KR101022519B1 - System and method for voice activity detection using vowel characteristic, and method for measuring sound spectral similarity used thereto - Google Patents

Abstract

The present invention relates to a speech section detection technology for speech recognition. The speech section detection system using the vowel feature according to the present invention includes a vowel feature storage unit for storing vowel feature information indicating a peak band at which a feature peak is located in a spectrum of a vowel. ; And determining whether the input sound corresponds to voice by using an average energy of a corresponding band corresponding to the peak band indicated by the stored vowel feature information in the spectrum of the input sound and a non-corresponding band except the corresponding band. Characterized in that it includes a speech section detection unit, to improve the speech section detection performance under various noise environments and SNR, as well as to reduce the amount of computation provides an advantage of improving the energy efficiency of the speech recognition system.

Description

System and method for voice activity detection using vowel characteristic, and method for measuring sound spectral similarity used

The present invention relates to a speech segment detection technique for speech recognition, and more particularly, to improve speech segment detection performance under various noise environments and SNRs by using characteristic spectral peaks of vowels, and to reduce the amount of computation. The present invention relates to a speech segment detection system and method for improving the energy efficiency of the apparatus and a method for measuring acoustic spectral similarity used therein.

Voice recognition is to identify linguistic meaning and content from speech spoken by humans through an automatic speech recognition system. Specifically, voice recognition is performed by inputting a speech waveform to identify a word or word string and extract meaning. Refers to the process. Such voice recognition technology is applied to various fields such as home electronics, mobile phones, security and authentication, and its application field and demand are expected to increase more rapidly.

Meanwhile, a voice recognition system that performs voice recognition can be broadly classified into a word speech system, a continuous speech recognition system, and a speaker recognition system. Speech recognition algorithms applied to the speech recognition systems generally include a speech section detection process, a feature extraction process, and a matching process.

Among the speech recognition processes, the voice activity detection (VAD) process is a technology for determining whether a specific sound generated corresponds to a human voice, and is mainly applied as a preprocessing step of speech recognition. In such speech section detection 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.

Existing techniques related to speech segmentation include machine learning methods that detect a speech segment after training the model using learning data about speech and sound other than speech, and features such as zero crossing rate, A method of detecting vocal intervals by modeling spectral entropy and searching for the appearance of a corresponding feature has been introduced.

However, in the conventional machine learning method, it is impossible for the model to learn all sounds other than the voice, and thus, there is a problem in that the speech section detection performance of the untrained sound is greatly reduced.

In addition, in the conventional method of modeling speech features, many noises other than speech also have similar characteristics to speech, and thus there is a problem of misrecognizing noise other than speech as speech.

In short, existing technologies have a fundamental problem of impairing speech segment detection performance in that they use features of the entire speech. The reason is that human voice is a very complex sound signal generated through the interaction of various articulation organs such as lungs, lips, tongue, and nasal cavity, so it is very difficult to find a single feature that covers the entire voice. This is because no clear feature of distinguishing sound signals other than speech has been revealed.

On the other hand, the conventional voice interval detection method shows a high recognition rate in a limited laboratory environment, while the other major factors that show a sharp drop in the recognition rate in a real life environment, such as home or office, there are two major.

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

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 causes the speech recognition system to identify unacknowledged acoustic signals and not to recognize acoustic signals other than speech signals as already learned speech signals.

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, as described above, it is virtually impossible to learn all the features of the vast sounds in advance in a real life environment.

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 calculation when detecting the voice section, and thus cannot provide a voice 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 not only to improve the speech segment detection performance under various noise environments and SNR by using the characteristic spectral peak of the vowel, but also to reduce the amount of computation to improve the energy efficiency of the speech recognition system. It is to provide a voice section detection system.

The second technical problem to be solved by the present invention is to improve the speech segmentation performance under various noise environments and SNRs by using characteristic spectral peaks of vowels, as well as to reduce the amount of computation and improve the energy efficiency of speech recognition systems. An interval detection method is provided.

The third technical problem to be solved by the present invention is to provide a method for measuring the acoustic spectral similarity used in the speech section detection system and method.

In order to solve the first technical problem as described above, the present invention provides a vowel feature storage unit for storing vowel feature information indicating a peak band in which a feature peak is located in a vowel spectrum; And determining whether the input sound corresponds to voice by using an average energy of a corresponding band corresponding to the peak band indicated by the stored vowel feature information in the spectrum of the input sound and a non-corresponding band except the corresponding band. Provided is a speech section detection system using a vowel feature including a speech section detector.

