KR100395222B1 - Voice Recognition System for Voice Mail Service (VMS) - Google Patents

Abstract

The present invention relates to a voice recognition system for receiving a corresponding service by commanding letters or numbers used in a voice mail service (VMS) function with voice. The speech recognition system of the present invention obtains pitch information by extracting a starting point and an end point for each input unit of an input speech, and simultaneously receives preprocessing means for extracting a feature vector of the input speech and user's selection mode information. And a recognition unit that performs word recognition using the detected feature vector and the set recognition target word, and controls the operations of the preprocessing means for extracting the feature vector by discriminating the voice and the voice based on the pitch information. Provides the user's selection mode information to the recognition unit, and includes a control unit for performing a control for transferring the recognition results of the recognition unit to the VMS main processor and a VMS main processor for providing the appropriate service to the user according to the recognition results performed by the recognition unit do. Accordingly, the present invention provides a rejection function for an input other than a voice by using pitch information of a target word of a user, and reduces a target word for each function by using information about a function being used by a user. And the recognition rate is improved. In particular, in the case of four-digit numbers used in the password authentication process, instead of receiving and confirming four-digit voices at the same time, it is implemented by performing recognition on each digit, thereby improving user convenience and recognition performance.

Description

Voice Recognition System for Voice Mail Service (VMS)

The present invention relates to a voice recognition system, and more particularly, to a voice mail service (VMS) for performing a function by inputting a command or a number, etc. used for a voice mail service (VMS) function, by voice. A voice recognition system.

In recent years, voice recognition systems have been applied in a very wide range of fields, and the trend of reconfiguring or improving the recognition device for new applications is on the rise. Among many applications of such voice recognition, voice-activate technology is rapidly developing to replace the existing mode selection function with voice input. In recent years, mobile phone or personal mobile communication has been rapidly developed, so the VMS function is used through various terminals as well as wired telephones. Therefore, a recognition device is needed to perform voice recognition by accurately distinguishing voice (noise) from voice and voice. Do. Existing voicemail service (VMS) can receive the service of the desired function by pressing the desired number in turn according to the announcement. Voice recognition is used to replace the numbers. Such a conventional voice recognition device has been implemented by DTW (dynamic time wrapping) method to reduce the recognition time.

1 is a view showing a conventional voice recognition device using a DTW method. The preprocessor 11 receives a user's voice and extracts a feature vector of a required voice. The DTW recognizer 13 reads a plurality of words to be recognized stored in the memory 15. Therefore, the closest word is output as a recognition result by comparing the input voice with a plurality of words to be recognized by a predetermined voice recognition program technique. The central control unit 17 provides the user with the corresponding service function according to the recognition result. If the user wants to input a number sound as a command, each time the user hears the "beep" sound provided by the system, the user speaks one number. The existing voice recognition device checks the result again after the user inputs the voice command and provides the corresponding service if it is correct.

As described above, the voice recognition device for the conventional voice mail (VMS) function is a very simple voice recognition device, which can recognize very few words, and does not process surrounding noise, and has a very low recognition performance. In addition, the conventional speech recognizer performs recognition after assuming that all the commands used are words to be recognized, and in the case of digit recognition, each digit is independently recognized by one digit after a 'beep' sound. However, this makes the user feel very uncomfortable, and when one number in the middle is wrongly recognized, it is troublesome to utter all numbers again from the beginning.

Accordingly, an object of the present invention is to provide a voice rejection function for a non-speech input using pitch information on a target word of a user, and perform voice recognition using information on a function being used by a user. It provides a voice recognition system for VMS).

Another object of the present invention is to isolate the speech recognition for each digit instead of the continuous word recognition method of receiving and confirming the speech corresponding to all the digits constituting the voice command in succession when recognizing the digits, The present invention provides a voice recognition system for a voice mail service (VMS) that improves user convenience and recognition performance by performing word recognition.

1 is a view showing a conventional speech recognition system,

2 is a view showing a voice recognition system according to a preferred embodiment of the present invention;

3 is a flowchart for explaining a process of determining sound and noise other than voice (noise);

※ Explanation of codes for main parts of drawing

20: preprocessing means 22: echo and noise canceling unit

24: end point extraction and pitch information detection unit 26: feature vector extraction unit

30: recognition unit 40: control unit

50: VMS main processing unit

A feature of the present invention for achieving the above object is a voice recognition system for a voice mail service (VMS), to remove the echo and noise of the input voice, the starting point for each input unit for the echo and noise is removed voice Extracts and end points, obtains pitch information, preprocessing means for extracting feature vectors of the input speech from the pitch information, receives selection mode information of the user, sets recognition target words, and detects the detected feature vectors and the set recognition objects. A recognition unit that performs word recognition using words, and controls the operation of the preprocessing means for extracting a feature vector by discriminating voice and non-voice sound based on the pitch information, and providing the user's selection mode information to the recognition unit. The controller performs a control for transferring the recognition result of the recognition unit to the VMS main processor, and the recognition performed by the recognition unit. According to and in the speech recognition system for voice mail service (VMS) to the user comprises a VMS main processor to provide the appropriate services.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

