acoustics, zero cross distribution of the acoustic signal, the level distribution and maximum power The feature extraction means obtains a result of the analysis by performing principal component analysis of a statistical distribution shape indicating acoustic features from a plurality of types of pre-processing sound source signals relating to audio signals and music signals in advance. The characteristic of the acoustic signal of the learning acoustic signal and the characteristic of the acoustic signal to be classified are acquired using an eigenvector .
[0009]
According to a second aspect of the present invention, in the acoustic classification apparatus according to the first aspect, the level distribution is given as a frequency distribution of relative levels with respect to a time average level in a frame, and the frequency distribution giving the maximum power is 1 The frame is shifted by predetermined L points, the appearance frequency of the frequency having the maximum power within each time window is obtained, and this is used as a physical statistic .
[0011]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0012]
FIG. 1 shows a functional configuration of an acoustic classification apparatus to which the present invention is applied. A general-purpose computer capable of executing a program such as a personal computer can be used as the acoustic analysis device. The components described below are realized by a CPU or the like executing a program described later.
[0013]
In FIG. 1, reference numeral 100 denotes a statistical property of sound for each predetermined time length extracted from a plurality of types of sound source signals such as a voice signal and a music signal, and a principal component analysis of the characteristic property is performed. It is a preprocessing unit for acquiring eigenvectors.
[0014]
200 acquires the acoustic feature of the learning acoustic signal whose classification content is known, acquires the main component of the learning acoustic signal feature using the eigenvector obtained by the pre-processing unit 100, and This is a learning unit that causes the neural network to learn such that the main component of the feature is input and the corresponding classification content is output.
[0015]
300 acquires the acoustic characteristics of the acoustic signal to be classified by the pre-processing unit 100 to be classified, acquires the main component of the acoustic signal to be classified using the eigenvector obtained by the pre-processing unit 100, The identification unit classifies the acoustic signals to be classified by inputting the principal components of the acoustic features to the neural network and receiving the output of the neural network.
[0016]
FIG. 2 shows the processing contents of a program for realizing the function of the preprocessing unit 100. Hereinafter, each process of FIG. 2 will be described.
[0017]
(1-1) Time division processing (step S10)
A plurality of types of pre-processing sound source signals such as audio signals and music signals are input from a microphone, converted from an analog signal to a digital signal in a general-purpose computer, temporarily stored in an internal memory, and then subjected to the following processing by the CPU. . That is, when obtaining a statistical feature quantity in the time-time axis direction of an acoustic signal collected during a predetermined time, the acoustic signal is a frame in which N blocks are arranged adjacent to each other as shown in FIG. It is divided in time. As will be described later, the acoustic signal at point K on the time axis is shifted by L points as shown in FIG. 5B within the frame for frequency analysis.
[0018]
(1-2) Extraction of acoustic features (Step S20)
In the present embodiment, statistical feature amounts of zero cross distribution, level distribution, and frequency distribution are used as feature amounts reflecting physical differences between voice and music. The characteristics of these distributions will be described below.
[0019]
(A) Zero-cross distribution This physical quantity is a distribution on the time axis of the number of times that the signal crosses the zero level in a certain time (one block). The zero-cross distribution is highly correlated with the frequency of an acoustic signal having a large power. For example, a signal with a dominant high frequency component tends to increase the number of zero crossings.
[0020]
In the present embodiment, the frequency distribution is obtained by calculating the number of zero crosses of adjacent N blocks for each M samples of one block and averaging them for one frame. The shape of the frequency distribution pattern of the number of cellos is adopted as a statistic.
[0021]
In the case of audio signals, in addition to vowels, consonants + vowels (number of zero crossings: small), friction sounds and plosives (number of zero crossings: many) are mixed, so the average frequency distribution of one frame has a large number of zero crossings and a small value. The shape separated into two poles. On the other hand, in the case of music, it shows the top of a single Yamagata.
[0022]
(B) Level distribution This physical quantity is calculated based on the relative level of the time average level of one block M sample sample to the time average level (0 dB) of the sound signal of one frame within the frame of this relative level. It is a thing. In the case of an audio signal, since there is often a silent section between the audio and audio signals, the distribution of the frequency distribution tends to be larger than that of music.