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

In one embodiment, the feature information generation unit, by dividing the entire spectral band of the vowels into a predetermined number of unit bands, and represents a unit band corresponding to the peak band in the spectrum of the vowels as 1 and other than the peak band. The vowel feature information representing a unit band corresponding to a valley band, which is a band, as 0 is generated.

In one embodiment, the voice section detector, in the spectrum of the input sound, the average energy for calculating the average energy of the corresponding band corresponding to the peak band of the stored vowel feature information and the non-corresponding band except the corresponding band A calculator; And a similarity measurer that measures spectral similarity between the vowel and the input sound by using an average energy difference between the corresponding band and the non-compliant band.

In example embodiments, the similarity measurer measures the spectral similarity using a Peak-Valley Energy Difference (PVED) value representing a value obtained by subtracting the average energy of the non-corresponding band from the average energy of the corresponding band.

The voice section detector, when the PVED value is greater than a predetermined threshold, determines the input sound as voice and detects the voice section as a voice section.

The voice section detector may be configured to calculate the PVED value with respect to the stored vowel feature information, and determine the input sound as voice when the maximum PVED value among the calculated PVED values is greater than the threshold value. do.

The voice segment detector may further include a hangover processor configured to apply a hang-over algorithm to the input sound when the PVED value is not greater than the threshold value. The detector determines the input sound as a voice when the input sound is processed by the hangover processor.

In order to solve the second technical problem as described above, the present invention provides a method for detecting a voice section of input sounds input to the system in a voice activity detection system, wherein a feature peak in a spectrum of a vowel A vowel feature storing step of storing vowel feature information indicating a located peak band; And determining whether the input sound corresponds to voice by using an average energy of a corresponding band corresponding to the peak band indicated by the stored vowel feature information in the spectrum of the input sound and a non-corresponding band except the corresponding band. Provided is a speech segment detection method using a vowel feature including speech segment detection.

In order to solve the third technical problem as described above, the present invention, a method for measuring the similarity (similarity) of the frequency spectrum between the pre-learned vowels and the input sound input to the system in the Voice Activity Detection system (Voice Activity Detection system) Calculating an average energy of a corresponding band corresponding to a peak band in which the spectral feature peaks of the vowels are located and a non-corresponding band excluding the corresponding band from the input sound spectrum; And measuring spectral similarity between the vowel and the input sound using an average energy difference between the corresponding band and the non-compliant band.

The present invention provides the advantage of improving the speech segment detection performance under various noise environments and signal-to-noise ratios (SNRs) by using the characteristic spectral peaks of the learned vowels without the feature extraction process for the input acoustic signal damaged by noise. to provide.

In addition, by reducing the amount of information processing and calculation in the speech section detection process, it provides an advantage that it can be easily applied to resource-constrained environment such as mobile terminal device, USN sensor node that is rapidly developing today.

Furthermore, by applying it as a preprocessing step of speech recognition, the speech recognition system can be further improved in energy efficiency by allowing the speech recognition system to remove unnecessary motion in the non-speech section.

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 illustrates a basic principle of a speech segment detection method using a vowel feature according to the present invention.

Referring to FIG. 1, the present invention solves the problem that it is difficult to extract features related to the entire speech in the existing technology, and to solve signal corruption and open-set problems. Without extracting the feature from the distorted input acoustic signal, the spectral feature 114 is extracted and stored from the pre-learned noiseless vowel signal 112 through the learning step 110, and the detection step 120 The stored spectral feature 114 is used to perform matching between the vowel signal 112 and the input acoustic signal 122 where the noise distortion has occurred. This is because signal corruption and open-set problems of the prior art were caused by performing certain matching using features 124 extracted from input acoustic signals 122 distorted by noise.

Accordingly, the present invention extracts only the spectral features of a vowel to perform matching between the vowel signal and the input acoustic signal, and introduces a new similarity measure called Peak-Valley Energy Difference (PVED).

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 speech segment detection system using a vowel feature according to the present invention.

3 is a flowchart illustrating an example of a speech segment detection method using a vowel feature according to the present invention.

2 and 3, first, the speech section detection system 200 according to the present invention includes a vowel feature storage unit 210 and a speech section detection unit 220, and includes an audio receiver 230 and a spectrum analyzer. 240 may further include.

