KR20040082756A - Method for Speech Detection Using Removing Noise - Google Patents

Abstract

PURPOSE: A speech extracting method through removal of non-voice is provided to remove non-voice such as noise and echo to efficiently extract speech. CONSTITUTION: An external speech signal is received and an estimated start point that is presumed as a voice is monitored and detected(201). It is judged whether a frame corresponding to the estimated start point satisfies a frame size boundary condition and an accumulation energy boundary condition(202). When the frame does not satisfy the conditions, the frame is decided as non-voice and removed(205). When the frame satisfies the conditions, the estimated start point is decided as a start point and an end point is detected to extract a speech signal(204).

Description

Speech extraction method using non-voice removal {Method for Speech Detection Using Removing Noise}

The present invention relates to a method for extracting speech by non-voice removal and a computer-readable recording medium having recorded thereon a program for realizing the method.

Speech recognition is a method of extracting language information included in a human voice, and refers to a technology for performing a specific operation by analyzing a feature of a voice input through a microphone, a headset, a telephone, and the like. Such voice recognition technology is closely related to real life, such as home automation, voice recognition toy, voice recognition language learner, voice recognition browser, voice recognition game, voice recognition mobile communication terminal, voice recognition appliances, stock trading system, It is widely used in various fields such as automatic guidance system.

In general, voice recognition is largely a voice input process for receiving a voice through a microphone, a headset, a telephone, a voice extraction process for extracting only voices without noise from the input voice, and a type of voice compression to filter human speech organs. Speech feature extraction process for finding coefficients and speech recognition process for recognition / misrecognition by recognition of speech using recognition algorithm. At this time, the voice signal input during the voice input process includes a signal generated by the user's voice, that is, voice, but may also include non-voice such as noise and echo. Noise refers to car noise, music sound, and noise signal from telephone network that may occur in the surrounding environment of the speaker, and echo refers to a signal in which the voice of the scenario announcement of an application service using voice recognition echoes through a communication circuit. Means.

However, since the general noise in speech recognition is relatively small compared to the size of the talker's voice, only the amount of energy is examined to extract the talker's voice. However, sparking noise (e.g. , The car horn sound is extracted along with the talker's voice because the energy of the unit frame is instantaneously loud. As described above, sparking noise having a large amount of energy can be removed by using a speech recognition algorithm. However, there is a problem that the load of the speech recognizer is increased and the processing time is long.

In addition, since echo has similar characteristics to the voice of the talker, it is difficult to identify the speech using a speech recognition algorithm using a noise model, and thus, a performance degradation of the speech recognizer occurs, resulting in a long processing time. In addition, as the platform of the voice recognizer is changed to a personal computer equipped with a voice processing board instead of dedicated hardware, a method for solving such a problem is required.

The present invention has been proposed to solve the above problems, and in order to efficiently extract speech by removing non-voice such as noise or echo in the speech extraction apparatus of the speech recognition system, the speech extraction method by non-voice removal And a computer readable recording medium having recorded thereon a program for realizing the method.

1 is a configuration diagram of an embodiment of a speech recognition system to which the present invention is applied.

Figure 2 is a flow diagram of an embodiment of a speech extraction method by non-voice removal according to the present invention.

Figure 3a and Figure 3b is a specific example of the speech extraction method by non-speech removal according to the present invention.

* Explanation of symbols on the main parts of the drawing

10: voice recognition system 11: voice input unit

12: A / D converter 13: voice extractor

14: voice recognition unit

According to an aspect of the present invention, there is provided a speech extraction method applied to an apparatus for extracting speech by non-speech, comprising: a first step of receiving and receiving a speech signal from an outside to monitor and detect an expected starting point estimated as speech; A second step of determining whether a frame corresponding to the expected starting point satisfies a frame size boundary condition and a cumulative energy boundary condition according to a voice feature; A third step of determining that the frame is non-voice and removing if not satisfied as a result of the determination of the second step; And a fourth step of determining the expected start point as a start point, detecting an end point, and extracting a voice signal when the determination result of the second step is satisfied.

The present invention also provides a speech extracting apparatus having a processor, comprising: a first function of receiving a speech signal from outside and monitoring and detecting an expected starting point estimated by speech; A second function of determining whether a frame corresponding to the expected starting point satisfies a frame size boundary condition and a cumulative energy boundary condition according to a voice feature; A third function of determining and removing the corresponding frame as non-voice if it is not satisfied as a result of the determination in the second function; And a computer-readable recording medium having recorded thereon a program for realizing the fourth function of determining the expected starting point as a starting point, detecting the end point, and extracting a voice signal if the determination result of the second function is satisfied.

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 a configuration diagram of an embodiment of a speech recognition system to which the present invention is applied.

As shown in FIG. 1, the voice recognition system 10 to which the present invention is applied includes a voice input unit 11 for receiving voice from the outside and an analog voice signal received from the voice input unit 11. A / D converter 12 for converting into a signal, the voice to extract the voice by removing the non-voice, such as noise and echo received from the audio signal converted into a digital signal from the A / D converter 12 It includes an extraction unit 13 and a speech recognition unit 14 for receiving the extracted voice from the speech extraction unit 13 to extract the speech features and to recognize the speech by a speech recognition algorithm. At this time, the speech extraction method according to the present invention is performed by the speech extraction unit 13.

