JPH04109300A - Impulsive sound discrimination device - Google Patents

Abstract

PURPOSE:To relax the observation conditions of an input sound and to discriminate the impulsive sound with high accuracy by providing a learning function which knows the category of an input pattern and updates a dictionary pattern. CONSTITUTION:The device is provided with an acoustic input part 1, a feature extraction part 2, a segment processing part 3, a recognition processing part 4, a learning pattern memory part 8, and a learning process part 9. Then when a recognition result obtained the recognition processing part 4 is received and a previously known result, a pattern stored in the learning pattern memory part 8 is read out to update a characteristic core. After the characteristic core is updated, a characteristic vector is calculated and the contents of a standard pattern memory part 5 are updated into a new dictionary. Thus, the learning function which knows the category of the input pattern and updates the standard pattern is provided. Consequently, the observation conditions of the input sound are relaxed and the impulsive sound can be identified with high accuracy.

Description

DETAILED DESCRIPTION OF THE INVENTION [Objective of the Invention (Industrial Application Field) The present invention relates to an impact sound identification device for identifying, for example, cannon sounds or the type of firearm.

(Prior Art) Development of an impact sound identification device that uses a voice recognition device to identify cannon sounds and the type of firearm has been progressing. FIG. 5 shows the configuration of a word identification device as a general speech recognition device.

In FIG. 5, audio is input after being converted into an electrical signal by a sensor such as a microphone, converted into a digital signal by an A/D (analog/digital) converter 11, and then input to a sound processing circuit 12. This acoustic processing circuit 12 converts the input audio signal into frequency space by band pass filter processing or Fourier transform, and detects a characteristic parameter time series. Here, as a preprocess, the speech pattern data in which words are formed is cut out from the time series and stored in the data buffer memory 13. The voice pattern data stored in the memory 13 is sequentially sent to the similarity calculation circuit 14. This similarity calculation circuit 14 calculates the similarity of the input speech pattern with word patterns registered in advance in the pattern dictionary file 15, and outputs the word name showing the maximum similarity as an identification result through the control circuit 16. .

Here, it is well known that highly accurate identification performance can be obtained when a composite similarity is used as the above-mentioned similarity. To explain this composite similarity method, samples of a certain category (for example, considering number recognition)
Collect data pronounced as , perform KL expansion, and use a plurality of eigenvectors of the covariance matrix in descending order of eigenvalues to form a dictionary for that category. Assuming that the dictionary for category L is (φ.“Lゝ), the composite similarity S is defined for the input cover turn f by the following equation.

SLL″ [f Σ (f ・ φ (Lゝ) 2/ II
f II 2rl where l f II is the norm of f, (f・φ.3L))
is the inner product of f and φ LL+.

By the way, in the case of voice recognition, the condition is that the position where the voice is generated is placed at a certain distance from the position of a sensor such as a microphone. For example, when the voice is generated toward a microphone or when inputting by phone. etc. in most cases. However, when identifying artillery batteries, etc., observations are often conducted outdoors, and there is no guarantee that the distance will be kept constant. Therefore, even if the voice recognition device having the above configuration is used, it is difficult to accurately identify impact sounds.

(Problems to be Solved by the Invention) As described above, conventional voice recognition devices require difficult observation conditions to be used for identifying impact sounds from artillery batteries, etc.
Even if it is a sound, it is difficult to identify it with high precision, as it depends greatly on the observation distance.

The present invention was made in order to solve the above-mentioned problems, and an object of the present invention is to provide an impact sound identification device that has relaxed observation conditions for input sounds and is capable of identifying impact sounds with high accuracy.

[Structure of the Invention] (Means for Solving the Problem) In order to achieve the above object, the present invention provides a pattern detection means for detecting an impact sound pattern consisting of a time series of characteristic parameters of an input sound, and a pattern detection means for detecting an impact sound pattern consisting of a time series of characteristic parameters of an input sound. storage means for storing the impact sound pattern stored in the storage means; similarity calculation means for calculating the similarity between the impact patterns stored in the storage means and a plurality of pre-registered standard patterns;
In an impact sound identification device that determines the input sound based on the similarity value obtained by this calculation means, the category information is input together with the input sound, and the impact sound pattern in the category is re-registered as a standard pattern. It is configured with a learning function.

(Function) The impact sound identification device with the above configuration has a learning function that updates the standard pattern by knowing the category of the incoming cover pattern, so for example, the incoming cover pattern of a specific category is updated by adding transformations based on the distance. By doing so, identification accuracy can be improved.

(Embodiment) An embodiment of the present invention will be described below with reference to FIGS. 1 to 4.

FIG. 1 shows its configuration. The acoustic input section 1 is equipped with a sensor such as a microphone and an A/D converter. The impact sound captured by the sensor is converted into an electrical signal, and the sound input section 1 is equipped with a sensor such as a microphone and an A/D converter.
/D converted and supplied to the feature extraction unit 2. The feature extractor 2 is composed of, for example, a group of 16 channel band-pass filters, and the output of each band-pass filter is sequentially supplied to the segment processor 3 at regular intervals. This segment processing section 3 extracts the impact sound part as a pattern by referring to the series of band-pass filter outputs and changes in sound volume, and the impact sound pattern obtained here is sent to the recognition processing section 4. . The recognition processing unit 4 compares the input impact sound pattern with standard patterns registered in advance from the memory unit 5, calculates the similarity of each, and displays the type name of the pattern showing the maximum similarity on the recognition result display. 6.