The vowel feature storage unit 210 stores vowel feature information indicating a peak band at which a characteristic peak is located in the frequency spectrum of the vowel through a learning step (S310 and S320).

In more detail, the vowel feature storage unit 210 may include a feature peak extractor 212 and a feature information generator 214 to perform a learning step on a vowel constituting a human voice. .

The feature peak extractor 212 extracts the feature peak from the spectrum of the vowel (S310). In this case, the feature peak may be extracted by applying various types of extraction methods. In an embodiment, the feature peak extractor 212 may extract a peak having an energy greater than a predetermined threshold among the spectral peaks of the vowel as a feature peak to simplify the calculation.

Human speech (voice or speech) is largely composed of consonants and vowels. In the case of consonants, the duration is very short and it is always characterized by vowels. On the other hand, vowels have a relatively long duration and each vowel has its own characteristic peaks known as formants in the spectrum.

4 shows the frequency spectrum of the 'a' vowel.

As shown in FIG. 4, it can be seen that each vowel has its own characteristic peak. The peak band where the characteristic peaks are present on the vowel spectrum has a significantly higher energy than the surrounding band so that it does not disappear easily due to noise distortion. Therefore, the vowel portion of the human voice is easy to detect even in a real life environment in which various noises occur, and if only the vowel portion can be detected, the consonant and The entire vowel can be detected. The feature peak extractor 212 extracts feature peaks of this vowel. 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 vowel feature information indicating a peak band in which the extracted feature peak is located in the spectrum of the vowel (S320).

5 illustrates a process of generating the vowel feature information.

Referring to FIG. 5, when the feature peak extractor 212 extracts four feature peaks from, for example, a particular collection of spectra 510, the feature information generator 214 may display four feature peaks. The vowel feature information 520 representing the peak bands is generated.

In one embodiment, the feature information generator 214 divides the entire spectral band of the vowel into a predetermined number of unit bands, and represents a unit band corresponding to the peak band in the vowel spectrum as 1 and The vowel feature information is generated by indicating a unit band corresponding to a valley band, which is a band other than the peak band, as 0. For example, the vowel feature information 520 may be generated in the form of a peak signature vector. That is, when the feature peak extractor 212 extracts feature peaks greater than or equal to the threshold value from an N-dimensional average spectrum obtained by applying a Discrete Fourier Transform (DFT) to training data about a vowel signal, the feature peak is extracted. The information generator 214 assigns 1 to each dimension corresponding to the peak band where the feature peak is located, and assigns 1 to each dimension corresponding to a valley band which is a band other than the peak band. By assigning zero, an N-dimensional peak signature vector may be generated.

As will be described again below, by generating and storing the vowel feature information as a binary value as described above, the amount of data stored in the vowel feature storage unit 210 can be reduced, as well as through the acoustic spectrum similarity measuring method according to the present invention. It is possible to reduce the amount of calculation and improve energy efficiency when detecting a voice segment.

As described above, the vowel feature storage unit 210 learns the vowel by storing the vowel feature information. In one embodiment, the learning steps (S310 to S320) by the vowel feature storage unit 210 may be performed at the time of production or production of the speech recognition system according to the intention of the producer or manufacturer. In this case, operations of the feature peak extractor 212 and the feature information generator 214 may be performed by an external device, and the vowel feature storage 210 may be configured to store only the vowel feature information. In addition, in one embodiment, it can be configured to be customized according to the user or operator of the speech recognition system. That is, before speech recognition, a user or operator may configure the speech section detection system 200 to individually learn the vowels in a noise-free environment. In this case, when the sound receiver 230 receives a voice signal of a waveform and delivers it to the spectrum analyzer 240, the spectrum analyzer 240 performs a vowel signal among the voice signals through frequency analysis. It can be configured to transfer to the vowel feature storage unit 210 by changing to a spectral vector.

Next, the voice section detection unit 220, in a voice activity detection (VAD) step for the input sound, a corresponding band corresponding to the peak band indicated by the stored vowel feature information on the spectrum of the input sound. ) And a voice section by detecting whether the input sound corresponds to voice using average energy of an irrelevant band excluding the corresponding band (S330 to S370).

In more detail, the voice interval detector 220 includes an average energy calculator 222 and a similarity measurer 224 to perform a VAD step on the input sound, and includes a hangover processor 226. It may further include.

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 receives the input sound signal through a frequency analysis process. Change to a spectral vector (S330). The spectrum analyzer 240 may perform the frequency analysis 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 (S340).

