KR101022516B1 - System and method for sound recognition using spectral peak, and method for measuring sound spectral similarity used thereto - Google Patents

Abstract

The present invention relates to a sound recognition technology, and an acoustic recognition system using a spectral peak according to the present invention includes: a sound register for acquiring a peak band in which a feature peak is located in a spectrum of a sound to be recognized; And a sound recognizing unit that recognizes or rejects the input sound as the registered sound by using an average energy of a corresponding band corresponding to the peak band of the registered sound in the spectrum of the input sound and a non-corresponding band except the corresponding band. It is characterized in that to improve the recognition rate for a single sound and multiple sounds in a noisy environment, as well as to provide an advantage of reducing the amount of calculation.

Description

System and method for sound recognition using spectral peak, and method for measuring sound spectral similarity used sir}

The present invention relates to a sound recognition technology, and more particularly, by using spectral peaks to improve the recognition rate for a single sound and multiple sounds in a noise environment, as well as to reduce the amount of computation can be easily applied to resource-constrained environments A sound recognition system and method and methods for measuring acoustic spectral similarity used therein.

Currently, the exact population of the deaf is unknown, but it is estimated that nearly 500 million people worldwide suffer from hearing impairment. Hearing impaired people have a lot of inconvenience in perceiving the impaired hearing organs due to congenital or acquired reasons. For example, hearing impaired people are very difficult to recognize the sound of the doorbell, telephone ringing, boiling water, fire alarm, car horn, etc. You may not be able to respond appropriately, and you may face dangerous situations.

In order to help the deaf people's daily lives, trained dogs may be used, but training and training dogs require a lot of money and time, and in reality, they do not meet the needs of the deaf. Therefore, there is an urgent need for research and development of acoustic recognition systems for the hearing impaired.

Meanwhile, sound recognition technology has been applied to various fields such as home appliances, mobile phones, security and authentication, and its application field and demand are expected to increase more rapidly.

In such acoustic recognition technology, it is very important to be able to maintain high recognition rate or accuracy in a laboratory environment during research and development as well as in a real life environment in which various noises exist.

However, the conventional acoustic recognition systems have a typical problem of showing a high recognition rate in a limited laboratory environment, but a sharp decrease in the recognition rate in a real living environment such as a home or an office.

There are two main factors that cause the degradation of the conventional acoustic recognition system.

One factor is signal corruption due to background noise and room reverberation. This signal impairment makes it difficult to match between a clean sound signal used in learning the acoustic recognition system and a sound signal including noise input in a real environment even in the same sound.

However, existing studies to solve the signal corruption problem can only reduce the effect of the signal damage to some extent only under certain conditions, and provide consistent and reliable performance when applied to real-life devices. Couldn't show The reason is that background noise that can occur in a real life environment is so diverse that not only all background noise can be predicted, but also that the characteristics of the background noise change with time. Thus, as SF Boll introduced in "Suppression of acoustic noise in speech using spectral subtraction" (IEEE Transactions on Acoustics, Speech, and Signal Processing, ASSP-27 (2), pp. 113-120, 1979.). Solving the signal impairment problem by applying algorithms for estimating background noise is limited.

Another factor is an open-set problem that allows the acoustic recognition system to identify unregistered unregistered acoustic signals and not recognize unregistered sounds as already learned registered sounds.

Among the existing techniques, voice activity detection (VAD) based methods in speech recognition, in particular, attempt to distinguish between speech and non-voice noises using the inherent features of speech signals. On the other hand, CM (Confidence Measure) based methods are to determine the reliability of the recognition result to reject the unregistered keywords. The recognition rejection processing of the unregistered sound may be performed before or after the recognition step.

However, existing studies to solve the open-set problem have not been able to achieve reliable performance in a real life environment. The reason is that in the real life environment, these kinds of unregistered sounds are very diverse, and it is virtually impossible to know the characteristics of these unregistered sounds in advance.

In short, the existing technology does not provide a direct solution that can provide stable recognition rate by solving signal corruption and open-set problems in a real life environment in which various noises occur.

In addition, the existing technology cannot reduce the amount of information processing and computation during the acoustic recognition process, and thus cannot provide a sound recognition technology that can be easily applied to resource-constrained environments such as the mobile communication system and the USN (Ubiquitous Sensor Network). There is a problem.

Accordingly, the first technical problem to be solved by the present invention is to provide a sound recognition system that can be easily applied to resource-constrained environments by reducing the amount of computation as well as improving the recognition rate for single and multiple sounds in a noisy environment. It is.

