KR20060064494A - Method for verifying speech/non-speech and voice recognition apparatus using the same - Google Patents

Abstract

The technical problem to be achieved by the present invention is to more clearly distinguish between speech and non-voice, it is possible to lower the load of the speech recognition unit, and to determine the non-voice signal as a speech signal and to reduce the error of speech recognition generated by speech recognition A non-speech verification method and a speech recognition apparatus using the same are provided.

The present invention provides a speech / non-voice verification unit that extracts a feature vector from input voice data and distinguishes whether the feature vector corresponds to speech or non-voice using a speech / non-voice model; And a voice recognition unit recognizing a voice from data corresponding to a section determined by the voice / non-voice verification unit as the voice.

Description

Speech / non-speech verification method and speech recognition apparatus using the same {Method for verifying speech / non-speech and voice recognition apparatus using the same}

1 is a view showing a speech recognition apparatus according to the prior art.

2 is a diagram illustrating a speech recognition apparatus according to an embodiment of the present invention.

FIG. 3 is a diagram for describing an example in which the voice recognition apparatus of FIG. 2 is connected through a network.

FIG. 4 is a diagram for describing a voice / non-voice verification method performed by the voice / non-voice verification unit 22 of FIG. 2.

FIG. 5 is a diagram for describing an initial modeling method of a voice / non-voice model used in a step corresponding to S43 of FIG. 4.

The present invention relates to a voice / non-voice verification method and a voice recognition device using the same, and more particularly, to a voice / non-voice verification method capable of reducing the load of a voice recognition unit requiring a lot of operations and a voice recognition device using the same. The invention relates to.

1 is a view showing a speech recognition apparatus according to the prior art. Referring to FIG. 1, a speech recognition apparatus according to the related art includes a speech endpoint detector 11 and a speech recognizer 12. The voice endpoint detecting unit 11 detects a start point and an end point of the voice signal section, and detects the voice section using, for example, short-time energy and zero crossing rate of the voice signal section. do. The voice recognition unit 12 performs a function of recognizing the voice in the voice section output from the voice endpoint detecting unit 11.

In the speech recognition apparatus according to the prior art having such a configuration, the voice end point detection unit 11 has a limitation in distinguishing between the voice signal and the non-voice signal. In particular, there is a problem in that it is not possible to distinguish between the living signal that can be commonly encountered in the surroundings such as mechanical sound and music. As described above, when the voice endpoint detector 11 does not distinguish the voice signal from the non-voice signal well and recognizes and outputs most of them as voice signals, the voice recognition unit 12 has a problem of performing a large amount of calculation. . In particular, when applied to a robot, it is necessary to determine whether the sound heard in the standby state is voice or non-voice, unlike the method of talking after pressing a button, that is, a push-button method. Therefore, when the voice and the non-voice can not be distinguished well, there is a problem that the rechargeable battery of the robot is quickly consumed by the frequent operation of the voice recognition unit 12. In addition, despite receiving a non-negative signal, there is a problem in that an error of speech recognition occurs by judging by voice and performing speech recognition.

Therefore, the technical problem to be solved by the present invention is to solve the above problems, and by clearly distinguishing between speech and non-voice, it is possible to lower the load of the speech recognition unit, and to cause a recognition error by receiving a non-voice signal. It is an object of the present invention to provide a speech / non-voice verification method and a speech recognition apparatus using the same.

As a technical means for achieving the above object, the first aspect of the present invention is to extract a feature vector from the input voice data, and to use the voice / non-voice model to determine whether the feature vector corresponds to speech or non-voice Voice / non-voice verification unit for distinguishing whether or not; And a voice recognition unit recognizing a voice from data corresponding to a section determined by the voice / non-voice verification unit as the voice.

In addition, a second aspect of the present invention includes the steps of (a) extracting a feature vector from the data in units of frames; (b) performing speech / non-voice determination on a frame-by-frame basis using the speech / non-voice model; (c) buffering speech / non-voice determination values of a plurality of consecutive frames by the length of the window; And (d) making a voice / non-voice decision on a window basis.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, embodiments of the present invention may be modified in various forms, and the scope of the present invention should not be construed as being limited by the embodiments described below. Embodiments of the present invention are provided to more fully explain the present invention to those skilled in the art.

2 is a diagram illustrating a speech recognition apparatus according to an embodiment of the present invention. Referring to FIG. 2, the voice recognition apparatus includes a voice endpoint detector 21, a voice / non-voice verification unit 22, and a voice recognition unit 23.

