KR100345402B1 - An apparatus and method for real - time speech detection using pitch information - Google Patents

Abstract

The present invention provides an apparatus and method for real-time detection of only voice from an input signal using pitch information, which is one of features of voice, in a system using voice as an input such as voice recognition and voice coding. There is a purpose.

According to the present invention, a voice input unit into which a voice signal is input, an A / D converter for converting an analog voice signal input by the voice input unit into a digital voice signal, and the A / D converter A voice detection device comprising a start point detector for detecting a start point of a voice in a converted digital voice signal, and an end point detector for detecting an end point of a voice in a digital voice signal converted by the A / D converter. Frame generation means for generating a frame having a predetermined length from the digital voice signal converted by the A / D converter; Threshold setting means for extracting samples from frames earlier in the frame generated by the frame generating means and determining a threshold value of the pitch using only energy values of the extracted samples; And a detecting means for detecting a starting point of a voice from frames of a voice signal which is subsequently input using the threshold values determined by the threshold setting means.

Description

An apparatus and method for real time speech detection using pitch information {An apparatus and method for real-time speech detection using pitch information}

The present invention relates to an apparatus and method for detecting only a voice of a signal input in real time in a system using voice as an input, such as voice recognition and voice coding, and in particular, pitch information, which is one of features of voice. The present invention relates to an apparatus and a method for detecting a voice part by accurately determining the presence or absence of voice even in a signal in which noise and voice coexist.

In general, the accurate separation of only pure speech parts from the speech signal from the ambient noise is very important in almost all areas of speech engineering such as speech recognition, speech coding and speech analysis. Recently, as speech recognition and coding techniques have been applied in real life, the importance of speech detection, which plays an important role in determining the performance of these functions, has been highlighted.

Traditionally, a method of detecting a speech part includes a method of using parameters in the time domain, a method of using parameters in the frequency domain, and a method of using the two domains together.

The method of using the time domain parameter is mainly a method of using short-term energy and zero crossing rate. In addition, a method of using the parameters of the frequency domain is a method of separating and using only a specific frequency region in which a human voice exists using a filter bank, and the method of using the parameters of the time and frequency domain together. First, only a specific frequency domain is separated, and then a time domain parameter is extracted and used.

These methods have their advantages and disadvantages. First, the method of using time-domain parameters has the advantage that the calculation is simple and easy to implement, while the strength of the input signal and the signal-to-noise ratio (Signal-To- The performance varies greatly depending on the size of the noise.

On the other hand, the method of using the parameters in the frequency domain and using the parameters in both the time and frequency domain is better than the case of using only the time domain parameters for the voice detection performance in a small SNR environment, but the calculation is complicated and time-consuming. However, due to the nature of the algorithm, there is a problem in that real-time implementation is difficult.

Therefore, the voice detection algorithm used in the current speech recognition system does not require real-time voice detection in the presence of noise depending on the characteristics of the system, and when accurate endpoint detection is required, the method of using the parameters in the frequency domain is relatively quiet. In case of real-time voice detection, the method of using time domain parameter is used.

As described above, in order to implement a voice detector operating in real time, a parameter in the time domain is mainly used. In this case, representative parameters are short-term energy and zero crossing rate. However, the zero crossing rate is such that the high frequency band is removed like a telephone line, and in the presence of channel noise, the characteristics of voice cannot be properly reflected. When the voice of 'ㅌ' and 'B', 'ㄹ' and 'ㅁ' are not detected, and the voice of lip and the tapping of the desk have a little energy. The case occurs.

In order to solve these problems, a method of detecting a start point and an end point of a voice using a pitch change rate and a normalized pitch change rate after determining the pitch threshold in the stable section of the voiced sound may be used. However, since this method finds the threshold of pitch in the stable section of voiced sound, it cannot be implemented in real time.

A real-time voice detection method using speech / non-voice discrimination that is generally used will be described as follows.