The impact sound pattern obtained by the segment processing section 3 is as follows:
The data is also supplied to the learning pattern memory unit 8 via the identification/learning mode changeover switch 7 . This learning pattern memory section 8 stores the impact sound pattern supplied during the learning mode as a learning pattern. The learning processing section 9 is activated by the learning mode designation, takes in the recognition result obtained by the recognition processing section 4, and reads out the pattern stored in the learning pattern memory section 8 if the recognition result differs from the previously known result. Update the characteristic kernel. When the characteristic kernel has been updated in this way, the eigenvector is calculated and the contents of the standard pattern memory section 5 are updated as a new dictionary.

The learning processing unit 9 may be configured to create a pattern by subjecting the learning pattern to a plurality of types of low-pass filter processing, and update the characteristic kernel using this pattern.

In the above configuration, the processing means will be explained below with reference to FIGS. 2 to 4.

That is, the impact sound identification device is characterized by having a learning function that updates the dictionary pattern (standard pattern) by knowing the category of the input pattern. This is a learning pattern for As a dictionary stored in the standard pattern memory section 5, a set of patterns lx, l i=1. ..., n
), define the following covariance matrix (characteristic kernel).

Kwx "XI" where X, T are the light chain vectors of XI. Item 20 represents the characteristics of the distribution of the pattern set and is called the characteristic kernel. A plurality of K eigenvectors are used in descending order of eigenvalue to form a dictionary pattern. When a pattern g belonging to the same category is misrecognized, it is assumed that the features of g are not reflected in K, and K is updated with +aggT to K' -, and multiple eigenvectors of K' are used in descending order of eigenvalue. and create a new dictionary pattern.

Furthermore, by creating g' by adding a modification to g according to the distance, and setting K' - to +αgg''+βg'g''', it is possible to update the dictionary pattern according to the distance.

To explain with an example, if the front of a cannon is observed at a position approximately 2.5 km away and a signal waveform as shown in Figure 2 is obtained, then the front of a cannon of the same type is observed at a position approximately 10 km away. In this case, the signal waveform is greatly deformed as shown in FIG. Therefore, g' is created by applying low-pass filter processing to data at a short distance, since the attenuation of high-frequency components generally increases as the distance increases. That is, when the data in FIG. 2 is subjected to low-pass filter processing at approximately 30 Hz, a waveform as shown in FIG. 4 is obtained, which approximates the waveform shown in FIG. 3.

Therefore, the impact sound identification device configured as described above can update the standard pattern according to the deformation of the pattern depending on the observation distance, so that the observation distance condition for the input sound can be relaxed and impact sounds can be identified with high accuracy. I can do it.

Although the above embodiments have been explained with a focus on observation distance, the present invention is not limited to this, and can also be applied to various conditions, such as when the transmission characteristics of sound waves change depending on weather conditions. Of course.

[Effects of the Invention] As described above, according to the present invention, it is possible to provide an impact sound identification device that has relaxed observation conditions for input sounds and can identify impact sounds with high accuracy.

[Brief explanation of the drawing]

FIG. 1 is a block circuit configuration diagram showing an embodiment of an impact sound identification device according to the present invention, and FIGS. 2 to 4 are waveform diagrams shown to explain impact sound pattern processing of the same embodiment, respectively. FIG. 5 is a block circuit diagram showing the configuration of a word identification device based on voice recognition used as a conventional impact sound identification device. DESCRIPTION OF SYMBOLS 1... Acoustic input section, 2... Feature extraction section, 3... Segment processing section, 4... Recognition processing section, 5... Standard pattern memory section, 6... Recognition result display section, 7. ··identification/
learning mode changeover switch, 8... learning pattern memory section, 9... learning processing section, 11... A/D converter,]
2... Acoustic processing circuit, 13... Data buffer memory, 14... Similarity calculation circuit, 15... Pattern dictionary file, 16... Control circuit. Applicant's Representative Patent Attorney Takehiko Suzue

Claims (2)

[Claims] (1) Regarding a pattern detection means for detecting an impact sound pattern consisting of a time series of characteristic parameters of an input sound, a storage means for storing an impact sound pattern obtained by this pattern means, and an impact pattern stored in this storage means In an impact sound identification device that determines an input sound based on a similarity calculation means that calculates the similarity with a plurality of pre-registered standard patterns, and a similarity value obtained by the calculation means, the input Enter the category information along with the sound,
An impact sound identification device having a learning function that re-registers an impact sound pattern in the category as a standard pattern. (2) The learning function predicts from the input sound a change in the feature parameter due to the difference between the observation position and the sound source position, creates a modified sound thereof, and adds it to the re-registered pattern. The impact sound identification device described.