The second technical problem to be solved by the present invention is to provide a sound recognition method that can be easily applied to resource-constrained environments by reducing the amount of computation as well as improving the recognition rate for single and multiple sounds in a noisy environment. .

The third technical problem to be solved by the present invention is to provide a method for measuring the similarity of the acoustic spectrum used in the acoustic recognition system and method.

In order to solve the first technical problem as described above, the present invention includes: a sound register for acquiring a peak band in which a feature peak is located in a spectrum of a sound to be recognized; And a sound recognizing unit that recognizes or rejects the input sound as the registered sound by using an average energy of a corresponding band corresponding to the peak band of the registered sound in the spectrum of the input sound and a non-corresponding band except the corresponding band. A sound recognition system using spectral peaks is provided.

The sound register may include: a feature peak extracting unit extracting a peak having an energy greater than a predetermined threshold among spectral peaks of the sound to be recognized as a feature peak; And a feature information generator for generating feature information indicating a peak band in which the extracted feature peak is located in the spectrum of the sound to be recognized.

The feature information generation unit may generate the feature information by setting a band except the overlap band as the peak band of the recognition target sound when there is an overlap band between bands in which feature peaks of the recognition target sounds are located. do.

The sound recognition unit may include: an average energy calculator configured to calculate an average energy of a corresponding band corresponding to the peak band of the registered sound and a non-compliant band except the corresponding band in the spectrum of the input sound; And a similarity measurer that measures the spectral similarity between the registered sound and the input sound using an average energy ratio between the corresponding band and the non-compliant band.

In example embodiments, the similarity measurer measures the spectral similarity using a peak domination ratio (PDR) value representing an average energy ratio of the corresponding band to the non-compliant band.

In one embodiment, the similarity measuring unit, using the normalized PDR (NPDR) value that normalized the PDR value by the average energy ratio of the peak band to the valley band other than the peak band in the spectrum of the registered sound The spectral similarity is measured.

The sound recognition unit recognizes the input sound as the registered sound when the PDR value is greater than a predetermined threshold.

In an embodiment, the sound recognizing unit calculates the PDR value for the registered sounds by the sound registration unit and sets the input sound when the maximum PDR value is greater than the threshold value among the calculated PDR values. Recognizes as a registered sound corresponding to the PDR value.

In order to solve the second technical problem as described above, the present invention, in the method of recognizing the input sound input to the system using the peak (peak) of the frequency spectrum in the sound recognition system (sound recognition system), Registering the sound to be recognized by acquiring a peak band in which a feature peak is located in a spectrum of sound; And a recognition step of recognizing or rejecting the input sound as the registered sound using the average energy of the corresponding band corresponding to the peak band of the registered sound and the non-corresponding band except the corresponding band in the spectrum of the input sound. A sound recognition method using spectral peaks is provided.

In order to solve the third technical problem as described above, the present invention provides a method for measuring the similarity (similarity) of the frequency spectrum between the pre-registered registration sound in the sound recognition system and the input sound input to the system. Calculating an average energy of a corresponding band corresponding to the peak band in which the characteristic peak of the registered sound of the registered sound is located and the non-corresponding band excluding the corresponding band from the input sound spectrum; And measuring spectral similarity between the registered sound and the input sound using an average energy ratio between the corresponding band and the non-compliant band.

 The present invention performs matching between two sounds by using spectral peak information of a pre-learned registered sound signal without extracting a feature of an input sound signal damaged by noise, thereby reducing background noise in a real life environment in which various noises occur. It provides the advantage of providing a stable recognition rate regardless.

In addition, it provides the advantage of improving the recognition rate for a single sound as well as multiple sounds.

Furthermore, by reducing the amount of information processing and computation in the sound recognition process, it provides an advantage that it can be easily applied to resource-constrained environments such as mobile terminal devices and USN sensor nodes that are rapidly developing today.

Prior to the description of the specific contents of the present invention, for the sake of understanding, an outline of the technical problem to be solved by the present invention is first presented.

1 shows a basic principle of a sound recognition method using spectral peaks according to the present invention.