Next, the similarity measurer 224 measures the spectral similarity between the vowel and the input sound by using an average energy difference between the corresponding band and the non-compliant band (S350).

As mentioned above, vowels in human speech have characteristic spectral peaks. Accordingly, the band of the spectrum of each vowel 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 a voice including the vowel, the mean energy of the corresponding band in the input sound spectrum shows a relatively large value regardless of the presence of noise, and the mean energy of the non-corresponding band is small. Will display the value.

Accordingly, the present invention intends to measure the similarity between the vowel 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 calculate the spectral similarity using a peak-valley energy difference (PVED) value representing a value obtained by subtracting the average energy of the non-corresponding band from the average energy of the corresponding band. It can be measured.

The PVED makes it possible to measure the similarity using only the peak feature vector (S) which is the vowel feature information of the vowel and the spectrum (X) of the input sound.

The PVED 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 of 1 for the peak band of the vowel and a binary value of 0 for the valley region.

When the PVED value is greater than a predetermined threshold (S352), the voice section detector 220 determines the input sound as a voice including the vowel and detects the input sound as a voice section (S360). However, when the PVED value is less than or equal to the threshold value (S352), the input sound is determined as noise or non-voice sound (S362).

According to an embodiment, when the PVED value is less than or equal to the threshold value, the voice interval detector 220 applies a hang-over algorithm to the input sound through the hangover processor 226 (S354). ).

The hangover algorithm refers to a process of processing together an intra-speech silence section of a voice generated during speech in a general voice recognition system. That is, the human voice sounds continuously occurring continuously when it is heard through the ear, but when the voice is analyzed in milliseconds (ms), there is a short silent period of several tens to several hundred ms or less in the middle of the voice. . Therefore, the hangover algorithm is to process the silence appearing in the middle of the voice together.

The present invention also applies the hangover algorithm. That is, after the application of the hangover algorithm by the hangover processing unit 226 (S354), the voice section detection unit 220 belongs to a section in which the input sound is processed as voice (S356). In operation S360, the signal is determined as a voice section.

6 shows the spectrum of the input sound with noise distortion.

As shown in FIG. 6, the background noise, as well as the background spectral peaks of the vowel sound, are partially buried due to channel distortions such as reverberation and short burst noise. Nevertheless, even if there are still characteristic spectral peaks of the vowel sound on the input acoustic spectrum, the existing technique causes misrecognition because it extracts features from the input acoustic spectrum to determine similarity with the vowel spectrum. 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.

7A to 7C show a histogram of relative average energy of corresponding and non-corresponding bands of the input sound spectrum calculated using the vowel feature information of FIG. 5.

As shown in FIGS. 7A and 7B, if the input sound is a voice including the vowel not only when there is background noise in the input sound but also when additional peaks are present, 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. 7C, 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 the misrecognition through the energy-based similarity measuring method, that is, the PVED-based similarity measuring method, and determines the similarity between the vowel and the input sound. Focus on the bands involved. As a result, the present invention can prevent a case where noise or non-voice sound is misrecognized as voice. This result is because the PVED value appears relatively large only if the characteristic peaks of the vowel appear on the spectrum of the input sound.

In one embodiment, when there are a plurality of vowel feature information stored in the vowel feature storage unit 210, the voice section detector 220 calculates the PVED value for each vowel feature information, and calculates the calculated PVED. When the maximum PVED value among the values is greater than the threshold value, the input sound may be determined as a voice having a vowel corresponding to the maximum PVED value. That is, if there were 200 vowel training data for the learning step, 200 vowel feature information, for example, peak signature vectors consisting of binary values would have been generated in the learning step. The PVED value of the input acoustic signal is measured for and the maximum PVED value is compared with the threshold value.

Then, when there is a subsequent input sound input to the voice interval detection system, the above-described voice interval detection process is repeated (S370).

When the speech section is detected from the input sounds by the speech section detecting system 200, the speech recognizer 250 performs a speech analysis, phoneme recognition, word recognition, sentence interpretation, and meaning extraction process using existing techniques. Can be done.

On the other hand, the present invention can be embodied as computer readable program codes 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.

TIMIT, NTIMIT (telephone), FFMTIMIT (far field microphone), and CTIMIT (cellular phone) corpus were used for the performance experiment of the present invention. Each sound was segmented into 32ms blocks at 10 ms intervals, and Fourier analysis was applied to each block through a Hamming window.

