KR20020049764A - Apparatus and method for speech detection using multiple sub-detection system - Google Patents

Abstract

PURPOSE: An apparatus and a method for detecting speech using multiple lower speech detection systems are provided to collect the operation results of multiple speech detecting systems that operate independently so as to obtain optimized detection performance. CONSTITUTION: An apparatus for detecting speech using multiple low speech detection systems includes a speech signal input unit(21), a plurality of detectors(22), a detection result combination unit(23), and a detection result output unit(24). The speech signal input unit receives an input speech signal transmitted from the outside. Each of the detectors detects the start point and end point of the input speech signal. The detection result combination unit combines the results obtained by the plurality of detectors to calculate a weighted average value. The detection result output unit outputs the detected result.

Description

Apparatus and method for speech detection using multiple sub-detection system}

The present invention relates to a voice detection technique using a plurality of lower voice detection systems. In particular, the weighted average value is obtained by varying weights for each system for a start point and an end point of a voice independently obtained by a plurality of lower voice detection systems. The present invention relates to a voice detection apparatus using a plurality of subordinate voice detection systems for detecting the start and end points of a conventional voice, and a method and a computer readable recording medium storing a program for realizing the method.

The present invention belongs to all fields using voice, such as speech recognition, speech synthesis, speech coding, and the like. In these fields, it is very important to accurately detect only a speech part in a signal containing speech.

In general, one of the main causes of malfunction of the voice recognizer in a real environment is that the voice detector outputs an incorrect voice detection result. In other words, when the non-voice is included in the speech region or the speech region is cut out, a misrecognition result is necessarily output through the recognizer.

Many methods for voice detection have been proposed. Each of these methods has their advantages, but none of them can perform well. Methods using multiple subsystems have also been presented, but this is to make judgments based on the results of each system at every moment and not to use the final detection result of each system.

Therefore, in the conventional methods, even if several systems are used, the performance of each system may not be derived, and there is a problem in that various systems cannot be used due to the constraint of real-time processing.

The present invention has been proposed to solve the above problems, the starting point and the end point obtained from several independent sub-voice detection systems in order to detect the voice part without the limitation of real-time processing in the system using the voice as an input A speech detection device using a plurality of subordinate voice detection systems for detecting the starting and ending points of a final voice, and a computer readable recording medium having recorded thereon a method and a program for realizing the above method. The purpose is to provide.

1 is an exemplary configuration of a general voice recognition system including a voice detection unit to which the present invention is applied.

2 is a block diagram of an embodiment of a voice detection apparatus using a plurality of lower voice detection systems according to the present invention;

3 is a flowchart illustrating a voice detection method using a plurality of lower voice detection systems according to the present invention.

Figure 4 is a flow diagram of an embodiment of a training process for determining the weight for each lower voice detection system in the voice detection method using a plurality of lower voice detection system according to the present invention.

* Explanation of symbols for the main parts of the drawings

11: voice input unit 12: voice detection unit

13 recognition unit 21 signal input unit

22: voice detector 23: detection result combination unit

24: detection result output unit

In order to achieve the above object, the present invention provides a voice detection apparatus using a plurality of lower voice detection systems, comprising: voice signal input means for receiving a voice input signal transmitted from the outside; A plurality of detection means for detecting a start point and an end point of the voice independently of the voice input signal transmitted through the voice signal input means; Detection result combining means for combining the detection results obtained through the plurality of detection means to obtain a weighted average value; And detection result output means for outputting detection results for the start point and the end point of the voice transmitted through the detection result combining means.

The present invention also provides a voice detection method applied to a voice detection apparatus using a plurality of lower voice detection systems, the method comprising: a first step of receiving a voice input signal transmitted from an external device; Detecting a start point and an end point of a voice independently of the input voice input signal; A third step of obtaining a weighted average value by combining the obtained detection results; And a fourth step of outputting the obtained weighted average value as a start point and an end point of the voice obtained by the entire system.

In addition, the present invention, the voice detection device using a plurality of lower voice detection system having a processor, the first function for receiving a voice input signal transmitted from the outside; A second function of detecting a start point and an end point of a voice independently of the input voice input signal; A third function of combining the obtained detection results to obtain a weighted average value; And a computer readable recording medium having recorded thereon a program for realizing a fourth function of outputting the obtained weighted average value as a starting point and an end point of the speech obtained by the entire system.

The present invention combines the results of voice detection systems that operate independently of each other, weighting each system to obtain a final result. The weight is determined to give the optimal voice detection performance during training.

The present invention is characterized by obtaining the detection performance superior to the detection performance of the previously announced voice detection system by combining the results of several independently operated voice detection systems to show the optimal detection performance. That is, the present invention does not aim at real time processing, but first, several lower voice detection systems independently detect the starting point of the voice. Next, a weighted average is obtained for several detected starting points. At this time, the weight for each detection system is determined in the training process. The same goes for the endpoint. The starting point and the end point obtained by weighted average are the starting point and the end point obtained from the entire voice detection system.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is an exemplary configuration of a general voice recognition system including a voice detection unit to which the present invention is applied.