The voice extractor 13 receives a voice signal converted into a digital signal from the A / D converter 12 and removes non-voice such as noise and echo. At this time, the sparking noise having instantaneous large energy satisfies the cumulative energy boundary condition because the unit frame energy is large, but has a limited frame size to distinguish it from the voice signal. In addition, the echo is similar to the speech and satisfies the frame size boundary condition, but since the energy accumulation value is distinguished from the speech signal, the echo is used to distinguish it from the speech signal. As such, since the voice extractor 13 removes the non-voice as much as possible, the load of the voice recognition unit 14 is reduced, and as a result, the voice processing performance can be improved.

2 is a flowchart illustrating a method for extracting speech by non-voice removal according to the present invention.

In the speech extraction method using non-voice removal according to the present invention, a process of monitoring an initial frame that can be expected to be largely speech, a process of determining speech and non-voice according to frame size boundary conditions and cumulative energy boundary conditions, and In the case where it is determined that the voice is determined, the process includes monitoring the frame and repeating the process.

First, an expected starting point is detected by receiving a voice from the outside and monitoring whether there is a first frame that can be predicted as a voice, that is, a frame satisfying an expected starting point determining condition (201). When the expected starting point is detected, it is determined whether the frame corresponding to the detected starting point satisfies the frame size boundary condition and the cumulative energy boundary condition according to the speech feature (202). As a result of the determination 202, if the voice characteristic is satisfied, the expected start point is determined as the start point, and the end point is detected to extract the voice signal (204). As a result of the determination 202, if the voice characteristic is not satisfied, the frame is determined to be non-voice, removed and initialized (205). That is, in the case of sparking noise, the cumulative energy boundary condition is satisfied but the frame size boundary condition is not satisfied, and in the case of echo, the frame size boundary condition is satisfied but the cumulative energy boundary condition is not satisfied. Therefore, it is determined that the voice feature does not satisfy the voice feature and is removed. After that, the process proceeds to “201” and the process for extracting the voice is resumed.

3a and 3b is a specific example of the voice extraction method by non-voice removal according to the present invention.

First, referring to the comparison conditions in each state, state 1 308 compares whether the frame counter (Env_Count) value of the audio signal data is smaller than the preset frame length (ENV_LEN) value. The state 2 310 compares whether the energy of the input speech unit frame is smaller than the lower energy limit ThreshEL set as a reference value. The state 3 314 and the state 4 316 compare whether the energy of the input voice unit frame is greater than the upper energy threshold ThreshEH set as a reference value. Finally, state 5 321 compares whether a zero crossing rate (ZCR) value of a voice unit frame is smaller than a preset zero crossing rate (ZCR) value. In this case, the zero crossing rate (ZCR) refers to the number of times that the signals of adjacent signals are different from each other. This value includes frequency information of the signal, and generally has a low value for voiced sound and a high value for unvoiced sound. In the following description, a specific embodiment of a voice extraction method by non-voice removal according to an embodiment of the present invention will be described.

First, the voice extracting unit 13 receives the voice data (301) and checks whether it is raw data (302), and if it is raw data, checks whether an expected starting point exists (303). As a result of the check 303, if the expected start point exists, it is determined whether the voice feature or the non-voice feature is satisfied, and if the voice feature is satisfied, the start point is determined, and if the non-voice feature is satisfied, the start point search is resumed (304 to 307). At this time, when the starting point does not satisfy both the voice feature and the non-voice feature, the reference value is initialized by setting a predetermined period as the noise feature section from the first frame of the voice data (308, 309).

Thereafter, the state 2 310 compares whether the unit frame energy of the voice signal data is lower than the lower energy threshold ThrEL set as a reference value, and if the unit frame energy is lower than the lower energy threshold ThrEL and the start point exists, the end point determination condition is satisfied. The end point is detected and voice signal extraction is completed (310 to 313).

On the other hand, if the unit frame energy of the voice signal data in state 2 310 is greater than or equal to the lower energy threshold ThreshEL, the upper energy limit (ThreshHL) in which the unit frame energy of the voice signal data is set to reference value in state 3 314 and state 316. ), And if the unit frame energy is equal to or higher than the upper energy limit, an expected starting point that satisfies the expected starting point determination condition is determined, and preparations are made to remove the non-voice section (316 to 320).

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.). Since this process can be easily implemented by those skilled in the art will not be described in more detail.

The present invention described above is capable of various substitutions, modifications, and changes without departing from the spirit of the present invention for those skilled in the art to which the present invention pertains. It is not limited by the drawings.

As described above, the present invention can reduce the load incurred in processing non-voice signals by providing a method of removing sparking noise and echoes having a large amount of energy in an instant, thereby improving speech recognition performance. It can be effected.

Claims (2)

In the speech extraction method applied to the speech extraction apparatus by non-voice removal, A first step of receiving and receiving a voice signal from the outside to monitor and detect an expected starting point estimated by the voice; A second step of determining whether a frame corresponding to the expected starting point satisfies a frame size boundary condition and a cumulative energy boundary condition according to a voice feature; A third step of determining that the frame is non-voice and removing if not satisfied as a result of the determination of the second step; And A fourth step of determining the expected start point as a start point and extracting a voice signal by detecting an end point if it is satisfied as a result of the determination of the second step; Speech extraction method by non-voice removal comprising a. In a speech extraction device having a processor, A first function of receiving a voice signal from the outside and monitoring and detecting an expected starting point estimated by the voice; A second function of determining whether a frame corresponding to the expected starting point satisfies a frame size boundary condition and a cumulative energy boundary condition according to a voice feature; A third function of determining and removing the corresponding frame as non-voice if it is not satisfied as a result of the determination in the second function; And A fourth function of determining the expected start point as a start point and detecting an end point and extracting a voice signal when the determination result of the second function is satisfied; A computer-readable recording medium having recorded thereon a program for realizing this.