JP2007292940A - Voice recognition device and voice recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device capable of accurately and rapidly recognizing whether sound indicated by a sound signal is voiceless sound or voiced sound. <P>SOLUTION: The voice recognition device 10 of the invention is provided with: a generating section 13 in which the sound signal along a time axis is segmented into a plurality of frames by shifting each segmentation start time; a calculation section 15 for calculating the number of zero crossing corresponding to each of the plurality of segmented frames; and a recognition section 17 in which it is recognized, based on the calculated number of zero crossing, whether a phoneme corresponding to each of the plurality of frames is the voiceless sound or the voiced sound, based on the result that the number of zero crossing for each of the plurality of frames is compared with the number of zero crossing in an adjoining frame to each of the plurality of frames. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a voice identification device and a voice identification method.

There is a voice identification technique for identifying whether a sound indicated by an input voice signal is an unvoiced sound or a voiced sound. Speech identification technology is used for speech recognition or speech synthesis. As such a voice identification technique, the following Patent Document 1 discloses that a voice signal in a predetermined section indicates an unvoiced sound or a voiced sound based on a zero-cross number of a voice signal in a predetermined section and a preset zero-cross number threshold. Techniques for indicating or identifying are described.
JP 2003-256000 A

  However, the number of zero crosses of the audio signal in the predetermined section indicating unvoiced sound may be larger or smaller than the number of zero crosses of the audio signal in the predetermined section indicating voiced sound. Therefore, when using the zero-cross number of the audio signal in a predetermined section, it is difficult to set an appropriate threshold value for identifying unvoiced sound and voiced sound.

  Also, a modified correlation method is known as a speech identification method. In the modified correlation method, an autocorrelation function is obtained for a waveform of an input speech signal, and then linear prediction analysis is performed to obtain linear prediction coefficients corresponding to orders. Thereafter, in the modified correlation method, the prediction residual is obtained, the peak of the prediction residual is detected, the time distance between the peaks is measured, and the voice is identified based on the measurement result.

  Since the modified correlation method requires such a complicated operation, the voice identification speed is slow. Therefore, the modified correlation method cannot cope with speech recognition and speech synthesis used in a real-time dialogue system.

  Therefore, an object of the present invention is to provide a speech identification device and a speech identification method that can accurately and quickly identify whether a phoneme indicated by a speech signal is an unvoiced sound or a voiced sound.

  As a result of research, the present inventors have found that an unvoiced sound signal has less periodicity than a voiced sound signal and has a large variation in the number of zero crossings. The present inventors recalled the present invention by advancing research focusing on the variation in the number of zero crossings.

  The voice identification device according to the present invention calculates a frame generation unit that cuts a voice signal along a time axis into a plurality of frames by shifting each cut-out start time, and calculates the number of zero crosses corresponding to each of the cut out frames. Based on the calculated zero-cross number, a zero-cross comparison unit that compares the zero-cross number in each of the plurality of frames with a zero-cross number in a frame adjacent to each of the plurality of frames, based on the calculated zero-cross number, Phoneme identifying means for identifying whether a phoneme corresponding to each of a plurality of frames is an unvoiced sound or a voiced sound.

  According to the speech identification method of the present invention, the frame generation unit extracts the audio signal along the time axis into a plurality of frames by shifting the extraction start time, and the zero-cross calculation unit is extracted in the first step. The second step of calculating the number of zero crosses corresponding to each of the plurality of frames, and the zero cross comparison means, based on the number of zero crosses calculated in the second step, closes the number of zero crosses in each of the plurality of frames to each of the plurality of frames. A third step of comparing with the number of zero crosses in the frame to be performed, a fourth step of identifying whether the phoneme corresponding to each of the plurality of frames is an unvoiced sound or a voiced sound based on the comparison result in the third step; Is provided.

  According to the present invention, audio signals are cut out into frames, and whether the phonemes corresponding to the plurality of frames are unvoiced sounds or voiced sounds is identified based on the result of comparing the number of zero crosses between adjacent frames. Therefore, the number of zero crosses of each frame can be accurately identified as compared with the case where the threshold is compared, and the number can be quickly identified as compared with the modified correlation method.