As shown in FIG. 1, the voice input unit 11 reads an input signal from a microphone or a signal stored in a file form and transmits the read signal to the voice detector 12 and the recognizer 13.

Then, the voice detector 12 detects the voice portion from the transmitted input signal, and transmits the start time and the end time information of the detected voice portion to the recognition unit 13.

Subsequently, the recognition unit 13 recognizes only the voice part of the input signal transmitted from the voice input unit 11 by using the detection result transmitted from the voice detector 12 and outputs the recognition result.

2 is a block diagram of an embodiment of a voice detection apparatus using a plurality of lower voice detection systems according to the present invention.

As shown in FIG. 2, a voice detection apparatus using a plurality of lower voice detection systems according to the present invention includes a signal input unit 21 and a signal input unit 21 for receiving a voice input signal transmitted from a voice input unit. Detection result combination unit for obtaining a weighted average value by combining the voice detector 22 for detecting the start and end points of the voice independently of the voice input signal transmitted through the voice detector 22 and the detection results obtained through the voice detector 22 ( 23 and a detection result output section 24 for outputting detection results for the start and end points of the voice transmitted through the detection result combination section 23.

Referring to the operation of the voice detection device of the present invention having the structure as described above in detail.

First, the input signal transmitted from the voice input unit is transmitted to the detector 22. In FIG. 2, an example using a plurality of voice detectors 221 to 22N is illustrated, but there are no limitations on the types and numbers of voice detectors used in actual applications, and any type of voice detectors may be used. In one embodiment, detection systems in the following list may be used.

First, the voice detector 1 221 uses a cepstrum as a feature vector for voice detection, and calculates an average value for this portion by assuming that the front part of the input voice is non-voice. In each frame, the difference between the non-voice part and the spectral value is obtained.

Second, voice detector 2 222 uses normalized energy and zero crossing rate as parameters for voice detection. At this time, if the energy and the zero crossing rate exceeds the threshold, it is determined that the voice, otherwise it is non-voice.

Third, the voice detector 3 223 generates a statistical model for the non-voice part and the voice part in advance, and obtains a probability value of the model for the non-voice part and the model for the voice part every frame of the input voice. Is detected.

Fourth, the voice detector 4 224 detects the voice part using the primary value C1 of the normalized Mel-Frequency Cepstral Coefficient. If the C1 value exceeds a predetermined threshold, it is determined to be a voice part.

In this way, the plurality of voice detectors 221 to 22N independently detect voice signals input from the voice input unit. The time information on the start point and the end point of the voice, which are the results detected by the plurality of voice detectors 221 to 22N, respectively, is transmitted to the detection result combination unit 23. The detection result combination unit 23 calculates a weighted average value of the transmitted detection results by Equation 1 and Equation 2 below. The weighted average value thus obtained is time information about the start and end points of the final speech portion detected by the entire speech detection system.

Here, S and E are the start point (time) and end point of the voice portion output as the final result of the entire detection system.

Is the weight that is applied to the i th detector in the process of combining the detection results of the respective detectors to finally obtain the starting point of speech. Is the weight applied to the i th detector in the process of combining the detection results of the respective detectors to finally find the end point of the voice, and these weights are determined in a separate training process.

And Is the start and end points obtained from the i th detector.

3 is a flowchart illustrating a voice detection method using a plurality of lower voice detection systems according to the present invention.

As shown in FIG. 3, the voice input signal is input from the voice input unit 301 to start detecting the start point of the voice independently through each voice detector 302, and the voice detector which detects the start point then detects the end point. (303).

Then, it is confirmed whether the beginning of speech is detected at all detectors (304), and if the beginning of speech is detected at all detectors, the individual speech results are combined to output the final speech starting point (305), and if not, all speech is detected. The process of checking whether the beginning of the voice is detected in the detector is repeated 303. At this time, when the start portion detection process is completed, the starting point of the speech is finally obtained by using Equation 1 and Equation 2 and output.

Then, it is checked whether all voice detectors have detected the end of the voice (306). If the end is detected, the individual voices are combined to output the final voice endpoint (307). Checking whether the end is detected is repeated 306. At this time, when the end detection process is completed, the final end point of the voice is finally obtained by using Equation 1 and Equation 2 and output.

4 is a flowchart illustrating a training process for determining a weight for each lower voice detection system in a voice detection method using a plurality of lower voice detection systems according to the present invention.

As shown in FIG. 4, the voice data used in the training process for determining the weight is a state where the start point and the end point of the voice unit are detected by hand. Each detector that receives training voice data (401) performs voice detection independently for each detector (402) and stores the performed results, respectively (403).

Subsequently, it is checked whether the voice detection process is finished for all training data (404), and when it is finished, using the stored detection result and label data, that is, the detection result obtained by hand, the following [Equation 3] and Equation 4 determines the weight for each detector (44).

here, Is a weight applied to the i-th detector in the process 305 of combining the detection results of the respective detectors to finally obtain the starting point of the speech by the voice detection system to which the present invention is applied. Is the weight applied to the i th detector in the step 307 of combining the detection results of each detector to finally find the end point of the voice.