First, the threshold value to be used for voice detection is determined using the initial appropriate number of frames that are input. Thereafter, the input frames are determined as voice or non-voice frames, respectively, using the threshold value. The method of detecting the start and end points of the speech using the frames determined as voice / non-voice uses the number of voice / non-voice frames (N1, N2) existing in a certain number of consecutive frames (P1, P2). do. That is, in the case of the voice start point, there are N1 frames in the P1 frames, and in the case of the voice end point, the frames determined to be non-voice are P2 frames after the start point of the voice is detected. N2 points in the field are determined as the start point and the end point of the voice, respectively. Here, the lengths of P1, P2, N1, and N2 may vary according to algorithms and parameters used, and a method of determining voice portion detection may also vary.

The real-time voice detection method can greatly shorten the time of voice processing due to the characteristics of the real-time processing. For example, suppose that you use a system that receives and processes a voice that has a length of 5 seconds. If the system's processing speed is 1 times the length of the voice, if you do not use real-time voice detection, 5 seconds of voice After receiving all the input, it must be processed, so after 5 seconds of voice input, the result comes out. However, if the system uses the real-time voice detection method, the processing must be performed immediately after the start point of the voice is detected, so that the result can be obtained immediately after the voice input is terminated.

The present invention has been made to solve the problems of the prior art as described above, by simply obtaining the pitch information which is one of the characteristics of speech in the process of obtaining the energy threshold in the time domain, and additionally used in the speech section detection, irregular noise It is an object of the present invention to provide a voice detection device and a method for accurately detecting a voice section in real time from an existing input signal.

1 is a flowchart illustrating a real-time voice detection process according to an embodiment of the present invention.

2 is a flowchart illustrating a speech / non-voice discrimination process of a frame using a pitch threshold alpha according to an embodiment of the present invention.

According to the present invention for achieving the object as described above, an audio input unit for inputting a voice signal, A / D conversion unit for converting an analog audio signal input by the voice input unit into a digital voice signal And an end point detector for detecting a start point of the voice in the digital voice signal converted by the A / D converter, and an end point detector for detecting an end point of the voice in the digital voice signal converted by the A / D converter. An apparatus for detecting speech, comprising: frame generating means for generating a frame having a predetermined length from a digital speech signal converted by the A / D converter; Threshold setting means for extracting samples from frames earlier in the frame generated by the frame generating means and determining a threshold value of the pitch using only energy values of the extracted samples; And a detecting means for detecting a starting point of a voice from frames of a voice signal which is subsequently input using the threshold values determined by the threshold setting means.

Also, a voice input step of inputting a voice signal, an A / D conversion step of converting an analog voice signal input in the voice input step into a digital voice signal, and a digital signal converted in the A / D conversion step In the speech detection method comprising a starting point detecting step of detecting the starting point of the speech from the speech signal, detecting the end point of the speech from the digital speech signal converted in the A / D conversion step, the starting point detecting step A first step of generating a frame having a predetermined length from the digital voice signal converted in the A / D conversion step; A second step of extracting samples from frames earlier in the frame generated in the first step and determining a threshold of a pitch using only energy values of the extracted samples; And a third step of detecting a starting point of a voice from frames of a voice signal which is subsequently input using the threshold values determined in the second step.

In addition, a first step of inputting a voice signal to the computer; A second step of converting the analog voice signal input in the first step into a digital voice signal; A third step of generating a frame having a predetermined length from the digital voice signal converted in the second step; After extracting the samples from the previous frame among the frames generated in the third step, the frame energy average value of the extracted samples is divided by the number of extracted samples, and the value obtained by multiplying a predetermined weight is obtained. A fourth step of determining a threshold; And a fifth step of determining the frame as a voice frame when the number of samples exceeding the threshold value of the pitch in one frame determined in the fourth step is n or more predetermined. Provided is a computer readable recording medium having recorded thereon.

The present invention further improves the accuracy of real-time speech detection by using pitch information, which is one of the features of speech, in order to detect a speech portion that cannot be detected accurately with only zero crossing rate and energy in an irregularly generated input signal. Let's do it. This method was devised based on the following experimental phonetic knowledge.

First, most human voice signals have power within 1 KHz, which is the low frequency range.

Second, pitch, which is characteristic of voice, especially voiced sounds, is produced by the tremor of the vocal cords.

Third, the most common error in the voice detection method using the short section energy as a threshold occurs because the length of the voiced sound section having an energy value above the threshold is short.

Fourth, the pitch is necessarily present in the voiced sound section having a small energy value.