  Further, in the speech identification device of the present invention, the phoneme identification means, when the difference value between the number of zero crosses in each of a plurality of frames and the number of zero crosses in a frame adjacent to each of the plurality of frames exceeds a predetermined threshold, It is also preferable to identify that the phoneme corresponding to the rear frame is an unvoiced sound.

  In the speech identification method of the present invention, in the fourth step, the phoneme identifying means sets a threshold value in which a difference value between the number of zero crosses in each of a plurality of frames and the number of zero crosses in a frame adjacent to each of the plurality of frames is set in advance. When exceeding, it is also preferable to identify that the phoneme corresponding to the rear frame is an unvoiced sound.

  Further, in the speech identification device of the present invention, the phoneme identification unit is configured to perform a backward operation when a difference value between the number of zero crosses in each of a plurality of frames and the number of zero crosses in a frame adjacent to each of the plurality of frames is smaller than a predetermined threshold. It is also preferable to identify that the phoneme corresponding to this frame is a voiced sound.

  In the speech identification method of the present invention, in the fourth step, the phoneme identifying means determines that a difference value between the number of zero crosses in each of a plurality of frames and the number of zero crosses in a frame adjacent to each of the plurality of frames is based on a predetermined threshold. It is also preferable to identify that the phoneme corresponding to the rear frame is a voiced sound when it is small.

  According to this preferable aspect, the number of zero crosses between adjacent frames can be compared more accurately.

  According to the present invention, since it is identified whether it is unvoiced sound or voiced sound based on the variation in the number of zero crosses of adjacent frames, it is possible to accurately and quickly identify whether the sound indicated by the sound signal is unvoiced sound or voiced sound.

  The knowledge of the present invention can be easily understood by considering the following detailed description with reference to the accompanying drawings shown for illustration only. Subsequently, embodiments of the present invention will be described with reference to the accompanying drawings. Where possible, the same parts are denoted by the same reference numerals, and redundant description is omitted.

  A voice identification system according to an embodiment of the present invention will be described with reference to FIG. FIG. 1 is a configuration diagram of a voice identification system according to the present embodiment. The voice identification system 1 according to the present embodiment is a system that includes a microphone 3 and a voice identification device 10 and identifies whether each sound constituting a word spoken by a person is voiced sound or unvoiced sound.

  The microphone 3 is a condenser microphone, samples vibrations of sounds spoken by a person, converts them into voice signals, and outputs them to the voice identification device 10. The voice identification device 10 is a device that identifies whether each sound constituting a word spoken by a person is voiced sound or unvoiced sound based on a voice signal input from the microphone 3.

  The voice identification device 10 according to the embodiment of the present invention will be described in more detail. The voice identification device 10 is a personal computer including a CPU, a memory, a power source, and an input / output interface unit as physical components. As shown in FIG. 2, the voice identification device 10 includes, as functional components, a detection unit 11, a generation unit (frame generation unit) 13, and a calculation unit (zero cross calculation unit) as illustrated in FIG. 2. ) 15 and an identification unit (zero cross comparison means, phoneme identification means) 17. FIG. 2 is a functional block diagram of the voice identification device according to the present embodiment. Subsequently, each functional component will be described.

  As shown in FIG. 3, the detection unit 11 is a part that detects an utterance section from the input voice signal S <b> 1. FIG. 3 is a graph showing a waveform of a voice signal input by the voice identification device according to the present embodiment. The detection unit 11 receives the audio signal S1 and detects the utterance interval by performing two-stage latching on both the beginning and the end of the utterance interval with respect to the amplitude of the audio signal.

  That is, in the waveform of FIG. 3, when the amplitude excitation occurs within a certain time from the point A where the amplitude rises (point B), the detection unit 11 sets the point A where the amplitude rises as the beginning of the utterance section. . In addition, the detection unit 11 sets the point D at which the amplitude has fallen as the ending of the utterance section when there is no excitation for 300 ms or more from the point D at which the amplitude has fallen. In the waveform shown in FIG. 3, the detection unit 11 does not recognize the point C at which the amplitude has fallen as the end of the utterance section because the amplitude is excited within 300 ms from the point C at which the amplitude has fallen.