2 is a diagram illustrating a voice recognition system for a voice mail service (VMS) of the present invention. The voice recognition system of the present invention receives a voice command, the preprocessing means 20 for performing preprocessing for voice recognition, the recognition unit 30 for recognizing the preprocessed voice, the controller 40 for controlling the voice recognition and VMS The main processing unit 50 is included. The preprocessing means 20 includes an echo canceling unit 22 having an echo cancel function for separating the input voice and the announcement broadcast voice and a noise cancel function for removing noise in the input voice, and an input. End point extraction and pitch information detection section 24 for extracting pitch information by extracting the beginning and the end of speech. The preprocessing means 20 also includes a feature vector extraction section 26 for calculating a feature vector to be used in the recognition section 28 for the output of the endpoint extraction and pitch information detection section 24.

The operation of the voice recognition system of the present invention having the above configuration will be described in detail.

In the embodiment of the present invention, the voice command to be input has a pulse code modulation (PCM) data format consisting of 8 ㎑-16 bits.

The echo and noise canceling unit 22 receives a voice command having this data format. The echo and noise control unit 22 extracts only the user's voice when the announcement broadcasted while the user uses the VMS function and the user's voice are mixed with each other and input into the system. The filter coefficients are repetitively changed to remove the announcement voice present in the voice, making it similar to the voice of the user. The noise canceling unit operates with a slight difference from the echo cancellation process for the input voice. The noise canceling unit removes the noise present in the speech by using a spectral substation method for every frame of the input speech, using the first 30 to 50 dB voice data of the input voice as initial information of the noise.

The end point extraction and pitch information detection unit 24 performs two functions, an end point extraction function and a pitch information extraction function. The end point extraction and pitch information detection unit 24 finds the 'start point' and 'end point' of the voice for each frame of the voice from which the echo is removed. In the section between the start point and the end point of the voice, pitch information is extracted using three frame data for each frame. In an embodiment, pitch information extraction uses a first order filter in a manner by autocorrelation. This pitch information refers to the average value, the variance, the difference between the maximum value and the minimum value of the pitch for the entire frame of the input voice, and the rate of change of the pitch value of the front and rear frames. In addition, the end point extraction and pitch information detection unit 24 transmits the extracted pitch information to the control unit 40. The control unit 40 controls the operation of the feature vector extraction unit 26 by using this pitch information to distinguish between voice and non-voice sound (noise). A detailed description of the division of the voice of the controller 40 into the sound (noise) other than the voice will be described later with reference to FIG. 3. The end point extraction and pitch information detection unit 24 outputs data only between the start point and the end point of the input voice command word to the feature vector extraction unit 26. The feature vector extractor 26 calculates the 13 th order MFCC, the 13 th order delta-MFCC, and the 13 th order delta-delta-MFCC using the end point extraction result, and outputs the result to the recognizer 30. The recognition unit 30 stores the feature vector for the voice data in an internal memory (not shown). The target word may be recognized through the beam search and the viterbi decoding on the feature vector. The recognition unit 30 outputs the recognition result of the target word to the control unit 40. The controller 40 compares the pitch information provided from the end point extracting and pitch information detecting unit 24 with the pitch information stored by the user, and the feature vector extracting unit 26 is present in the existing pitch information range. Control to keep running. Otherwise, the feature vector extraction unit 26 is controlled to not operate. At the same time, the control signal is output to the VMS main processor 50 so as to transmit the announcement to the user again, and the control signal is output so that the echo and noise canceling unit 22 is operated to prepare to receive the user's input voice.

In addition, the control unit 40 receives all information on the current recognition mode from the VMS main processing unit 50 and outputs it to the recognition unit 30, and at the same time sends the recognition preparation command signal to the echo and noise canceling unit 22. Output The VMS main processing unit 50 outputs information on which function a user uses among the plurality of VMS functions to the control unit 40. The controller 40 stores this information in an internal memory (not shown). The recognition unit 30 performs recognition using all the information on the currently selected function or the previously selected function by using the recognition target word and a model parameter corresponding thereto. In particular, when the currently selected function is a function such as a password input, at this time, the 4-digit tone becomes the target word. In the present invention, the recognition of the four digit number uses the isolated word recognition method instead of the continuous digit sound recognition method, and proceeds by receiving the recognition result by uttering one digit. The numbers successfully recognized at each digit are temporarily stored in a memory (not shown) of the recognition unit 30, and if the digits are incorrect at each digit, the user is requested to re-create only the wrong digit again. Therefore, if all four-digit recognition is successful, the entire recognition is successful and the password is authenticated.

