KR100334238B1 - Apparatus and method for detecting speech/non-speech using the envelope of speech waveform - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

The present invention relates to a speech / non-voice discrimination apparatus and method using envelope information of speech waveforms.

2. The technical problem to be solved by the invention

The present invention provides a voice / non-voice discrimination apparatus for determining whether an input signal is voice / non-voice using frequency characteristics of a voice envelope, and a method and a computer-readable recording medium recording a program for realizing the method. Its purpose is to.

3. Summary of Solution to Invention

The present invention provides a speech / non-voice discrimination apparatus using envelope information of speech waveforms, comprising: storage means for storing an input speech signal; Envelope detection means for extracting an envelope with respect to the signal stored in said storage means; Frequency converting means for converting the envelope signal into a frequency domain; And voice / non-voice discrimination means for determining whether voice / non-voice is performed according to the ratio of the total frequency domain energy and the specific frequency domain energy of the envelope signal converted into the frequency domain by the frequency conversion means.

4. Important uses of the invention

The present invention is used in the field of speech processing.

Description

Apparatus and method for detecting speech / non-speech using the envelope of speech waveform}

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech / non-voice discrimination apparatus and method for determining whether a signal detected as speech is a real speech in a system using speech as an input such as speech recognition and speech coding. A voice / non-voice discrimination apparatus for accurately determining whether a voice or a non-voice is used even in a signal in which noise and voice coexist using frequency information, and a method and a computer-readable recording medium recording a program for implementing the method. will be.

In general, one of the main factors causing the malfunction of the speech recognizer in a real environment is the malfunction of the speech detector. If noise is falsely detected as voice, it is inevitably forced to output a misrecognition result. When the recognition result is output even though the voice is not input, the users of the voice recognition system have no doubt about the reliability of the system. Thus, the task of determining whether the voice / non-voice is to the signal detected as voice in the voice detector can be of considerable importance depending on the application of the system.

Conventional methods for determining speech / non-voice include a method of making a voice model and a non-voice model having characteristics in a frequency domain and using which model the input signal is close to and a method of determining using pitch information in a time domain. have. That is, conventional speech / non-voice discrimination apparatuses make a model of noise and speech separately to determine whether the probability of which model is high or to determine the pitch when the voiced sound interval is satisfied to satisfy a given condition. Have been drunk. However, since these methods must first extract relevant information every frame, there is a problem in that a large amount of computation is required.

In addition, the method using the conventional model has a problem in that the performance of the system is significantly reduced when the environment in which the actual system operates is severely different from the environment in which the model is made. In other words, the method of using noise and speech models has to be made in advance, and it is difficult to guarantee stable performance when there are many differences between the environment when the model is created and when the actual discrimination system operates. have.

In addition, in the method of using the pitch-related information, the problem is how insensitive and stable the conditions used for the extraction of the pitch and the voice judgment. In other words, the method of using the pitch is difficult to pull out the pitch correctly in a real environment.

SUMMARY OF THE INVENTION The present invention has been proposed to solve the above-mentioned problems. The present invention provides a speech / non-voice discrimination apparatus and a method for determining the speech / non-voice of an input signal using frequency characteristics of a speech envelope and a program for realizing the method. Its purpose is to provide a computer-readable recording medium having recorded thereon.

That is, the present invention down-samples the signal extracted by the envelope detector to reduce the calculation amount, and uses the energy ratio of the specific frequency domain to the energy of the entire frequency domain with respect to the frequency-converted envelope. It is an object of the present invention to provide a voice / non-voice discrimination apparatus for determining non-voice and a method thereof and a computer-readable recording medium recording a program for realizing the method.

1 is an exemplary view of a speech recognition system to which a speech / non-voice discrimination apparatus according to the present invention is applied.

2 is a block diagram of an embodiment of a speech / non-voice discrimination apparatus using envelope information of a speech waveform according to the present invention;

3 is a flowchart illustrating a method for determining a voice / non-voice using envelope information of a voice waveform according to the present invention;

* Explanation of symbols for the main parts of the drawings

11 microphone 12 voice detector

13: recognizer 14, 22: envelope detector

15: envelope analyzer 21: buffer

23: down-sampling 24: frequency converter

25: voice / non-voice discrimination unit

An apparatus of the present invention for achieving the above object comprises: a speech / non-voice discrimination apparatus using envelope information of speech waveforms, comprising: storage means for storing an input speech signal; Envelope detection means for extracting an envelope with respect to the signal stored in said storage means; Frequency converting means for converting the envelope signal into a frequency domain; And voice / non-voice discrimination means for determining whether voice / non-voice is performed according to a ratio of total frequency domain energy and specific frequency domain energy of the envelope signal converted into the frequency domain by the frequency conversion means.

In addition, the apparatus of the present invention, by reducing the number of samples of the envelope by downsampling the envelope signal extracted by the envelope detection means, in order to reduce the amount of calculation in the frequency conversion means It further comprises a down sampling means for transferring to the conversion means.

