JPH07199995A - Discriminating device for sound source - Google Patents

Abstract

PURPOSE:To eliminate the influence of a noise in the case where the name of a sound source is determined from a received acoustic signal. CONSTITUTION:The received acoustic signal is drawn 10 into a sound print image and a narrow-band signal is detected through a line extracting process 11 to reduce the influence of the noise. Further, the signal is classified 12 into higher harmonic groups on the basis of relation among extracted lines and then a cepstrum (logarithmic transformation 4 and inverse Fourier transformation 5) is performed; and pattern matching 7 is performed, so signals from plural sound sources can be determined at the same time.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sound source identification device,
In particular, the present invention relates to a device for identifying a sound source from a signal processing result of an acoustic signal.

[0002]

2. Description of the Related Art Conventionally, a sound source identification device of this type obtains a cepstrum (a Fourier transform of an acoustic signal, a logarithmic transform, and then an inverse Fourier transform) from the acoustic signal, as is done, for example, in speech recognition. The sound source was identified by matching the temporal change with a standard pattern created based on past data.

FIG. 2 is a block diagram showing an example of a conventional sound source identification device. In the figure, the microphone 1 converts an acoustic signal into an analog electric signal. The A / D conversion circuit 2 converts an analog electric signal into a digital signal. The Fourier transform circuit 3 obtains a power spectrum of the acoustic signal by frequency-analyzing the acoustic signal converted into a digital signal. The logarithmic conversion circuit 4 logarithmically converts the signal amplitude of the power spectrum. The inverse Fourier transform circuit 5 performs an inverse Fourier transform on the logarithmically transformed acoustic signal. The standard pattern circuit 6 stores a standard pattern created by performing the above operation on the acoustic signals accumulated in the past as a database. The pattern standard circuit 7 compares the stored standard pattern with the pattern obtained by performing the above-described processing on the acoustic signal observed by the microphone 1, and outputs the sound source name to the closest pattern.

Next, the operation will be described. Microphone 1
The acoustic signal input by means of the characteristics of the sound source includes the frequency of the vibration of the sound source and the frequency characteristic given until the vibration is transmitted from the sound source to the microphone 1. In the power spectrum obtained by the A / D conversion circuit 2 and the Fourier transform circuit 3, the product of the power spectrum of the sound source vibration and the frequency characteristic given at the stage when the vibration is transmitted appears. The power spectrum is logarithmically converted by the logarithmic conversion circuit 4, so that the power spectrum of the sound source vibration that has appeared in the form of a product and the frequency characteristic of the transfer path are converted into a sum form. In this way, after separating the two features possessed by the sound source, collation with the stored standard pattern is performed.

[0005]

In this conventional sound source identification apparatus, when the signal-to-noise ratio (S / N) is small (when noise is large), the result of the inverse Fourier transform (input to the pattern matching circuit) is noise. It was difficult to discriminate because of the large change due to the influence of and the large difference from the standard pattern.
Further, when a plurality of sound source signals are mixed in the acoustic signal input to the microphone, the signals cannot be separated for each sound source, which makes it difficult to identify.

An object of the present invention is to identify a sound source with a high reliability even when a signal-to-noise ratio (S / N) of an acoustic signal is small and a plurality of sound source signals are observed mixedly. To provide the equipment.

[0007]

Therefore, in the sound source identification device of the present invention, the noise component is removed from the frequency analysis result, and the acoustic signal input from the microphone and the line extraction circuit for extracting only the signal component are Fourier-transformed. Fourier transform circuit and logarithmic transform circuit and inverse Fourier transform to separate the frequency characteristic in the transmission by classifying the signal line that has a harmonic relationship from the variation characteristic of the signal frequency to classify the signal for each sound source And an inverse Fourier transform circuit for performing pattern matching with a standard pattern.

[0008]

The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the present invention.
First, the microphone 1 converts an acoustic signal into an analog electric signal and inputs it to the A / D conversion circuit 2. A / D conversion circuit 2
Converts the input analog electric signal into a digital signal. The Fourier transform circuit 3 frequency-analyzes the acoustic signal converted into a digital signal to obtain a power spectrum. The moving average circuit 8 and the time integration circuit 9 remove noise components from the power spectrum. Sound print image drawing circuit 1
0 draws a voice print image in which the intensity of the signal amplitude of the power spectrum is represented by shading. The line extraction circuit 11 performs a signal peak detection process 11 for detecting the maximum amplitude point of the power spectrum, a peak tracking process 112 for tracking the detected peak in the time direction, and a tracking in the true line direction when the line is branched. It comprises a line branching determination process 113 for scanning and a line determination process 114 for determining whether the tracked line is a signal or noise. The harmonic detection circuit 12 classifies the extracted signal line into those having a relation of an integral multiple of a certain fundamental frequency 1
21 and a harmonic determination processing 122 for determining that there is a harmonic relationship based on the temporal variation characteristic. The logarithmic conversion circuit 4 logarithmically converts the amplitude of the power spectrum. The inverse Fourier transform circuit 5 performs an inverse Fourier transform on the logarithmically transformed signal. The pattern matching circuit 7 outputs the sound source name of the closest standard pattern by matching the standard pattern 6 stored as a database with the pattern obtained by the above operation. In the illustrated embodiment, the harmonic detection circuit 12 and the logarithmic conversion circuit 4 can be arranged in any order.