The voice end point detector 21 detects a start point and an end point of the voice signal section. For example, the voice endpoint detection unit 21 detects the voice section by using a short-time energy and a zero crossing rate of the audio signal. do. The voice endpoint detection unit 21 is an optional component and can achieve the object of the present invention by using only a voice recognition device including only the voice / non-voice verification unit 22 and the voice recognition unit 23. However, when the voice endpoint detection unit 21 is additionally included, the load of the voice / non-voice verification unit 22 may be reduced.

The voice / non-voice verification unit 22 performs a function of finally detecting the voice section and the non-voice section by using a feature vector on the data of the voice section output from the voice endpoint detector.

Examples of feature vectors extracted by the speech / non-voice verification unit 22 include filter bank energy, pitch, amount of change of energy value between filter banks in a frame, amount of change of filter bank energy values between frames, mel filter bank coefficients, and the like. . The speech / non-voice verification unit 22 distinguishes whether the feature vector corresponds to speech or non-voice using the speech / non-voice model. When the voice / non-voice verification unit 22 detects a voice section, the voice / non-voice verification unit 22 outputs at least one of voice data corresponding to the voice section and a feature vector corresponding to the voice section to the voice recognition unit 23. If the voice / non-voice verification unit 22 outputs the feature vector to the voice recognition unit 23, and the voice recognition unit 23 performs voice recognition using the voice recognition unit 23, the voice recognition unit 23 performs the feature. The advantage is that it does not have to include a separate component for extracting the vector.

The speech recognizer 23 performs a function of recognizing the speech in the speech section using at least one of the speech data and the feature vector output from the speech / non-voice verifier 22.

FIG. 3 is a diagram for describing an example in which the speech recognition apparatus of FIG. 2 is connected through a network. Referring to FIG. 3, the speech recognition apparatus includes a speech recognition server 31 and at least one client 32A, 32B, and 32C.

The voice recognition server 31 is connected to at least one client 32A, 32B, 32C through communication, and includes at least a voice recognition unit 23.

Each client 32A, 32B, 32C is connected to the voice recognition server 31 through communication, and has microphones 33A, 33B, 33C, voice endpoint detectors 21A, 21B, 21C, and voice / non-voice verification unit ( 22A, 22B, 22C). The clients 32A, 32B, 32C may preferably be robots.

In the speech recognition apparatus represented in the figure, the microphones 33A, 33B, and 33C should be located at the clients 32A, 32B, and 32C, and the speech recognition unit 23 should be located at the speech recognition server 31. The voice endpoint detectors 21A, 21B, 21C and the voice / non-voice verifiers 22A, 22B, 22C may be located at the clients 32A, 32B, 32C as shown in the figure, and the voice endpoint detectors 21A, 21B, 21C) may be located on the clients 32A, 32B, 32C, and the voice / non-voice verification units 22A, 22B, 22C may be located on the voice recognition server 31, and the voice endpoint detectors 21A, 21B, 21C and All of the voice / non-voice verification units 22A, 22B, and 22C may be located in the voice recognition server 31. When the voice endpoint detectors 21A, 21B, and 21C are located in the voice recognition server 31, a separate voice endpoint detector may be provided for each client 32A, 32B, and 32C, and a plurality of clients 32A, 32B, One voice endpoint detector may be provided for 32C). In addition, when the voice / non-voice verification units 22A, 22B, and 22C are located in the voice recognition server 31, a separate voice / non-voice verification unit may be provided for each client 32A, 32B, and 32C. One voice / non-voice verification unit may be provided for the clients 32A, 32B, and 32C. When only the microphones 33A, 33B, 33C are located at the clients 32A, 32B, 32C, the clients 32A, 32B, 32C must communicate with the voice recognition server 31 at all times or frequently. The problem is that it puts a lot of load. This problem is particularly acute in applications such as robots that do not operate in a push-button manner. Accordingly, the voice endpoint detectors 21A, 21B, 21C are located at the clients 32A, 32B, 32C, and the voice / non-voice verification units 22A, 22B, 22C are located at the voice recognition server 31, or as shown in the figure. It is more preferable that the voice endpoint detection sections 21A, 21B, 21C and the voice / non-voice verification sections 22A, 22B, 22C are located at the clients 32A, 32B, 32C because the communication load can be reduced. If the voice endpoint detectors 22A, 22B, 22C are not used, the location of the voice / non-voice verifiers 22A, 22B, 22C at the clients 32A, 32B, 32C can reduce the communication load. More preferred.