  After detecting the utterance section, the detection unit 11 performs pre-emphasis on the detected speech signal S2 in the utterance section, as shown in FIG. 4, and generates the speech signal S3. FIG. 4A is a graph showing the waveform of the audio signal S2 before pre-emphasis. FIG. 4B is a graph showing the waveform of the audio signal S3 after pre-emphasis. Pre-emphasis is to emphasize the high frequency range of the audio signal. The amplitude of the audio signal S3 after pre-emphasis is larger than the amplitude of the audio signal S2 before pre-emphasis.

上式（１）において、ＨＳＲ［］はプリエンファシス前の音声信号を示し、ＨＰｒＳＲ［］はプリエンファシス後の音声信号を示す。 Specifically, the detection unit 11 performs pre-emphasis of the audio signal S2 by the following equation (1).

In the above equation (1), HSR [] indicates an audio signal before pre-emphasis, and HPrSR [] indicates an audio signal after pre-emphasis.

  FIG. 5A is a graph showing the spectrum of the audio signal S2 before pre-emphasis. FIG. 5B is a graph showing the spectrum of the audio signal S3 after pre-emphasis. As shown in FIG. 5, a higher frequency signal is amplified by pre-emphasis. The powers “la” and “lb” at specific frequencies in the audio signal before pre-emphasis are emphasized to become larger powers “la ′” and “lb ′”. The detection unit 11 outputs the audio signal S3 subjected to pre-emphasis to the generation unit 13.

  The generation unit 13 is a part that generates a plurality of frames using the audio signal S3. As illustrated in FIG. 6, the generation unit 13 generates a plurality of frames by multiplying the audio signal S3 by 30% for each 512 sampling while multiplying the function H (Humming window). FIG. 6 is a graph showing the waveform and function H of the audio signal after the shift.

上式（２）において、Ｐｒｅｃｉｓｅはサンプリング数（５１２）を示し、ｐは次数を示す。生成部１３は、音声信号Ｓ３の時系列順に、フレームを生成すると共に、時系列順にフレームにフレームナンバーを割り当てる。すなわち、フレームナンバーＮ（Ｎは正の整数）のフレームは、フレームナンバーＮ−１のフレームにおける後方部分の７０％ほどを含み、音声信号Ｓ３においてフレームナンバーＮ−１のフレームに後続する部分に対応する信号を３０％ほど含む。生成部１３は、生成した複数のフレームを算出部１５へ出力する。 Specifically, the generation unit 13 generates a frame by the following expression (2).

In the above formula (2), Precise indicates the sampling number (512), and p indicates the order. The generation unit 13 generates frames in the time sequence of the audio signal S3 and assigns frame numbers to the frames in the time sequence. That is, the frame with the frame number N (N is a positive integer) includes about 70% of the rear portion of the frame with the frame number N-1, and corresponds to the portion following the frame with the frame number N-1 in the audio signal S3. About 30% of the signals to be transmitted. The generation unit 13 outputs the generated plurality of frames to the calculation unit 15.

The calculation unit 15 is a part that calculates the number of zero crosses corresponding to each frame. The calculation unit 15 removes noise of each frame and calculates the number of zero crosses corresponding to each frame. The number of zero crossings is the number of intersections between the waveform of the signal from which noise is removed and the horizontal axis of the graph. First, the calculation unit 15 calculates an autocorrelation function shown in FIG. 7A for each frame. FIG. 7A shows a frame autocorrelation function. Specifically, the calculation unit 15 obtains the following equation (3) when given the steady time series data {x (t) | t = 0,..., N−1} where XtXt−τ + 1 average is 0. To calculate the autocorrelation function.

  Further, the calculation unit 15 calculates the average autocorrelation function shown in FIG. 7B by moving and averaging the calculated autocorrelation functions in a triple sequence. FIG. 7B shows an average autocorrelation function obtained by taking a moving average of the autocorrelation function of FIG. The calculation unit 15 calculates the number of zero crosses of each frame using an average autocorrelation function. The calculation unit 15 outputs the calculated number of zero crosses and the frame number of the corresponding frame to the identification unit 17.