Referring to FIG. 1, the present invention extracts features from an input acoustic signal in which noise distortion occurs in a sound recognition process in order to solve signal corruption and open-set problems of the prior art. Without the process, the spectral feature 114 is extracted from the pre-learned noise-free registration sound signal 112 through the learning step 110 and the registration using the extracted spectral feature 114 in the recognition step 120. Matching is performed between the acoustic signal 112 and the input acoustic signal 122 where the noise distortion has occurred. The reason is that signal corruption and open-set problems of the prior art are caused by performing certain matching using features 124 extracted from the input acoustic signal 122 distorted by noise.

Accordingly, the present invention extracts only the spectral features of the registered sound to perform matching between the registered sound signal and the input sound signal, and introduces a new similarity measure (Peak Domination Ratio) for this purpose.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings in order to clarify the solutions of the technical problems of the present invention. However, in describing the present invention, when it is determined that the detailed description of the related known technology or configuration may make the gist of the present invention unclear, the detailed description thereof will be omitted. In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to intention or custom of a user, an operator, or the like. Therefore, the definition should be based on the contents throughout this specification.

2 shows an example of a sound recognition system using spectral peaks in accordance with the present invention.

3 is a flowchart illustrating an example of a sound recognition method using spectral peaks according to the present invention.

2 and 3, first, a sound recognition system 200 according to the present invention includes a sound register 210 and a sound recognizer 220, a sound receiver 230, and spectrum analysis. The unit 240 may further include a storage unit 250 and an output unit 260.

The sound register 210 acquires a peak band at which a characteristic peak of a recognition sound spectrum is located in a registration step of recognizing a target sound by learning. The target sound is registered (S310 to S330).

In more detail, the sound register 210 may include a feature peak extractor 212 and a feature information generator 214 to perform a registration step of the recognition target sound.

The feature peak extractor 212 extracts the feature peak from the spectrum of the sound to be recognized (S310). In this case, the feature peak may be extracted by applying various types of extraction methods, and in one embodiment, the feature peak extractor 212 may be configured to predetermine the spectral peaks of the recognition sound to simplify the calculation. Peaks with energy greater than the threshold can be extracted as feature peaks.

In general, the sounds that humans need to recognize in everyday life each have a unique frequency spectrum form, and there are characteristic peaks in these spectra. In particular, mechanical sounds, such as doorbell sounds, boiling kettle sounds, telephone rings, etc., have distinctive feature peaks on the spectrum.

4 shows frequency spectra of mechanical sound.

Referring to FIG. 4, it can be seen that there are unique feature peaks in each sound spectrum. For example, when the threshold is set as shown in FIG. 4, four doorbell sounds 410, two water boiling kettle sounds 420, seven phone ringing sounds 430, and seven features. There will be a peak. The feature peak extractor 212 extracts these feature peaks. Of course, as described above, not only simple energy-based peak extraction schemes, but also more complex algorithms may be used.

The feature information generator 214 generates feature information indicating that the extracted feature peak indicates a peak band in which the feature peak is located in the spectrum of the sound to be recognized (S320).

5 shows a process of generating the feature information.

Referring to FIG. 5, when the feature peak extractor 212 extracts four feature peaks from, for example, a doorbell acoustic spectrum 510, the feature information generator 214 may include four peaks at which the four feature peaks are located. Generate feature information 520 representing the bands.

In one embodiment, the feature information 520 may be generated in the form of a peak signature vector. That is, when the feature peak extractor 212 extracts a feature peak equal to or greater than the threshold value from the N-dimensional average spectrum obtained from the training data about the acoustic signal to be recognized, the feature information generator 214. Is assigned to each dimension corresponding to the peak band where the feature peak is located, and 0 is assigned to each dimension corresponding to a valley band, which is a band other than the peak band, to obtain an N-dimensional peak feature vector. Can be generated.

In one embodiment, the feature information generation unit 214 is characterized in that the band except for the overlapping band as the peak band of the recognition target sound when there is an overlapping band between bands in which feature peaks of the recognition target sounds are located. Information can be generated.

6A and 6B illustrate a process of generating the feature information when overlapping bands exist between peak bands of sound to be recognized.

As shown in FIG. 6A, when an overlapping band exists between peak bands of sound to be recognized, an overlapping portion exists between respective feature information (peak feature vectors) corresponding thereto. That is, the B band portion overlaps between the first feature information 610 and the third feature information 630, and the A band portion overlaps between the second feature information 620 and the third feature information 630.

In this case, as shown in FIG. 6B, the feature information generation unit 214 generates only the bands excluding the overlap band portions A and B as peak bands of the sound to be recognized and generate the feature information.