FIG. 4 is a diagram for describing a voice / non-voice verification method performed by the voice / non-voice verification unit 22 of FIG. 2. Referring to FIG. 4, the voice / non-voice verification method includes a first buffering step S41, a feature vector extraction step S42, a model improvement step S43, a frame classification step S44, and a second buffering step S45. And window classification step S46.

In the first buffering step S41, an operation of outputting the input voice data in one frame unit is performed. The length of the frame is a length suitable for extracting the feature vector. For example, the length of the frame may be 20 ms, and the frame may be obtained by overlapping the front frame and the rear frame by 10 ms.

In the feature vector extraction step (S42), for example, at least one of a feature vector such as a filter bank energy, a pitch, an amount of change of energy values between filter banks in a frame, an amount of change of filter bank energy values between frames, and a mel filter bank coefficient is extracted. The filter bank energy refers to a value obtained by obtaining energy of required frequency bands among frequency bands of voice data. The filter is used to extract only the required frequency band, and the filter bank is a set of filters. The filter bank energy is expressed in the form of one real value per filter. The amount of change in the energy value between filter banks in a frame means the amount of energy difference between filter banks in one frame and is expressed by the number less than the number of filters. The amount of change in the filter bank energy value between frames means a difference of energy values from the same filter between adjacent frames among frame columns over time. Mel filter bank coefficients are also referred to as MFCC (Mel Frequency Cepstrum Coefficient), by using the filter of the frequency band to Mel frequency band to obtain an energy value and take an Inverse Fast Fourier Transform again The value obtained.

In the model improvement step S43, the remodeling of the voice / non-speech model is performed, and the remodeling is performed with the optimized model using an adaptive technique. As an adaptive technique, at least one of a method such as an eigen voice, a maximum likelihood linear regression (MLLR), a maximum A-Posterior (MAP), or the like may be used. Further, in improving the voice / non-voice model, remodeling can be done on-line. As such, when the voice / non-voice model is improved, the voice / non-voice decision may be more accurate. The model improvement step S43 is an optional step and may be performed as necessary.

In the frame classification step S44, the voice / non-voice decision is made on a frame-by-frame basis using the voice / non-voice model. In the determination of speech / non-speech, when statistical modeling is used, at least one of methods such as Gaussian Mixture Model (GMM), Hidden Markov Model (HMM), Support Vector Machine (SVM), and Neural Network (NN) Determining speech / non-voice using one method, and modeling by rule-based method, determines voice / non-voice for a frame using rules. In outputting the voice / non-voice determination value, the determination value may be output by a hard determination method, that is, a method of simply outputting one of a value corresponding to voice and a value corresponding to non-voice. soft) The determination value may be output by a determination method, i.e., a method of outputting a degree close to voice or non-voice by real.

In the second buffering step S45, the decision value for the voice / non-voice performed in the frame classification step S44 is buffered for the consecutive frames by the length of the window. A window is a set of consecutive frames that uses an appropriate size, such as 300ms to 1000ms. The feature vector extracted in the feature vector extraction step S42 in the second buffering step S45 may be additionally buffered.

In the window classification step S46, a voice / non-voice decision is finally made in the window unit. In performing the voice / non-speech determination, it may be determined using a rule-based method, that is, a threshold value, or may be determined using a statistical method, that is, a classifier GMM, HMM, SVM, NN, or the like.

By performing such a step, it is possible to determine whether the input data is speech or non-voice, and in response thereto, a speech section may be determined and the corresponding data (voice signal and / or feature vector) may be delivered. In particular, in determining the voice / non-voice, by using the feature vector, the voice / non-voice is more accurately compared to the voice endpoint detection unit using short-time energy and / or zero crossing rate. You can decide.

FIG. 5 is a diagram for describing an initial modeling method of a voice / non-voice model used in a step corresponding to S43 of FIG. 4. The initial modeling method of the speech / non-speech model is performed before the speech / non-speech verification method represented in FIG. 4, and the speech / non-speech model obtained after the performance is used in the speech / non-speech verification method. The modeling process is to elaborate each model using feature vectors for speech / non-voice, preferably off-line. Referring to FIG. 5, the initial modeling method of the speech / non-voice model includes a buffering step S51, a feature vector extraction step S52, and a speech / non-voice modeling step S53.

In the buffering step S51, the input voice data is buffered and then output in one frame unit. The length of the frame is a length suitable for extracting the feature vector, and may be the same as the frame length used in the first buffering step of FIG. 4.

In the feature vector extraction step (S52), for example, feature vectors such as filter bank energy, pitch, change amount of energy value between filter banks in a frame, change amount of filter bank energy values between frames, and mel filter bank coefficients are extracted.