This will be described in detail as follows.

Human voice signals are mostly concentrated within 1 KHz, the low frequency range, and the pitch, which is characteristic of voiced sounds, exists between 100 and 250 Hz on average, including both males and females. Such pitch information is an important feature to distinguish voiced and unvoiced voices and voices and non-voices.

100 Hz represents a period of 1/100 second, 250 Hz represents a period of 1/250 seconds, which is expressed in msec, 10 ms to 4 ms. That is, since the length of a frame used for speech recognition usually uses about 10 ms to 20 ms, this means that at least 1 to 5 pitches exist in one frame.

Expressed again as a sample, since a sampling frequency of about 8 kHz to about 16 kHz is generally used for speech recognition, one frame includes 80 to 320 samples. This means that the number of samples forming one pitch is 16 to 320 samples. The values of adjacent samples forming one pitch have a characteristic of not showing a sudden change.

In the process of calculating the frame energy within one frame using this characteristic, the existence of pitch is determined by using the energy values of several adjacent samples without any additional calculation, and in parallel with other criteria such as energy or zero crossing rate. Used for voice / non-voice judgment.

In order to achieve the above object, the present invention has a structure in which a voice / non-voice determination process of a frame using pitch is added to a real-time voice section detection method using voice / non-voice discrimination for each frame.

In the following, with reference to the accompanying drawings, a preferred embodiment according to the present invention will be described in detail.

1 is a flowchart illustrating a real-time voice detection process according to an embodiment of the present invention.

First, in step S100, when a signal is input to an input device such as a microphone, in step S101, the signal is converted from an analog signal to a digital signal by an A / D converter, and then in step S102, a unit of speech processing is suitable. To form a frame having a length. Subsequent steps are performed frame by frame.

Subsequently, in step S103, it is determined whether or not the threshold values used for voice detection are set using the frames in front of the input signal. If not, in step S104, the threshold values are set, and then the step is set. Returning to S100 and if the setting is made, in step S105, after determining the voice / non-voice for each frame, in step S106, whether the starting point of the voice has been detected from the frames of the subsequent input signal using the determined threshold values. Judge. In this case, in general, the thresholds used for the voice detection in the real-time voice detection method are obtained in the initial frames of the input signal, which is made on the assumption that there is silence in front of the input signal.

If the result of the determination in step S106 is not detected, after detecting the start point of the voice in step S107, the process returns to the step S102, and if it is detected, it is determined in step S108 whether the end point of the voice has been detected. do.

As a result of the determination in step S108, if no detection is made, the flow returns to step S102. If detection is made, in step S109, after detecting the end point of speech, the input of the speech is automatically stopped.

In this process, the step of applying the present invention is the step S104 of determining a threshold value for voice detection, and the step S105 of determining an input frame as voice / non-voice using the determined threshold value.

Looking at the case in which the present invention is applied in the above two steps as follows.

First, the threshold using the pitch used in the present invention is referred to as 'alpha'. The alpha value is obtained by dividing the average energy of the frames used to determine the threshold by the number of samples constituting a frame. The value obtained by multiplying the average energy value of the samples in the frame by the weight w is alpha, which can be formulated as in Equation 1 below. The weight w is a value determined by repeated experiments.

2 is a flowchart illustrating a voice / non-voice discrimination process of a frame using a pitch threshold alpha according to an embodiment of the present invention.

As a method of determining whether a subsequent input frame is a voice frame or a non-voice frame using alpha calculated above, all samples existing in one frame are sequentially compared with an alpha value.

First, in step S201, the counter variable Counter is set to 0, and the variable Cur_S indicating the number of samples currently being worked is set to 0, and in step S202, when the sample value is input, in step S203, the Cur_S After performing the operation of = Cur_s + 1, in step S204, it is determined whether the variable Cur_S representing the total number of samples in the current frame is greater than Total_S.

As a result of the determination in step S204, if Cur_S is greater than Total_S, in step S210, the current frame is judged to be a non-voice frame, and then, if Total_S is greater than Cur_S, if in step S205 the sample value is greater than the alpha already calculated Determine whether or not.