As will be described again below, by generating the feature information as described above, the recognition rate of each sound can be improved, as well as the recognition rate of multiple sounds that occur simultaneously.

Next, the feature information generation unit 214 registers the recognition target sound as a registration sound by storing the feature information in the storage unit 250 (S330). The storage unit 250 may be configured to be integrated with the sound register unit 210 or may be configured outside the sound register unit 210.

In addition, the registration step (S310 to S330) by the sound registration unit 210 may be performed at the time of production or production of the acoustic recognition system according to the intention of the producer or manufacturer, but customizing according to the use of the user or operator This can be configured to be possible. That is, the user or the operator may register the recognition target sound to be recognized through the sound recognition system in a noise-free environment before the recognition of the input sound. In this case, when the sound receiver 230 receives a recognition target sound signal of a waveform and transmits it to the spectrum analyzer 240, the spectrum analyzer 240 performs the frequency analysis to recognize the sound. The signal may be converted into a spectrum vector and transmitted to the sound register 210.

Next, the sound recognizing unit 220, in the step of recognizing the input sound, a relevant band corresponding to the peak band of the registered sound on the spectrum of the input sound and the spectrum of the input sound The input sound is recognized as the registered sound or rejected using the average energy of an irrelevant band excluding the corresponding band (S340 to S380).

In more detail, the sound recognizer 220 may include an average energy calculator 222 and a similarity measurer 224 to perform the recognition of the input sound.

First, when the sound receiver 230 receives an input sound signal of a waveform and transmits it to the spectrum analyzer 240, the spectrum analyzer 240 converts the input sound signal into a spectrum vector through a frequency analysis process. To change (S340). The spectrum analyzer 240 may perform the frequency analysis by using a Discrete Fourier Transform (DFT).

The average energy calculator 222 calculates average energy per unit band of the corresponding band and the non-compliant band (S350).

Next, the similarity measurer 224 measures the spectral similarity between the registered sound and the input sound using an average energy ratio between the corresponding band and the non-compliant band (S360).

Recognized sounds according to the invention, in particular mechanical sounds such as doorbell sounds, have characteristic spectral peaks. The reason is that the sharp peaks in the spectrum produce unique mechanical sounds, such as beep sounds. Accordingly, the band of each acoustic spectrum may be divided into a peak band in which the feature peak is located and a valley band which is a band other than the peak band.

Of particular note is that the valley band is considerably wider than the peak band. That is, the peak band corresponds to a band in which acoustic energy is concentrated in a narrow band, whereas the valley band corresponds to a band in which acoustic energy is scattered in a very wide band. As a result, the average energy of the peak band represents a relatively large value, and the average energy of the valley band represents a relatively small value.

Similarly, if the input sound corresponds to the same sound as the registered sound, the average energy of the corresponding band in the input sound spectrum shows a relatively large value regardless of the presence of noise, and the average energy of the non-corresponding band is small. Will display the value.

Accordingly, the present invention intends to measure the similarity between the registered sound and the input sound using the average energy of the corresponding band and the non-compliant band.

In one embodiment, the similarity measurer 224 may measure the spectral similarity using a peak domination ratio (PDR) value representing an average energy ratio of the corresponding band to the non-compliant band.

The PDR makes it possible to measure the similarity using only the peak feature vector (S) of the registered sound and the spectrum (X) of the input sound.

The PDR value may be represented as in Equation 1.

Equation 1 may be specifically calculated as Equation 2.

In Equation 2, N represents the dimensions of the S and X, X [k] represents the k-dimensional energy of the input acoustic spectrum (X), S [k] is k of the peak feature vector (S) Represents a dimensional binary value. That is, S [k] has a binary value 1 for the peak band of the registered sound and a binary value 0 for the valley area.

If the PDR value is equal to or greater than a predetermined threshold value (S362), the sound recognition unit 220 recognizes the input sound as the registered sound (S370). However, when the PDR value is smaller than the threshold (S362), the input sound is judged as noise or unregistered sound and rejected recognition (S372).

7A and 7B show the spectrum of the input sound with noise distortion.

7A shows the spectrum of a buzzer sound damaged by background noise.