In the speech / non-speech modeling step S53, a statistical method may be used as a model making method, or a rule setting method for a rule-based model may be used, or a hybrid method of two methods may be used. . Statistical methods include vector quantization, Gaussian modeling, etc. Discriminative learning can be used to improve their discrimination.

Tables 1 and 2 expressed below are tables for comparing the performance of the speech recognition apparatus according to the embodiment of the present invention and the speech recognition apparatus according to the prior art.

vocalization Recognition error Refuse input Error rate 359 359 0 100%

vocalization Recognition error Refuse input Error rate 359 62 297 17%

Table 1 is a diagram showing the degree to which an error occurs when the voice recognition device according to the prior art recognizes this as a voice when noise other than voice is input. As can be seen from Table 1, when inputting 359 noises, the input was not rejected because it was recognized as non-voice, and all of them were recognized as voice and a recognition error occurred. Therefore, the error rate corresponds to 100%.

Table 2 is a diagram showing the extent to which an error occurs when the voice recognition device according to the present invention recognizes this as a voice when noise other than voice is input. As shown in Table 2, when 359 noises were input, 297 times were denied by recognizing it as non-voice, and 62 times when recognition errors occurred by recognizing it as voice. Therefore, the error rate is 17%, and it can be clearly seen that the speech recognition apparatus according to the present invention has an improved effect of reducing the speech recognition error by removing the non-voice noise.

Voice / non-voice verification method and voice recognition device using the same according to the present invention by clearly distinguishing the speech section and the non-voice section, compared to the prior art, the load of the speech recognition unit that requires a lot of operations and the time required for speech recognition It has the advantage of reducing the error of recognition by receiving a non-voice signal.

In addition, the voice / non-voice verification method and the voice recognition apparatus using the same in accordance with the present invention loads the voice recognition server when the voice recognition server to recognize the voice signals generated from a plurality of clients (such as robots at home) or The advantage is that the number can be reduced and the recognition performance can be improved.

Claims (15)

A voice / non-voice verification unit which extracts a feature vector from input voice data and distinguishes whether the feature vector corresponds to voice or non-voice using a voice / non-voice model; And And a speech recognizer configured to recognize a speech from data corresponding to a section determined by the speech / non-voice verifier as a speech. The method of claim 1, And the feature vector is at least one of a filter bank energy, a pitch, an amount of change of energy values between filter banks in a frame, an amount of change of filter bank energy values between frames, and a mel filter bank coefficient. The method of claim 1, And a data corresponding to the section determined as the voice is at least one of a feature vector and voice data. The method according to any one of claims 1 to 3, The voice / non-voice verification unit is located in the client, the voice recognition unit is located in the voice recognition server. The method according to any one of claims 1 to 3, And a voice end point detector to detect a voice section using at least one of a short time energy and a zero crossing rate, and transmit data corresponding to the voice section to the voice / non-voice verification unit. The method of claim 5, wherein The voice endpoint detecting unit is located in the client, the voice recognition unit is located in the voice recognition server, the voice / non-voice verification unit is located in the client or the voice recognition server. (a) extracting a feature vector from frame-based data; (b) performing speech / non-voice determination on a frame-by-frame basis using the speech / non-voice model; (c) buffering speech / non-voice determination values of a plurality of consecutive frames by the length of the window; And (d) making voice / non-voice decisions on a window-by-window basis. The method of claim 7, wherein And the feature vector is at least one of a filter bank energy, a pitch, an amount of change of energy values between filter banks in a frame, an amount of change of filter bank energy values between frames, and a mel filter bank coefficient. The method of claim 7, wherein In determining the voice / non-voice in step (b), the voice / non-voice verification method is determined using at least one of a statistical method and a rule-based method. The method of claim 7, wherein And (c) additionally buffering the feature vector. The method of claim 7, wherein In determining the voice / non-voice in step (d), the voice / non-voice verification method is determined using at least one of a statistical method and a rule-based method. The method according to any one of claims 7 to 11, Performed before step (a) and (e) buffering and outputting the input voice data into data in units of frames. The method according to any one of claims 7 to 11, After the step (a) (f) remodeling the speech / non-speech model using an adaptive technique using the feature vector. The method according to any one of claims 7 to 11, Performed before step (a) (g) further comprising a voice / non-voice model initialization step. The method of claim 14, Step (g) Buffering input voice data for initialization in units of frames; Extracting a feature vector for initialization from input voice data for initialization of the buffered frame unit; And Creating a speech / non-speech model using the feature vector for initialization.

Images