[0023]
(C) Frequency distribution giving maximum power (hereinafter, frequency distribution)
Amplitude frequency characteristics obtained by FFT (fast Fourier analysis) of the time waveform of the K point that has been windowed are rearranged on the logarithmic frequency axis so as to correspond to hearing, and the physical quantity representing the time is the frequency that gives the maximum power. It was. Each frame was shifted by L points, the appearance frequency of the frequency having the maximum power within each time window was determined, and this was adopted as a physical statistic. In the case of an audio signal, the level distribution is represented on a document separated into two poles of a high frequency and a low frequency as in the zero cross distribution. By using this statistic, it is possible to classify music signals with a narrower range like a single instrument and those with a wider range like an orchestra by the shape of the distribution pattern, so that music signals can be further classified. .
[0024]
(1-3) Statistics calculation and averaging process (steps S30 and S40)
Since the acoustic feature amount described above is extracted from the acoustic signal (learning acoustic signal and acoustic signal to be classified) for each frame, the acoustic statistical feature amount is obtained using the acoustic feature amount extracted from a plurality of frames of the acoustic signal. Calculate and average statistics for each acoustic signal.
[0025]
(1-4) Principal component analysis processing and eigenvector storage (step 50, S60)
In this embodiment, the statistical amount of one frame is a multidimensional vector, but in order to reduce the number of dimensions of the acoustic feature amount, the principal component analysis process, more specifically, the feature of the statistical distribution shape is maintained. As it is (for example, factor loading of 90% or more), a process of extracting the main component is performed. In addition, the eigenvector obtained as a result of the principal component analysis is stored in a storage unit inside the apparatus so that it can be used as a transformation matrix for obtaining the principal component of the statistic in the subsequent processing.
[0026]
FIG. 3 shows processing contents for realizing the function of the learning unit 200 of FIG. 1 by a program.
[0027]
A learning acoustic signal whose contents are input from a microphone or the like is converted from an analog signal to a digital signal and then processed by a general-purpose computer.
[0028]
(2-1) Time division process to statistic calculation process (steps S100 to S120)
The time division process, the acoustic feature quantity extraction process, and the statistic calculation process have the same processing contents as the pre-process described above, and a common program (subroutine program) can be used for these processes.
[0029]
(2-2) Principal component extraction process (step S130)
To obtain principal components using eigenvectors, multivariate analysis, which is a known technique, is performed. The multivariate analysis is described in detail in, for example, “Sadao Ishimura, Multivariate analysis that can be easily understood, Chapter 4 (Tokyo Book)”.
[0030]
(2-3) Learning of Neural Network The neural network is made to learn the main components of the characteristics of the learning acoustic signal obtained in the above processing steps and the known classification data indicating the type content of the learning acoustic signal.
[0031]
The neural network and its learning method are well known, but since they relate to the invention, they will be briefly described.
[0032]
A typical configuration of the neural network is shown in FIG. FIG. 1 shows an input layer, 2 an intermediate layer, and 3 an output device. Neurons can be used for each layer. The neuron is represented by a function in which the correlation between the input signal and the output signal is determined in advance (in this embodiment, a sigmoid function that is mathematically handled is used). Therefore, when an input signal having a certain value is input to the input layer 1, an output signal having a predetermined value is output from the output layer 3. Therefore, the main component of the feature obtained from the learning acoustic signal is input to the input layer 1, and the correlation function of each neuron (both the transfer function is set so that data indicating the classification content is output from the output layer 3. Learning the coefficient). There are many proposals as a learning method, but a typical example is that a certain coefficient is initially given to the correlation function of each neuron, an input signal is given to the input layer 1 of the neural network, and the output from the output layer 3 Calculate the output signal. If the output signal value is far from the target classification data value, try changing the initial value little by little until the output signal value reaches the target classification data value. Repeat with mistakes. When the value of the output signal falls within the allowable range, the coefficient of the correlation function used at that time becomes the learning result. The above learning process can be realized by executing a computer program. When a plurality of sets of input signals are given, the coefficient of the correlation function is detected so that the output signal becomes the value of the corresponding classification data when each of the input signals is given.