As shown in FIG. 7A, some peaks in the acoustic spectrum are buried due to background distortions such as channel distortions, such as reverberation, as well as still on the input acoustic spectrum despite background noise. Even when there are characteristic spectral peaks of the buzzer sound, the existing technology causes misrecognition because it extracts a feature from the input sound spectrum to determine similarity with a registered sound. The reason is that the similarity of peaks existing in a narrow band in the existing technology may be meaningless depending on the similarity of valleys existing in a relatively very wide band. In particular, a method of determining similarity with the same weight over the entire spectral band, such as an Euclidean distance based method, tends to show a result distorted by these background noises.

FIG. 7B shows the spectrum of the doorbell sound damaged by background noise and short burst noise.

As shown in FIG. 7B, unrelated peaks may appear in the spectrum of the input sound due to short burst noises such as door closing sounds. One of the major problems with spectral peak based methods is that misrecognition can easily occur due to the addition of these unrelated peaks.

As a result, in the existing technology, the change in the relative magnitude of the spectral peaks and the addition of unrelated peaks due to the background noise distorts the characteristics of the input acoustic signal, resulting in a sudden decrease in performance.

8A to 8C show the histograms of the relative average energy of the corresponding and non-corresponding bands of the input acoustic spectrum.

As shown in Figs. 8A and 8B, if the input sound is the same sound as the registered sound, not only when there is background noise in the input sound but also when there are additional peaks, the corresponding band in the spectrum of the input sound. It can be seen that the average energy (black bar) of has a relatively large value compared to the average energy (white bar) of the non-corresponding band. However, as shown in FIG. 8C, if the input sound is only noise, it can be seen that the average energy difference between the corresponding band and the non-compliant band is not large in the spectrum of the input sound.

Accordingly, the present invention eliminates the feature extraction process from the input acoustic signal that causes misrecognition through the energy-based similarity measuring method, that is, the PDR-based similarity measuring method, and measures the similarity between the registered sound and the input sound. This allows you to focus on specific related bands. As a result, the present invention can prevent not only the case where the noise or the unregistered sound is mistaken as the registered sound, but also the case where the registered sound is mistaken as the other registered sound. This result is because the PDR value appears relatively large only when the characteristic peaks of the registered sound appear on the spectrum of the input sound.

In one embodiment, when there are a plurality of registration sounds by the sound registration unit 210, the sound recognition unit 220 calculates the PDR value for each of the registration sounds, the maximum of the calculated PDR value When the PDR value is larger than the threshold value, the input sound may be recognized as a registered sound corresponding to the maximum PDR value.

In addition, the sound recognition unit 220 normalizes the PDR value to simplify the calculation and efficiency by performing a recognition process using a single threshold value for sounds representing various levels of PDR values. Can be used.

In one embodiment, the similarity measuring unit 224 of the sound recognition unit 220, the PDR value by the average energy ratio (self PDR) of the peak band to the valley band in the spectrum of the registered sound. The spectral similarity may be measured using a normalized PDR (NPDR) value normalized to. That is, NPDR value for the similarity measurement between the i-th registered sound spectrum (R _i) and the input sound spectrum (X) may be invoked acid as shown in Equation 3.

In Equation 3, S _i is a feature vector peak (peak signature vector) of the registered spectrum (R _i). In one embodiment, in the registration step by the sound registration unit 210, the self PDR value for each registered sound is generated and stored in the storage unit 250, the recognition by the sound recognition unit 220 In the step, the NPDR value may be calculated using the feature information and the self PDR value stored in the storage unit 250.

On the other hand, in the acoustic spectral similarity measurement of the present invention, the characteristic that unrelated peaks of the input acoustic spectrum not related to the registered sound have less influence than the related peak (peak corresponding to the characteristic peak of the registered sound) occurs simultaneously. Also, the recognition rate is improved in recognition of multiple sounds.

9 shows the frequency spectrum of multiple sounds.

As shown in Fig. 9, the multiple sound is composed of a simultaneous doorbell sound and a telephone ringing sound. For example, when calculating the NPDR for the doorbell sound among the registered sounds, the feature peaks of the doorbell sound have more influence on the NPDR value than the feature peaks of the ringtone sound. That is, the multi-acoustic signal shows a high NPDR value for the doorbell sound despite the ringing tone. Of course, this principle applies equally to the calculation of NPDR for ringing tones. In one embodiment, the sound recognition unit 220 may recognize that all of the registered sounds are present when the multiple sound represents an NPDR value of a predetermined value or more for a specific registration sound. Therefore, the acoustic recognition system according to the present invention can recognize all the respective acoustic signals in the multiple acoustic signals.