Next, the operation of the embodiment shown in FIG. 1 will be described. It is considered that the sound emitted from a single sound source is obtained by multiplying a single frequency by a transfer function depending on the characteristics of the material and shape of the sound source. It is considered that the acoustic signal observed by the microphone 1 includes a plurality of sounds emitted from such a single sound source and also includes weekly noise.

Here, it is assumed that the influence of the propagation path and the observation error party have little influence. According to the above assumption,
The observed signal is approximated by Equation 1 shown below.

X (n) = f ₁ (n) × g ₁ (n) + f ₂ (n) × g ₂ (n) + ... + f _k (n) × g _k (n) + N (Equation 1) X (N): observed signal f _k (n): sound source signal g _k (n): transfer function N: noise The A / D conversion circuit 2 discretizes the observed signal, and the Fourier transform circuit 3 calculates the power spectrum of the observed signal. Ask Koutor. The moving average circuit 8 and the time integration circuit 9 average out the power spectrum in the frequency direction and the time direction to remove the influence of noise and observation error. Depending on the observation situation, it is possible without the moving average circuit and the integrating circuit.

Next, in the line detection circuit 11, the peak (maximum point) of the obtained amplitude of the power spectrum is tracked in the time direction. Tracking is performed under the condition that the signal frequency does not change rapidly. When the signal lines branch and intersect in the middle, the true signal run-in is tracked under the judgment conditions such as the similarity of the signal characteristics.
Next, it is judged whether the line is a signal or noise from the integrated energy of the time continuity and the signal strength of the tracked line. In the harmonic detection circuit 12, the signals having the harmonic relationship are in an integral multiple of the same fundamental frequency, and have the same time variation characteristic of the frequency (when the fundamental frequency fluctuates, the signal frequencies having the harmonic relationship are The observed signals, which are multiple sound source signals, are separated for each single sound source based on the rule that the relationship of being an integral multiple of the fundamental frequency is maintained, that is, f ₁ × g ₁ , f ₂ × in Expression 1. f ₃ is separated into .... further logarithmic transformation by the logarithmic converter circuit 4, it separates the signal frequency (f _k) and the transfer function (g _k). separated signal with the transfer function inverse Fourier transform circuit 5 Inverse Fourier transform is performed, and the pattern matching circuit 7 identifies what the sound source is by matching a plurality of standard patterns created in advance by the same operation with the pattern obtained by the above operation for the observed acoustic signal. The matching method uses the squared error of the two patterns as an evaluation function, the source name of the standard pattern with the smallest error as the source name of the input signal, and a neural network that has learned the standard pattern. There are methods such as identifying the sound source name.

By using a plurality of microphones and utilizing directivity, it is possible to improve the separation performance of a plurality of sound source signals and the establishment of identification.

[0014]

As described above, since the sound source identification device according to the present invention performs the line extraction processing on the input signal,
The probability of sound source identification is high even when the noise is large relative to the signal level. Further, since the signal is separated for each sound source by the harmonic detection processing, even when a plurality of sound source signals are observed mixedly, it is possible to identify them simultaneously.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

FIG. 2 is a block diagram of a conventional sound source identification device.

[Explanation of symbols]

 1 Microphone 2 A / D conversion circuit 3 Fourier transform circuit 4 Logarithmic transform circuit 5 Inverse Fourier transform circuit 6 Standard pattern 7 Pattern matching circuit 8 Moving average circuit 9 Time integration circuit 10 Voice print image drawing circuit 11 Line extraction circuit 111 Signal peak detection Circuit 112 Peak tracking processing 113 Line branch determination processing 114 Line determination processing 12 High frequency detection circuit 121 Basic frequency detection processing 122 High frequency determination processing

Claims (3)

[Claims] 1. A microphone for inputting an acoustic signal,
An A / D conversion circuit that discretizes the input acoustic signal, a Fourier transform circuit that performs a Fourier transform, a moving average circuit that performs a moving average in the frequency direction, a time integration circuit that integrates in the time direction, and a signal that forms a line A line extraction circuit for extracting, a harmonic detection circuit for classifying lines having a harmonic relationship from the signal frequency variation characteristic, a logarithmic conversion circuit for logarithmically converting the amplitude of the classified lines, and an inverse Fourier transform circuit for inverse Fourier transform And a standard pattern circuit for accumulating the standard pattern, and a pattern matching circuit for matching the pattern obtained by the operation with the standard pattern. 2. The sound source identification circuit according to claim 1, wherein the line extraction circuit includes means for detecting an amplitude maximum point of the power spectrum and means for tracking the detected peak in the time direction. 3. The received acoustic signal is converted into voice print information,
A sound source identification method characterized by extracting lines of voiceprint information, classifying frequency relationships between the extracted lines, performing cepstrum processing, and performing pattern matching.