In addition, to compare and evaluate the performance of the present invention, the existing VAD schemes for the same test data, that is, "A Silence Compression Scheme for G.729 Optimized for Terminals Conforming to ITU-T V.70" (ITU-T Rec. Method introduced in G.729 Annex B, 1996. (G.729), "Digital Cellular Telecommunications System (Phase 2+); Voice Activity Detector (VAD) for Adaptive Multi Rate (AMR) Speech Traffic Channels" (GSM 06.94 Two methods (AMR1 and AMR2) introduced in v7.1.1 (ETSI EN 301 708), 1999.), A. Davis et al., 2 described "A statistical voice activity detection using low-variance spectrum estimation and an adaptive threshold" ( IEEE Transactions (LVS) as introduced in on Audio, Speech, and Language Processing , vol. 14, no.2, pp.412-424, 2006.), and JL Shen et al. "Robust entropy-based endpoint detection for speech" recognition in noisy environments "( Proceeding of International Conference on Spoken Language Processing , paper 0232, 1998.) was used.

Figure 8 shows the results of the speech section detection performance of the present invention for the four corpus (corpus).

As shown in FIG. 8, it can be seen that the speech segment detection performance of the present invention (PVED) is significantly superior to the existing schemes on average as well as for each corpus.

9 shows a test result of speech section detection performance of the present invention according to various SNR levels.

10 illustrates a test result of speech section detection performance of the present invention according to various background noises.

In order to measure the detection performance of the present invention according to various SNR levels and background noises, 7 SNR levels of background noises were generated in 5dB steps from 0dB to 30dB and mixed with the entire TIMIT corpus to generate test data. The background noise includes 10 background noises: white noise, pink noise, and 8 noises in Aurora-2 corpus (airports, exhibitions, restaurants, streets, Underground passages, cars, trains, and multiple crosstalk noise) were used. That is, test data of 441000 (441000 = TIMIT 6300 x noise type 10 x SNR level 7) were used.

9 and 10, the speech segment detection performance of the present invention (PVED) is, of course, on average. It can be seen that each SNR level and background noise is significantly superior to the conventional methods.

As described above, the present invention provides the advantage of improving speech segment detection performance under various noise environments and SNRs by using the characteristic spectral peaks of the learned vowels without the feature extraction process for the input acoustic signal corrupted by noise. In addition, by reducing the amount of information processing and calculation in the speech section detection process, it provides an advantage that it can be easily applied to resource-constrained environment such as mobile terminal device, USN sensor node that is rapidly developing today. Furthermore, by applying it as a preprocessing step of speech recognition, the speech recognition system can be further improved in energy efficiency by allowing the speech recognition system to remove unnecessary motion in the non-speech section.

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 voice interval detection method using a vowel feature according to the present invention.

2 is a block diagram illustrating an example of a speech section detection system using a vowel feature according to the present invention.

3 is a flowchart illustrating an example of a speech segment detection method using a vowel feature according to the present invention.

4 is a diagram illustrating a frequency spectrum of an 'a' vowel.

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

6 is a diagram illustrating a spectrum of input sound in which noise distortion occurs.

7A to 7C are histograms showing relative average energies of corresponding and non-corresponding bands of the input sound spectrum calculated using the vowel feature information of FIG. 5.

8 to 10 are diagrams showing the test results of the speech section detection performance of the present invention.

Claims (19)