In such multiple acoustic recognition, a misrecognition problem may occur when each component acoustic signal has spectral peaks present in the same band. For example, when the doorbell sound and the ringtone sound have feature peaks in the same band, the NPDR value for the doorbell sound may appear high even though the ringtone sound is recognized. Similarly, the NPDR value for the ringing tone may be high even when the doorbell sound is recognized.

Accordingly, in order to solve this problem, the present invention provides a common peak band in the peak band when generating the feature information (eg, peak feature vector) related to the sound to be recognized as described above with reference to FIGS. 6A and 6B. Exclude. As a result, the characteristic information on the registered sound can each exhibit a unique peak band, and it is possible to eliminate the influence of the common peak band on the NPDR value.

Then, when there is a subsequent input sound input to the sound recognition system, the above-described sound recognition process is repeated (S380).

When the input sound is recognized as the registered sound, the output unit 260 outputs a recognition result in various forms. In one embodiment, the output unit 260 may include a variety of display devices, sound devices, haptic interface, and the like.

The present invention can also be embodied as computer readable program code on a computer readable recording medium. When the present invention is executed through software, the constituent means of the present invention are code segments for performing necessary tasks. The program or code segments may be stored on a processor readable medium or transmitted by a computer data signal coupled with a carrier on a transmission medium or network.

The computer-readable recording medium includes all kinds of recording devices for storing data that can be read by a computer system. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disks, optical data storage devices, and the like. The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

Hereinafter, the significant effects of the present invention are verified.

Registration and recognition were performed using a doorbell, a boiling kettle, and a telephone ring sound as test data for performance evaluation of the present invention. Each sound was sampled at 16 kHz, and 128 ms blocks were each processed with a 64 ms overlap. Fourier analysis was applied through a Hamming window to obtain 1024 spectral vectors for each block. In addition, all vectors for each sound were averaged to obtain an average spectrum.

In addition, the Euclidean distance based method for comparing the entire spectral band was selected as a performance comparison target of the present invention. Since the energy levels of the training data and the experimental data may be different, vector normalization was applied to each of the spectral vectors before calculating the Euclidean distance.

10 is a graph showing the recognition performance test results of the present invention according to the SNR of the short-range test data.

To demonstrate the recognition performance of the short range test data of the present invention, the doorbell, kettle, and ringtones are recorded 10 times in a noise-free environment, and the test data is generated by mixing the background noises having an SNR level of 10 dB to -20 dB. It was. The background noises include white noise, pink noise, and eight noises at Aurora-2 corpus, ie airports, exhibitions, restaurants, streets, underground passages, cars, trains. And multiple crosstalk noise. That is, test data of 2400 pieces (2400 = sound type (3) x recording number (10) x SNR level (8) x noise type (10)) was used.

As shown in FIG. 10, the present invention (NPDR), unlike the Euclidean distance-based method (Normalized Euclidean), can be seen that satisfactory performance in a noisy signal as well as a noisy signal. This is because the similarity measuring method of the present invention weights the related peak band as compared to the wide valley band, thereby minimizing the effect of background noise spread over the entire spectral band. The Euclidean distance based scheme significantly reduces the accuracy to 33.3% when the SNR level is -15 dB or less, which is no different from random selection.

In Fig. 11A, the Euclidean distance of far-field test data for registered and unregistered sounds is shown as a histogram.

In FIG. 11B, NPDR values of far-field test data relating to registered and unregistered sounds are shown as histograms.

In order to prove the recognition performance of the present invention for remote test data, 10 recordings of the three types of sound sources were performed in 90 configurations in which the indoor structure (for example, home, office, etc.), the sound source, and the microphone positions are different. Was performed. That is, 2700 test data (2700 = sound type (3) x recording frequency (10) x composition type (90)) were used.

As shown in FIG. 11A, in the case of the Euclidean distance-based scheme, it can be seen that the Euclidean distance (black bar) of the registered sound overlaps a lot with the Euclidean distance (white bar) of the unregistered sound. This result indicates that the Euclidean distance cannot be used as an effective reference value for rejecting recognition of unregistered sounds.

On the other hand, as shown in Figure 11b, in the present invention, the NPDR value of the registered sound is clearly distinguished from the NPDR value of the unregistered sound. These results clearly indicate that the NPDR value of the present invention is an effective reference value for determining whether a given sound corresponds to a registered sound.

As described above, since the Euclidean distance overlaps the output of the registered sound with the output of the unregistered sound at a high level, the Equal Error Rate (EER) of the Euclidean distance method is 63.5%, which is quite high. In contrast, the EER of the present invention is only 0.2%.