  Based on the comparison between the number of zero crosses corresponding to the frame of frame number N-1 and the number of zero crosses corresponding to the frame of frame number N, the identification unit 17 determines whether the sound indicated by the frame of frame number N is an unvoiced sound or a voiced sound. It is a part to identify.

  More specifically, when the difference between the number of zero crosses corresponding to the frame of frame number N-1 and the number of zero crosses corresponding to the frame of frame number N is greater than a predetermined value, the identification unit 17 The sound indicated by the frame is identified as an unvoiced sound. In addition, the identification unit 17 determines the sound indicated by the frame of the frame number N when the difference between the number of zero crosses corresponding to the frame of the frame number N-1 and the number of zero crosses corresponding to the frame of the frame number N is smaller than a predetermined value. Is identified as a voiced sound.

  More specific description will be given with reference to FIG. FIG. 8 is a diagram showing the number of zero crossings in each frame. The area where the number of zero crosses of the frame is relatively fluctuating from “13” “20” “24” “30” “25” “30” “39” “8” from the left in FIG. The frame area A is not stable.

  The difference between the number of zero crosses in each frame included in the frame area A and the number of zero crosses in the immediately preceding frame is relatively large. Therefore, the identification unit 17 identifies the sound indicated by each frame included in the frame area A as an unvoiced sound. That is, the identification unit 17 identifies a sound indicated by a frame with a large variation in the number of zero crosses as an unvoiced sound. Note that the waveform in the frame area A shown in FIG. 8 indicates a consonant “s” which is an unvoiced sound.

  From the left in FIG. 8, the area where the number of zero crosses of the frame is relatively stable from the left in FIG. 8 is “7”, “8”, “8”, “8”. It is.

  The number of zero crosses in each frame included in the frame area B is relatively small from the number of zero crosses in the immediately preceding frame. Therefore, the identification unit 17 identifies the sound indicated by each frame included in the frame region B as a voiced sound. That is, the identification unit 17 identifies a sound indicated by a frame with a small variation in the number of zero crosses as a voiced sound. The waveform in the frame region B shown in FIG. 8 indicates a vowel “a” that is a voiced sound.

  FIG. 9 is a graph showing fluctuations in the number of zero crossings in a frame. The horizontal axis indicates the frame number, and the vertical axis indicates the number of zero crossings of the frame. Each of the curve X1 and the curve X2 indicates a variation in the number of zero crossings of the audio signal.

  Since the region Y1 of the curve X1 has a large variation in the number of zero crossings, the identification unit 17 identifies each sound indicated by the frame included in the region Y1 as an unvoiced sound. In addition, since the variation in the number of zero crosses is small in the area Y2 of the curve X2, the identification unit 17 identifies each sound indicated by the frame included in the area Y2 as a voiced sound. In this way, the identification unit 17 detects a variation in the number of zero crosses for each frame, and identifies whether the sound indicated by the frame is an unvoiced sound or a voiced sound.

  The unvoiced sound has less periodicity than the voiced sound, and therefore the fluctuation of the number of zero crosses per frame is large. In addition, the voiced sound has a periodicity as compared with the unvoiced sound, so that the change in the number of zero crosses for each frame is small. Therefore, as described above, by detecting a variation in the number of zero crosses for each frame, it is possible to identify whether the sound indicated by the frame is an unvoiced sound or a voiced sound.

  Subsequently, with reference to FIG. 10, the operation of the voice identification device 10 when identifying a sound based on a voice signal will be described, and the voice identification method according to the present embodiment will be described. FIG. 10 is a flowchart showing the operation of the voice identification device according to the present embodiment.

  When the identification process is started, the voice identification system 1 shifts to the recording mode and “.wav file” is input (S21). When the recording mode is entered, the speech section of the input voice signal S1 is detected by the detection unit 11 (S22). When the speech signal S2 in the speech section is detected, pre-emphasis on the waveform of the speech signal S2 is performed by the detection unit 11 (S23).

  When the audio signal S2 is pre-emphasized, a frame is generated by the generation unit 13 based on the pre-emphasized audio signal S3 (S24). When a frame is generated, an autocorrelation function is calculated for each frame by the calculation unit 15 (S25). When the autocorrelation function is calculated, the moving average of the autocorrelation function is calculated by the calculation unit 15 (S26).