Is the starting point of the voice obtained from the i th detector for the j th input voice during the training process, Is the end point of the voice obtained by the i th detector for the j th input voice during the training process.

Is the starting point of the voice manually obtained for the jth input voice during training. Is the endpoint of the voice manually obtained for the jth input voice during training.

Here, N is the total number of lower voice detectors, and M is the size of the training data (number of speeches).

The method of the present invention as described above may be implemented as a program and stored in a computer-readable recording medium (CD-ROM, RAM, ROM, floppy disk, hard disk, magneto-optical disk, etc.).

The present invention described above is not limited to the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes are possible in the art without departing from the technical spirit of the present invention. It will be apparent to those of ordinary knowledge.

As described above, the present invention combines the results of independently operating voice detection systems, weights each system to obtain a final result, and the weight is obtained in a training process to obtain an optimal voice detection performance. Therefore, the present invention has the effect of obtaining a detection performance superior to that of the previously announced voice detection system by combining the results of several independently operated voice detection systems to show the optimal detection performance.

Claims (7)

In the voice detection device using a plurality of lower voice detection system, Voice signal input means for receiving a voice input signal transmitted from the outside; A plurality of detection means for detecting a start point and an end point of the voice independently of the voice input signal transmitted through the voice signal input means; Detection result combining means for combining the detection results obtained through the plurality of detection means to obtain a weighted average value; And Detection result output means for outputting detection results for the start point and end point of the voice transmitted through the detection result combining means Voice detection apparatus using a plurality of lower voice detection system comprising a. The method of claim 1, The weighted average value is, A voice detection apparatus using a plurality of subordinate voice detection systems, which is obtained by the following equation and is time information on the start and end points of the final voice portion detected by the entire voice detection system. , Where S and E are the starting point (time) and ending point of the speech part output as the final result of the whole detection system, Is the weight applied to the i th detector in the process of combining the detection results of the detectors to finally find the starting point of the speech. Is the weight applied to the i th detector in the process of combining the detection results of the detectors to finally find the end point of the voice, And Is the start and end points of the i th detector for the input voice) The method according to claim 1 or 2, The weight is, A voice detection device using a plurality of subordinate voice detection systems, which is determined for each detector by the following equation, and is determined in a process of obtaining from training data. , (here, Is the weight that the voice detection system applies to the i th detector in the process of combining the detection results of the detectors to finally find the starting point of the voice, Is the weight applied to the i th detector in the process of combining the detection results of each detector to finally find the end point of the voice, Is the starting point of the voice obtained from the i th detector for the j th input voice during the training process, Is the end point of the voice obtained from the i th detector for the j th input voice during training. Is the starting point of the voice manually obtained for the jth input voice during training. Is the endpoint of the voice manually obtained for the jth input voice during training, N is the total number of lower voice detectors, and M is the size of the training data (number of speeches). In the voice detection method applied to a voice detection device using a plurality of lower voice detection system, A first step of receiving a voice input signal transmitted from the outside; Detecting a start point and an end point of a voice independently of the input voice input signal; A third step of obtaining a weighted average value by combining the obtained detection results; And A fourth step of outputting the obtained weighted average value as a start point and an end point of the voice obtained by the entire system; Voice detection method using a plurality of lower voice detection system comprising a. The method of claim 4, wherein The weighted average value is, A voice detection method using a plurality of subordinate voice detection systems, which is obtained by the following equation and is time information on the start and end points of the final voice portion detected by the entire voice detection system. , Where S and E are the starting point (time) and ending point of the speech part output as the final result of the whole detection system, Is the weight applied to the i th detector in the process of combining the detection results of the detectors to finally find the starting point of the speech. Is the weight applied to the i th detector in the process of combining the detection results of each detector to finally find the end point of the voice, And Is the starting point and ending point from the i th detector) The method according to claim 4 or 5, The weight is, A method of detecting a voice using a plurality of lower voice detection systems, wherein each detector is obtained by the following equation, and is determined through a process of obtaining from training data. , (here, Is the weight that the voice detection system applies to the i th detector in the process of combining the detection results of the detectors to finally find the starting point of the voice, Is the weight applied to the i th detector in the process of combining the detection results of each detector to finally find the end point of the voice, Is the starting point of the voice obtained from the i th detector for the j th input voice during the training process, Is the end point of the voice obtained from the i th detector for the j th input voice during training. Is the starting point of the voice manually obtained for the jth input voice during training. Is the endpoint of the voice manually obtained for the jth input voice during training, N is the total number of lower voice detectors, and M is the size of the training data (number of speeches). In the voice detection device using a plurality of lower voice detection system having a processor, A first function of receiving a voice input signal transmitted from the outside; A second function of detecting a start point and an end point of a voice independently of the input voice input signal; A third function of combining the obtained detection results to obtain a weighted average value; And A fourth function of outputting the obtained weighted average value as a start point and an end point of the voice obtained by the entire system; A computer-readable recording medium having recorded thereon a program for realizing this.