Figure 12 shows the results of the recognition performance test of the present invention for the single acoustic and multi-acoustic data.

In order to prove the recognition performance of the present invention with respect to multiple sounds simultaneously occurring, four kinds of mixed signals (buzzer + kettle, doorbell + telephone, kettle + telephone, doorbell + kettle + telephone) in 60 indoor environments are provided. Ten recordings were performed. That is, 2400 pieces of test data (2400 = mixed signal 4 x recording number 10 x configuration type 60) were used.

As shown in FIG. 12, it can be seen that the recognition result OUTPUT of the present invention for each used input sound INPUT shows a very high accuracy. That is, the accuracy of single acoustic recognition is 97.5%, the accuracy of multiple acoustic recognition is 92.8%, and the accuracy of rejection of unregistered acoustic recognition is 98.9%.

It is particularly noteworthy that the present invention exhibits very high accuracy even in multi-acoustic recognition, and that at least some of the constituent sounds can be recognized with high accuracy even when not all the sounds constituting the multi-sound are recognized. .

As described above, the present invention provides a real life environment in which various noises are generated by performing matching between two sounds using spectral peak information of a pre-learned registered sound signal without a feature extraction process for an input sound signal damaged by noise. Provides the advantage of providing a stable recognition rate independent of background noise. In addition, it provides the advantage of improving the recognition rate for a single sound as well as multiple sounds. Furthermore, it reduces the amount of information processing and computation in the acoustic recognition process, and provides the advantage that it can be easily applied to resource-constrained environments such as mobile terminal devices and USN sensor nodes that are rapidly developing today.

So far, the present invention has been described with reference to the embodiments. However, one of ordinary skill in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential technical spirit of the present invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. That is, the true technical scope of the present invention is shown in the appended claims, and all differences within the equivalent scope will be construed as being included in the present invention.

1 is a view showing the basic principle of the acoustic recognition method using the spectral peak according to the present invention.

2 is a block diagram showing an example of an acoustic recognition system using spectral peaks according to the present invention.

3 is a flowchart illustrating an example of a sound recognition method using spectral peaks according to the present invention.

4 shows frequency spectra of mechanical sound.

5 is a diagram illustrating a process of generating feature information of the present invention.

6A and 6B illustrate a process of generating feature information of the present invention when there is an overlapping band between peak bands of sound to be recognized.

7A and 7B are diagrams illustrating a frequency spectrum of input sound in which noise distortion occurs.

8A to 8C are histograms showing relative average energies of corresponding and non-corresponding bands of the input acoustic spectrum.

9 is a diagram illustrating a frequency spectrum of multiple sounds.

10 is a graph showing the recognition performance test results of the present invention according to the SNR of the short-range test data.

FIG. 11A is a diagram showing a histogram of Euclidean distance of far-field test data regarding a registered sound and an unregistered sound. FIG.

FIG. 11B is a histogram of NPDR values of remote test data relating to registered and unregistered sounds. FIG.

12 is a view showing the results of the recognition performance test of the present invention for the single acoustic and multi-acoustic data.

Claims (20)