On the other hand, the method of the present invention, the speech / non-voice discrimination method applied to the speech / non-voice discrimination apparatus using the envelope information of the speech waveform, comprising: a first step of storing a speech input signal; Extracting an envelope of the stored input signal; Converting the envelope signal into a frequency domain; And determining a speech / non-voice based on a ratio of total frequency domain energy and specific frequency domain energy of the envelope signal converted into the frequency domain.

In addition, the method of the present invention, downsampling the extracted envelope signal in order to reduce the amount of calculation when converting the envelope signal of the time domain to the frequency domain in the third step of the envelope of the envelope; The method may further include a fifth step of moving to the third step after reducing the number of samples.

On the other hand, the present invention, in order to determine whether the voice / non-voice using the envelope information of the audio waveform, a voice / non-voice determination device having a processor, the first function of storing a voice input signal; A second function of extracting an envelope with respect to the stored input signal; A third function of converting the envelope signal into a frequency domain; And a fourth program for realizing a fourth function for determining voice / non-voice according to a ratio of total frequency domain energy and specific frequency domain energy of the envelope signal converted into the frequency domain. do.

In addition, the present invention, in order to reduce the amount of calculation when converting the envelope signal in the time domain to the frequency domain in the third function, by performing downsampling on the extracted envelope signal to sample the envelope (sample) Provided is a computer-readable recording medium having recorded thereon a program for further realizing the fifth function, which reduces the number and then transitions to the third function.

Briefly looking at the gist of the present invention.

Generally speaking, a good voice system that should operate in a real environment should be excellent in terms of stability as well as stability and insensitivity to the surrounding environment, and should be as simple as possible. In order to satisfy this problem, the present invention proposes a speech / non-voice discrimination apparatus and method using speech envelope information.

That is, there is always a voiced sound section in the voice, and generally longer than the voiced sound section. Since the envelope of speech formed by the energy change of the voiced sound section or the repetition of the voiced sound and the unvoiced sound is due to the physical motion of the articulator, the energy is concentrated in the limited frequency region. And, even if the noise envelope is mixed with moderate noise, its shape does not change significantly, and this characteristic has great strength in terms of insensitivity to the surrounding environment, and also related to the phoneme length. The energy is concentrated.

Therefore, the present invention makes it possible to accurately discriminate voice / non-voice even in a noisy input signal by using a small calculation amount and a simple algorithm by using this feature of the voice envelope.

In conclusion, most of the envelopes of the human voice signal are energized in a specific frequency region, and vary depending on the speed of speech and the types of spoken phonemes. Therefore, in the present invention, after calculating the envelope of the signal received by the voice detector using this characteristic, it is converted to the frequency domain, and then the energy distribution is examined to determine whether the voice or non-voice.

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 view of a speech recognition system to which a speech / non-voice discrimination apparatus according to the present invention is applied.

First, a signal input by an input device such as the microphone 11 passes through the voice detector 12. The voice detector 12 here is a conventional general voice detector. The voice detector 12 transmits the input signal from the microphone 11 or the like to the next blocks from the time when it is determined to be voice, but does not transmit the input signal to the next block when the voice is finished. The signal passing through the voice detector 12 is input to the recognizer 13 and the envelope detector 14. The recognizer 13 is a general recognizer including a feature vector extractor.

The envelope detector 14 stores a signal input from the voice detector 12, and then detects an envelope from the stored signal. Then, the detected envelope is passed to the envelope analyzer 15, and the envelope analyzer 15 analyzes the input envelope to determine whether or not it is negative. The result of the determination is passed to the voice detector 12 and the recognizer 13. The envelope detector 14 and the envelope analyzer 15 operate independently of the recognizer 13.

When the voice detector 12 receives the determination result from the envelope analyzer 15 that the signal input from the microphone 11 is not voice, the voice detector 12 does not pass the signal from the microphone 11 to the next block and detects the beginning of the voice. Switch to the initial mode for.

When the recognizer 13 is in a recognition ready state and receives a signal determined to be a voice from the voice detector 12, the recognizer 13 recognizes the signal after performing feature vector extraction and outputs a recognition result. During the recognition, upon receiving the determination result from the envelope analyzer 15 that the signal input from the voice detector 12 is not voice, the recognition process is immediately stopped and the initial recognition ready state is entered again.

2 is a block diagram of an embodiment of a speech / non-voice discrimination apparatus using envelope information of a speech waveform according to the present invention.

Voice / non-voice discrimination apparatus using the envelope information of the voice according to the present invention, the buffer 21 for temporarily storing the input signal detected by the voice detector 12, with respect to the input signal stored in the buffer 21 Envelope detector 22 for extracting the envelope, down sampling to reduce the number of samples of the envelope by performing downsampling on the envelope signal extracted by the envelope detector 22 to reduce the amount of calculation in frequency conversion A sampling unit 23, a frequency converter 24 for converting the downsampled envelope signal down-sampled into the frequency domain, and a frequency converter 24 converted into the frequency domain. For an envelope signal, a speech / non-speech plate for judging speech / non-voice using the ratio of the energy of a specific frequency domain to the energy of the entire frequency domain of the envelope. And a unit (25).