  When the moving average is calculated, the zero cross number of the frame with frame number 0 is calculated by the calculation unit 15 (S27). The number of zero crosses of the frame with the frame number i is calculated by the calculation unit 15 (S28).

  When the discriminator 17 determines that the difference between the number of zero crosses of the frame with the frame number i-1 and the number of zero crosses of the frame with the frame number i is greater than a predetermined value (YES in S29), the sound indicated by the frame with the frame number i The identification unit 17 identifies that the (phoneme) is an unvoiced sound (S30).

  When the discriminating unit 17 determines that the difference between the number of zero crosses of the frame of frame number i-1 and the number of zero crosses of the frame of frame number i is smaller than a predetermined value (NO in S29), the sound indicated by the frame of frame number i The identification unit 17 identifies that the (phoneme) is a voiced sound (S31).

  When the sound is identified as unvoiced or voiced, the frame number i is added (S32). If the identification target frame is not the last frame of the utterance section (NO in S33), the process returns to step S28 and is repeated until the identification target frame becomes the final frame of the utterance section.

  If the identification target frame is the last frame of the utterance section (YES in S33), the identification process is terminated. In this way, it is discriminated whether the sound indicated by the speech signal in the utterance section is voiced sound or unvoiced sound.

  According to the present embodiment, the sound indicated by the frame of frame number N based on a comparison between the number of zero crosses corresponding to the frame of frame number N-1 and the number of zero crosses corresponding to the frame of frame number N is unvoiced sound or voiced sound. To identify. Therefore, the sound indicated by the frame can be identified based on the variation of the number of zero crossings. That is, it is possible to more easily and accurately identify whether the sound indicated by the audio signal is an unvoiced sound or a voiced sound.

  Further, according to the present embodiment, when the identification unit 17 determines that the difference between the number of zero crosses corresponding to the frame with the frame number N-1 and the number of zero crosses corresponding to the frame with the frame number N is greater than a predetermined value, The sound indicated by the N frames is identified as an unvoiced sound. In this case, since the sound indicated by the frame included in the region where the variation in the number of zero crosses is relatively large can be identified as an unvoiced sound, the sound can be accurately identified.

  Further, according to the present embodiment, when the identification unit 17 determines that the difference between the number of zero crosses corresponding to the frame of frame number N-1 and the number of zero crosses corresponding to the frame of frame number N is smaller than a predetermined value, The sound indicated by the N frames is identified as a voiced sound. In this case, since the sound indicated by the frame included in the region where the variation in the number of zero crosses is relatively small can be identified as voiced sound, the sound can be accurately identified.

  A discrimination test was conducted on Japanese consonants using the voice discrimination device 10. FIG. 11 is a table showing the discrimination test results using the voice discrimination apparatus according to the present embodiment. In FIG. 11, “◯” indicates that the identification by the voice identification device 10 is accurate, and “X” indicates that the identification by the voice identification device 10 is incorrect. According to the discrimination test result of FIG. 11, the target rate of voice discrimination using the voice discrimination device 10 is about 90%.

  The hit rate of the modified correlation method known as a conventional speech identification method is about 80%. In the modified correlation method, an autocorrelation function is obtained for a waveform of an input speech signal, and then linear prediction analysis is performed to obtain linear prediction coefficients corresponding to orders. The linear prediction coefficient for the order is a standard indicating the degree to which the current sample value can be expressed by correlation with a plurality of past sampling values. Thereafter, the modified correlation method obtains a prediction residual, detects a peak of the prediction residual, measures a time distance between the peaks, and performs identification based on the measurement result. Since the modified correlation method requires such a complicated operation, identification takes time.

  According to the discrimination test result of FIG. 11, the voice identification device 10 of this embodiment discriminates whether the sound indicated by the voice signal is an unvoiced sound or a voiced sound with an accuracy equal to or higher than the case where the conventional modified correlation method is used. can do. In addition, the speech identification device 10 of the present embodiment can identify whether the sound indicated by the speech signal is an unvoiced sound or a voiced sound by a simpler and quicker method than when the modified correlation method is used.