A sound register for acquiring a peak band at which a feature peak is located in a spectrum of a recognition sound and registering the recognition sound; And A sound recognition unit that recognizes or rejects the input sound as the registered sound using an average energy of a corresponding band corresponding to the peak band of the registered sound in the spectrum of the input sound and a non-corresponding band except the corresponding band; , And the sound register recognizes each of the multiple sounds simultaneously generated by excluding overlapping bands of feature peaks of the sound to be recognized. The method of claim 1, The sound registration unit, A feature peak extracting unit extracting a peak having an energy greater than a predetermined threshold among the spectral peaks of the recognition sound as a feature peak; And And a feature information generator configured to generate feature information indicating a peak band in which the extracted feature peak is located in the spectrum of the sound to be recognized. The method of claim 2, The feature information generator, when there is an overlap band between bands in which feature peaks of recognition sounds are located, generates the feature information by setting a band except the overlap band as the peak band of the recognition sound. Acoustic Recognition System Using Peaks. The method of claim 1, The sound recognition unit, An average energy calculator configured to calculate an average energy of a corresponding band corresponding to the peak band of the registered sound and a non-compliant band except the corresponding band in the spectrum of the input sound; And And a similarity measurer for measuring the spectral similarity between the registered sound and the input sound using an average energy ratio between the corresponding band and the non-compliant band. The method of claim 4, wherein And the similarity measurer measures the spectral similarity using a peak domination ratio (PDR) value representing an average energy ratio of the corresponding band to the non-compliant band. The method of claim 5, The similarity measuring unit measures the spectral similarity using a normalized PDR (NPDR) value in which the PDR value is normalized by an average energy ratio of the peak band to a valley band other than the peak band in the spectrum of the registered sound. A sound recognition system using spectral peaks, characterized in that. The method of claim 5, And the sound recognition unit recognizes the input sound as the registered sound when the PDR value is larger than a predetermined threshold value. The method of claim 7, wherein The sound recognition unit calculates the PDR value for the registered sounds by the sound registration unit, and registers the input sound corresponding to the maximum PDR value when the maximum PDR value among the calculated PDR values is greater than the threshold value. An acoustic recognition system using spectral peaks, characterized in that the recognition by sound. In a method of recognizing an input sound input to the system using a peak of the sound spectrum in a sound recognition system, Registering the sound to be acquired by acquiring a peak band in which a feature peak is located in a spectrum of the sound to be recognized; And A recognition step of recognizing or rejecting the input sound as the registered sound using an average energy of a corresponding band corresponding to the peak band of the registered sound in the spectrum of the input sound and a non-corresponding band except the corresponding band; , And the registering step recognizes each of the multiple sounds simultaneously generated by excluding overlapping bands of the feature peaks of the recognition target sound. 10. The method of claim 9, The registration step, Extracting, as a characteristic peak, a peak having an energy larger than a predetermined threshold among spectral peaks of the recognition sound; And And generating feature information indicating a peak band in which the extracted feature peaks are located in the spectrum of the recognition target sound. The method of claim 10, The generating of the characteristic information may include generating the characteristic information by using a band except the overlapping band as the peak band of the recognition target sound when there is an overlapping band between bands in which feature peaks of the recognition target sounds are located. A sound recognition method using spectral peaks. 10. The method of claim 9, The recognition step, Calculating an average energy of a corresponding band corresponding to the peak band of the registered sound and a non-corresponding band excluding the corresponding band from the spectrum of the input sound; And And measuring spectral similarity between the registered sound and the input sound using an average energy ratio between the corresponding band and the non-compliant band. The method of claim 12, The measuring similarity may include measuring the spectral similarity using a peak domination ratio (PDR) value representing an average energy ratio of the corresponding band to the non-compliant band. . The method of claim 13, In the similarity measuring step, the spectral similarity is measured using a normalized PDR (NPDR) value in which the PDR value is normalized by an average energy ratio of the peak band to a valley band other than the peak band in the spectrum of the registered sound. Sound recognition method using a spectral peak, characterized in that the step of. The method of claim 13, And the recognizing step is recognizing the input sound as the registered sound when the PDR value is larger than a predetermined threshold. The method of claim 15, In the recognizing step, the PDR value is calculated for the registered sounds according to the registration step, and when the maximum PDR value among the calculated PDR values is greater than the threshold value, the input sound corresponds to the maximum PDR value. A sound recognition method using spectral peaks, characterized in that the step of recognizing the sound. A method for measuring the similarity (similarity) of a frequency spectrum between a registered sound registered in advance in a sound recognition system and an input sound input to the system, Calculating an average energy of a corresponding band corresponding to the peak band in which the characteristic peak of the registered sound is located in the spectrum of the input sound and a non-corresponding band excluding the corresponding band; And Measuring a spectral similarity between the registered sound and the input sound using an average energy ratio between the corresponding band and the non-compliant band, The spectrum of the registered sound is a spectrum spectral similarity measurement method, characterized in that the spectrum except the overlapping band between the characteristic peaks of the multiple sounds occurring at the same time. The method of claim 17, The similarity measuring step is measuring the spectral similarity by using a peak domination ratio (PDR) value representing the average energy ratio of the corresponding band to the non-corresponding band. The method of claim 18, In the similarity measuring step, the spectral similarity is measured using a normalized PDR (NPDR) value in which the PDR value is normalized by an average energy ratio of the peak band to a valley band other than the peak band in the spectrum of the registered sound. A method of measuring acoustic spectral similarity, characterized in that the step of. A computer-readable recording medium having recorded thereon a program for executing a method according to any one of claims 9 to 19 with a computer.

Images