A vowel feature storage unit for storing vowel feature information indicating a peak band in which a feature peak is located in a pre-learned noiseless spectral spectrum; And A voice for detecting a voice section by determining whether the input sound corresponds to voice using the average energy of the corresponding band corresponding to the peak band indicated by the stored vowel feature information in the spectrum of the input sound and the non-corresponding band except the corresponding band. Speech segment detection system using a vowel feature including a segment detector. The method of claim 1, The vowel feature storage unit, A feature peak extracting unit extracting a peak having an energy larger than a predetermined threshold among the spectral peaks of the vowel as a feature peak; And And a feature information generator for generating vowel feature information indicating a peak band in which the extracted feature peak is located in the spectrum of the vowel. The method of claim 2, The feature information generation unit divides the entire spectral band of the vowel into a predetermined number of unit bands, and represents a unit band corresponding to the peak band as 1 in the spectral of the vowel and represents a valley band other than the peak band. and a vowel feature information representing a unit band corresponding to zero). The method of claim 1, The voice section detection unit, An average energy calculator configured to calculate an average energy of a corresponding band corresponding to the peak band of the stored vowel feature information and a non-corresponding band excluding the corresponding band from the spectrum of the input sound; And And a similarity measurer for measuring the spectral similarity between the vowel and the input sound by using an average energy difference between the corresponding band and the non-compliant band. The method of claim 4, wherein The similarity measuring unit measures the spectral similarity using a peak-valley energy difference (PVED) value representing a value obtained by subtracting the average energy of the non-corresponding band from the average energy of the corresponding band. Voice segment detection system. The method of claim 5, The voice section detector, when the PVED value is greater than a predetermined threshold value, the voice section detection system using the vowel feature, characterized in that for determining the input sound as a voice section. The method of claim 6, The voice section detector may calculate the PVED value with respect to the stored vowel feature information and determine the input sound as a voice when the maximum PVED value among the calculated PVED values is greater than the threshold value. Voice segment detection system using features. The method of claim 6, The voice section detector further includes a hangover processor configured to apply a hang-over algorithm to the input sound when the PVED value is not greater than the threshold value. The voice section detection unit, the voice section detection system using the vowel feature, characterized in that for determining the input sound as a voice when the input sound is processed by the hangover processing unit. In a voice activity detection system (Voice Activity Detection system) for detecting a voice interval of the input sound input to the system, A vowel feature storage step of pre-store vowel feature information indicating a peak band at which a feature peak is located in a pre-learned noiseless spectral spectrum; And A voice for detecting a voice section by determining whether the input sound corresponds to voice using the average energy of the corresponding band corresponding to the peak band indicated by the stored vowel feature information in the spectrum of the input sound and the non-corresponding band except the corresponding band. Speech segment detection method using a vowel feature comprising a segment detection step. 10. The method of claim 9, The vowel feature storing step, A feature peak extracting step of extracting a peak having an energy greater than a predetermined threshold among the spectral peaks of the vowel as a feature peak; And And a feature information generating step of generating vowel feature information indicating a peak band in which the extracted feature peak is located in the spectrum of the vowel. The method of claim 10, The generating of the characteristic information may include classifying the entire spectral band of the vowels into a predetermined number of unit bands, indicating a unit band corresponding to the peak band as 1 in the spectral band of the vowel, and having a valley band that is a band other than the peak band ( generating a vowel feature information indicating a unit band corresponding to a valley band of 0 by using a vowel feature. 10. The method of claim 9, The voice section detection step, Calculating an average energy of the corresponding band corresponding to the peak band of the stored vowel feature information and the non-corresponding band excluding the corresponding band from the spectrum of the input sound; And And a similarity measuring step of measuring a spectral similarity between the vowel and the input sound by using an average energy difference 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-valley energy difference (PVED) value representing a value obtained by subtracting the average energy of the non-corresponding band from the average energy of the corresponding band. Speech segment detection method using vowel feature. The method of claim 13, The voice section detecting step, the voice section detection method using the vowel feature characterized in that the step of detecting the input sound to the voice section when the PVED value is greater than a predetermined threshold value. The method of claim 14, The voice section detecting step may include calculating the PVED value with respect to the stored vowel feature information, and determining the input sound as a voice when the maximum PVED value among the calculated PVED values is greater than the threshold value. Speech segment detection method using a vowel feature. The method of claim 14, The voice section detecting step may further include a hangover processing step of applying a hang-over algorithm to the input sound when the PVED value is not greater than the threshold value. The voice section detection step, the voice section detection method using the vowel feature, characterized in that the step of determining the input sound as a voice when the input sound is processed by the hangover processing step. In the method for measuring the similarity (similarity) of the frequency spectrum between the noise-free vowels pre-trained in the Voice Activity Detection system and the input sound input to the system, Calculating an average energy of the corresponding band corresponding to the peak band in which the spectral feature peaks of the vowels are located and the non-corresponding band excluding the corresponding band from the spectrum of the input sound; And And measuring spectral similarity between the vowel and the input sound using an average energy difference between the corresponding band and the non-compliant band. The method of claim 17, The measuring similarity may include measuring the spectral similarity using a peak-valley energy difference (PVED) value representing a value obtained by subtracting the average energy of the non-corresponding band from the average energy of the corresponding band. Acoustic spectral similarity measurement method. A computer-readable recording medium having recorded thereon a program for executing a method according to any one of claims 9 to 18 by a computer.

Images