[0033]
FIG. 4 shows processing contents for realizing the function of the identification unit 300 of FIG. 1 by a program.
[0034]
(3-1) Time division processing to principal component extraction processing (steps S200 to S230)
The processing contents of the time division processing to the statistical processing are the same as the learning processing. However, the difference is that the acoustic signal input to the acoustic classification device is an unknown acoustic signal.
[0035]
(3-2) Neural network calculation process Since the correlation function (more precisely, coefficient) of each neuron of the neural network is determined in the above learning process, the value of the output signal is calculated using these correlation functions. . As is well known, when a neural network learns the relationship between an input signal and an output signal and then gives the input signal to the neural network, it outputs an output signal corresponding to the learning input signal similar to the input signal. It has the nature of Therefore, the classification data value obtained as a result of the calculation becomes the classification identification result. The classification result may be displayed on a display of an acoustic classification device (general-purpose computer) or may be printed out.
[0036]
It is not limited to embodiment described above, A various deformation | transformation is possible according to a use. For example, the number of neurons constituting the neural network may be appropriately determined according to the application. The acoustic classification can be used for high-speed search of an acoustic database, automatic indexing of acoustic signals and image search using this index as side information, preprocessing for speech recognition processing, and automatic monitoring of broadcast audio.
[0037]
【The invention's effect】
As described above, according to the present invention, it is possible to speed up the classification process by using a neural network for the classification of acoustic signals, and the acoustic zero-cross distribution, level as acoustic features Since statistical properties such as distribution and frequency distribution are used, it is also possible to classify music by a single musical instrument or music such as orchestra, and obtain discrimination performance comparable to human judgment. Furthermore, it is possible to make various kinds of acoustic signals to be identified. As a result, it is possible to contribute to the automatic monitoring of broadcast sound, the automatic indexing and searching of the sound database, and the efficiency of searching using multimedia. .
[Brief description of the drawings]
FIG. 1 is a block diagram showing a functional configuration of an embodiment of the present invention.
FIG. 2 is a flowchart showing the contents of preprocessing.
FIG. 3 is a flowchart showing the contents of a learning process.
FIG. 4 is a flowchart showing the contents of identification processing.
FIGS. 5A and 5B are explanatory diagrams for explaining signal division processing on a time axis. FIGS.
FIG. 6 is an explanatory diagram for explaining a neural network.
[Explanation of symbols]
100 Pre-processing unit 200 Learning unit 300 Identification unit

Claims (2)

An acoustic classification device for classifying acoustic signals,
A feature extraction means for extracting a statistical property of the sound for each predetermined time length as an acoustic feature from the acoustic signal;
Learning means for acquiring acoustic features of a learning acoustic signal whose classification contents are known by the feature extraction means, and learning the neural network so that the acquired acoustic features are input and corresponding classification contents are output;
The acoustic feature of the acoustic signal to be classified is acquired by the feature extraction unit, the acquired acoustic feature is input to the neural network, and the output of the neural network is received to classify the acoustic signal to be classified. comprising an identification means that,
The acoustic statistical properties are zero cross distribution, level distribution and frequency distribution giving maximum power of the acoustic signal;
The feature extraction means performs a principal component analysis of a statistical distribution shape indicating acoustic features from a plurality of types of pre-processing sound source signals related to a speech signal and a music signal in advance, and uses the eigenvector obtained as a result of the analysis, An acoustic classification apparatus characterized by acquiring acoustic characteristics of a learning acoustic signal and characteristics of the classification target acoustic signal . 2. The acoustic classification apparatus according to claim 1, wherein the level distribution is given as a frequency distribution of a relative level with respect to a time average level in a frame, and the frequency distribution giving the maximum power is a predetermined L point in one frame. An acoustic classification device characterized by shifting each time, obtaining an appearance frequency of a frequency having the maximum power within each time window, and adopting this as a physical statistic .

Images