As a result of the determination in step S205, if the sample value is less than or equal to alpha, the counter is reset to 0 again in step S206, and then returned to step S202. If the sample value is greater than alpha, the counter is returned to step S207. After increasing by one, in step S208, it is determined whether or not the counter is larger than n preset.

As a result of the determination in step S208, if the counter is larger than n, the current frame is determined to be an audio frame and then terminated.

The n value depends on the sampling frequency used for A / D and is usually set between 3 and 8. That is, if n is set to 5, when five or more sample values exceeding the alpha value exist in one frame, the frame is determined as a voice frame.

Speech / non-speech judgment of a frame using alpha is used as an important data for determining the starting point of speech along with judgment using other thresholds used together.

The present invention as described above is recorded on a computer-readable recording medium, and can be processed by a computer.

As described in detail above, the present invention provides a method for detecting a speech section in real time using the speech / non-voice determination result for each frame, and adds pitch information existing in the speech section as a threshold for detecting the speech section with a very small amount of calculation. By using it, it is possible to accurately detect only a voice section from an input signal such as irregularly generated non-voices such as a hard disk drive noise of a notebook and a speaker's lips.

The technical spirit of the present invention has been described above with reference to the accompanying drawings, but this is by way of example only and not by way of limitation to the present invention. In addition, it is obvious that any person skilled in the art may make various modifications and imitations without departing from the scope of the technical idea of the present invention.

Claims (10)

A voice input unit to which a voice signal is input, an A / D converter for converting an analog voice signal input by the voice input unit into a digital voice signal, and a digital voice signal converted by the A / D converter In the speech detection device comprising a start point detection unit for detecting the start point of the voice at the end, and an end point detection unit for detecting the end point of the voice in the digital voice signal converted by the A / D converter, The starting point detector, Frame generation means for generating a frame having a predetermined length from the digital voice signal converted by the A / D converter; Threshold setting means for extracting samples from frames earlier in the frame generated by the frame generating means and determining a threshold value of the pitch using only energy values of the extracted samples; Detection means for detecting the frame as a voice start point if the number of samples in one frame exceeding the threshold values determined by the threshold value setting means is n or more predetermined; Voice detection device comprising a. The method of claim 1, The threshold setting means, An apparatus for detecting speech, comprising: determining a threshold alpha of a pitch using average energy of samples in a frame having a predetermined length as shown in [Equation 1] below. [Equation 1] Where w is a predetermined value by repeated experiments. delete The method according to claim 1 or 2, The n value is set to an integer of any one of 3 to 8. A voice input step of inputting a voice signal, an A / D conversion step of converting an analog voice signal input in the voice input step into a digital voice signal, and a digital voice signal converted in the A / D conversion step A voice detection method comprising: a start point detection step of detecting a start point of a voice from an end point; and an end point detection step of detecting an end point of a voice from a digital voice signal converted in the A / D conversion step. The starting point detection step, A first step of generating a frame having a predetermined length from the digital voice signal converted in the A / D conversion step; A second step of extracting samples from frames earlier in the frame generated in the first step and determining a threshold of a pitch using only energy values of the extracted samples; A third step of detecting the frame as a voice start point if the number of samples in one frame exceeding the threshold values determined in the second step is n or more predetermined; Voice detection method comprising a. The method of claim 5, The second step, A voice detection method characterized in that the threshold value of the pitch alpha is determined using the average energy of the samples in the frame of a predetermined length as shown in [Equation 2] below. [Equation 2] Where w is a predetermined value by repeated experiments. delete The method according to claim 5 or 6, The n value is set to an integer of any one of 3 to 8. On your computer, A first step of inputting a voice signal; A second step of converting the analog voice signal input in the first step into a digital voice signal; A third step of generating a frame having a predetermined length from the digital voice signal converted in the second step; A fourth step of extracting samples from the frame of the front part of the frame generated in the third step, and then determining a threshold of the pitch according to Equation 3 below; And a fifth step of determining the frame as a voice frame when the number of samples exceeding the threshold value of the pitch in one frame determined in the fourth step is n or more predetermined. A computer-readable recording medium that has been recorded. [Equation 3] Where w is a predetermined value by repeated experiments. The method of claim 9, The fifth step, And n is set to an integer of any one of 3 to 8. A computer-readable recording medium having recorded thereon a program.