  In the above-described embodiment, the number of zero crosses of adjacent frames is a comparison target. However, the embodiment of the present invention is not limited to this, and the number of zero crosses of adjacent (near) frames can be compared. That's fine. In this embodiment, the frame is generated while shifting by 30%. However, for example, it is also preferable to perform the framing more finely by reducing the shift amount. When the frame interval is shortened in this way, two or more previous frames may be used as a comparison reference within a range that does not inhibit the correlation of data. In this way, it is also preferable to use the number of zero crossings of frames close to each other, that is, frames that are close enough not to inhibit the correlation of data as a comparison.

It is a lineblock diagram of the voice identification system concerning this embodiment. It is a functional block diagram of the voice identification device according to the present embodiment. It is a graph which shows the waveform of the audio | voice signal which the audio | voice identification apparatus which concerns on this embodiment input. It is a graph which shows the waveform of the audio | voice signal before and after pre-emphasis. It is a graph which shows the spectrum of the audio | voice signal before and after pre-emphasis. It is a graph which shows the waveform of the audio | voice signal after a shift, and a Hamming window. It is a graph which shows the autocorrelation function of a flame | frame. It is a figure which shows the number of zero crosses of each frame. It is a graph which shows the fluctuation | variation of the zero crossing number of a flame | frame. It is a flowchart which shows operation | movement of the audio | voice identification apparatus which concerns on this embodiment. It is a table | surface which shows the identification test result using the audio | voice identification apparatus which concerns on this embodiment.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Voice identification system, 3 ... Microphone 10 ... Voice identification device, 11 ... Detection part, 13 ... Generation part, 15 ... Calculation part, 17 ... Identification part.

Claims (6)

A frame generation means for cutting out the audio signal along the time axis into a plurality of frames by shifting each cut-out start time;
Zero-cross calculating means for calculating the number of zero-crosses corresponding to each of the plurality of cut out frames,
Zero cross comparison means for comparing the number of zero crosses in each of the plurality of frames with the number of zero crosses in a frame adjacent to each of the plurality of frames based on the calculated number of zero crosses;
Phoneme identification means for identifying whether the phoneme corresponding to each of the plurality of frames is an unvoiced sound or a voiced sound based on the comparison result;
A voice identification device comprising:   The phoneme identification unit corresponds to a rear frame when a difference value between the number of zero crosses in each of the plurality of frames and the number of zero crosses in a frame adjacent to each of the plurality of frames exceeds a predetermined threshold. The speech identification device according to claim 1, wherein the phoneme is identified as an unvoiced sound.   The phoneme identifying means is configured to detect a phoneme corresponding to a rear frame when a difference value between a number of zero crosses in each of the plurality of frames and a number of zero crosses in a frame adjacent to each of the plurality of frames is smaller than a predetermined threshold. The voice identification device according to claim 1, wherein is identified as a voiced sound. A first step in which the frame generation means cuts the audio signal along the time axis into a plurality of frames by shifting each cut-out start time;
A second step in which the zero cross calculating means calculates the number of zero crosses corresponding to each of the plurality of frames cut out in the first step;
Based on the number of zero crosses calculated in the second step, a third step in which zero cross comparison means compares the number of zero crosses in each of the plurality of frames with the number of zero crosses in a frame adjacent to each of the plurality of frames;
A fourth step in which the phoneme identifying means identifies whether the phoneme corresponding to each of the plurality of frames is an unvoiced sound or a voiced sound based on the comparison result in the third step;
A voice identification method comprising:   In the fourth step, the phoneme identifying means, when a difference value between the number of zero crosses in each of the plurality of frames and the number of zero crosses in a frame adjacent to each of the plurality of frames exceeds a predetermined threshold value, The speech identification method according to claim 4, wherein the phoneme corresponding to the rear frame is identified as an unvoiced sound.   In the fourth step, the phoneme identifying means determines whether the difference between the number of zero crosses in each of the plurality of frames and the number of zero crosses in a frame adjacent to each of the plurality of frames is smaller than a predetermined threshold. The speech identification method according to claim 4, wherein the phoneme corresponding to the frame is identified as a voiced sound.

Images