Looking at the specific operation with reference to FIG.

3 is a flowchart illustrating a method for determining a voice / non-voice using envelope information of a voice waveform according to the present invention.

First, signals input from the voice detector 12 are temporarily stored in the buffer 21 for a predetermined length (preferably 1 to 10 seconds) (31, 32). The envelope detector 22 then extracts the envelope 33 with respect to the input signal stored in the buffer 21 (33).

Next, downsampling is performed on the extracted envelope signal in order to reduce the amount of calculation in converting the envelope signal in the time domain into the frequency domain, thereby reducing the number of samples of the envelope (34). Thereafter, the downsampled envelope signal is converted into the frequency domain (35).

Next, it is determined whether the envelope signal converted into the frequency domain is speech / non-voice using Equation 1 below (36). In other words, in the case of a negative envelope, it will have a high energy in a specific frequency region, so when the energy ratio in a specific frequency region with respect to the energy of the entire frequency region of the envelope is alpha (α), Therefore, it is determined whether the voice or non-voice.

α = (energy from frequency f1 to f2 in the frequency-converted envelope) / total energy of the frequency-converted envelope

According to Equation (1), the energy ratio α in a specific frequency region with respect to the total frequency domain energy of the envelope signal is determined, and if the value is greater than or equal to the threshold, it is determined to be negative (37). 38).

Thereafter, the determination result is sent to the recognizer 13 and the voice detector 12. When the determination result is non-voice, the voice detector 12 and the recognizer 13 are switched to the initial mode.

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

The present invention as described above, by using the frequency characteristics of the speech envelope can accurately determine whether the input signal speech / non-voice, and due to the characteristics of the simple algorithm and the speech envelope affects the environment in which the actual discriminating device operates There is an effect of ensuring a stable performance less received.

Claims (8)

In a speech / non-voice discrimination apparatus using envelope information of an audio waveform, Storage means for storing an input speech signal; Envelope detection means for extracting an envelope with respect to the signal stored in said storage means; Frequency converting means for converting the envelope signal into a frequency domain; And Speech / non-voice discrimination means for determining whether speech or non-voice according to the ratio of the total frequency domain energy and the specific frequency domain energy of the envelope signal converted into the frequency domain by the frequency conversion means Speech / non-voice discrimination apparatus using the envelope information of the speech waveform comprising a. The method of claim 1, In order to reduce the amount of calculation in the frequency converting means, downsampling is performed on the envelope signal extracted by the envelope detecting means to reduce the number of samples of the envelope and deliver the downsampled means to the frequency converting means. Speech / non-voice discrimination apparatus using the envelope information of the speech waveform further comprising. The method according to claim 1 or 2, The voice / non-voice discrimination means, The ratio of the total frequency domain energy and the specific frequency domain energy of the envelope signal is determined. If the ratio is greater than or equal to the threshold, the voice is judged as voice, and if it is less than or equal to the threshold, the voice / ratio using the envelope information of the voice waveform is determined. Voice discrimination device. In a speech / non-voice discrimination method applied to a speech / non-voice discrimination apparatus using envelope information of a speech waveform, A first step of storing a voice input signal; Extracting an envelope of the stored input signal; Converting the envelope signal into a frequency domain; And A fourth step of determining whether voice / non-voice is performed according to a ratio of total frequency domain energy and specific frequency domain energy of the envelope signal converted into the frequency domain; Speech / non-voice discrimination method using the envelope information of the speech waveform comprising a. The method of claim 4, wherein In order to reduce the amount of calculation when converting the envelope signal in the time domain to the frequency domain in the third step, downsampling is performed on the extracted envelope signal to reduce the number of samples of the envelope. 5th step to 3rd step Speech / non-voice discrimination method using the envelope information of the speech waveform further comprising. The method according to claim 4 or 5, The fourth step, The ratio of the total frequency domain energy and the specific frequency domain energy of the envelope signal is determined. If the ratio is greater than or equal to the threshold, the voice is judged as voice, and if it is less than or equal to the threshold, the voice / ratio using the envelope information of the voice waveform is determined. Voice discrimination method. In order to determine whether speech or non-voice using the envelope information of the speech waveform, to a speech / non- speech discrimination apparatus having a processor, A first function of storing a voice input signal; A second function of extracting an envelope with respect to the stored input signal; A third function of converting the envelope signal into a frequency domain; And A fourth function of determining whether voice / non-voice is based on a ratio of total frequency domain energy and specific frequency domain energy of the envelope signal converted into the frequency domain A computer-readable recording medium having recorded thereon a program for realizing this. The method of claim 7, wherein In order to reduce the amount of calculation when converting the envelope signal in the time domain to the frequency domain in the third function, down sampling is performed on the extracted envelope signal to reduce the number of samples of the envelope. Fifth function to pass to third function A computer-readable recording medium that records a program for further realization.