3 is a flowchart for explaining the operation of the control unit 40 to determine sound and noise other than voice (noise). The controller 40 receives pitch information from the endpoint and the pitch information detector 24 (step 310). Then, the pitch information of the existing rescued word stored in the memory (not shown) is compared with the pitch information of the currently input voice (step 320). Next, it is determined whether the pitch information of the current voice is included in a certain valid system with an error from the most approximate pitch information among the existing pitch information of the constructed word (step 330). If the pitch information of the voice command is included in a certain valid system with an error of the most approximate pitch information among the existing pitch information of the constructed word, it is determined as a word (step 340), and the process proceeds to step 390 to continue the feature vector extraction. To control. However, if the pitch information of the voice is not included in a certain valid system, the error with the most recent pitch information among the pitch information of the constructed word is compared with the pitch information of the existing built noise (step 350). . In step 360, if the error with the pitch information of the existing noise is included in a certain valid system is determined to be noise (step 380), and outputs a control signal to stop the feature vector extraction to the feature vector extraction unit 26 (Step 400). However, in step 360, if the error with the pitch information of the previously constructed noise is not included in a certain valid system, it is determined as a pseudo word (step 370), and the feature vector extraction unit 26 is continued to continue feature vector extraction. To control. Here, the pitch information for each constructed word is newly updated by the recognition result. However, no update is made for words determined as similar words. In addition, even in the case of being judged as noise, the pitch information for the noise is newly updated.

As described above, the voice recognition system for the voice mail service (VMS) of the present invention allows a user to conveniently use a voice mail service (VMS) that has been previously serviced, and the pitch information of the target word of the user. It provides a rejection function for non-speech inputs, and reduces the target word for each function by using information on the function being used by the user, thereby improving recognition speed and recognition rate. In particular, in the case of four-digit numbers used in the password authentication process, instead of receiving and confirming four-digit voices at the same time, it is implemented by performing recognition on each digit, thereby improving user convenience and recognition performance.

Claims (5)

In the voice recognition system for voice mailbox service (VMS), Preprocessing means for removing echo and noise of the input voice, extracting start and end points for each input unit of the voice from which the echo and noise have been removed, and obtaining pitch information, and extracting feature vectors of the input voice from the pitch information; A recognition unit configured to receive a user's selection mode information, set a recognition target word, and perform word recognition using the detected feature vector and the set recognition target word; It controls the operation of the preprocessing means for extracting the feature vector by discriminating the sound and the sound other than the voice based on the pitch information, providing the user's selection mode information to the recognition unit, and transfers the recognition result to the VMS main processor. A control unit which performs a control to perform the control; And A voice recognition system for a voice mail service (VMS) comprising a VMS main processing unit for providing an appropriate service to the user according to the recognition result performed in the recognition unit. The method of claim 1, wherein the pretreatment means An echo and noise canceller having an echo cancel for separating the input voice and the announcement voice and a noise cancel function for removing noise in the input voice; An end point extraction and pitch information detection unit for extracting the 'start point' and 'end point' of the input voice and extracting pitch information using the 'start point' and 'end point'; And And a feature vector extracting unit for calculating a feature vector to be used in the recognition unit with respect to the extracted input voice end point and the pitch information. The apparatus of claim 1 or 2, wherein the controller compares the pitch information of the current voice input with the pitch information of a previously constructed word to convert the word into a word when an error of the most approximate pitch information is included in a certain valid system. Discriminating and controlling the feature vector extracting unit to continuously extract the feature vector. Otherwise, the input voice is discriminated as noise when the error of the most approximate pitch information is included in a certain valid system. Voice recognition system for voice mailbox service (VMS), characterized in that the control to stop the feature vector extraction of the feature vector extraction unit. The apparatus of claim 2, wherein the end point extracting and pitch information detecting unit searches for the 'start point' and 'end point' of the voice for each input unit with respect to the input voice, and a predetermined number of frames for each frame in a section between the start point and the end point of the voice. Speech recognition system for voice mail service (VMS), characterized in that to extract the pitch information using the data. 5. The method of claim 4, wherein the end point extraction and pitch information detection unit uses a first-order filter in an autocorrelation method to obtain an average value, a variance, a difference between a maximum value and a minimum value of the pitch, and the front of the entire frame of the input voice. Voice recognition system for voice mail service (VMS), characterized in that for extracting the pitch information including the rate of change of the pitch value of the rear frame.