KR19980028644A - Speech recognition device and method - Google Patents

Abstract

An apparatus and method are disclosed for performing speech recognition using phase information. The analog speech signal sampled in the speech recognition apparatus according to the present invention is converted into a digital signal by an analog digital converter and then applied to a spectrum analyzer and a phase analyzer. The spectrum analyzer and the phase analyzer perform spectral analysis and phase analysis from the digital speech signal to generate spectral patterns and phase patterns, respectively. The spectral patterns for speech signals to be recognized differently are stored in the spectral template / model, and the phase patterns are stored in the phased template / model by grouping those having the same spectrum. The first pattern matching unit outputs speech data corresponding to the output of the spectrum analyzer among the spectral patterns stored in the spectral template / model as first speech recognition data and is applied to the second pattern matching unit. The second pattern matching unit selects one group from each group of the phased template / model according to the first speech recognition data, and stores the speech corresponding to the output of the phase analyzer among the phase patterns belonging to the selected group. The data is output as the second speech recognition data which is the final recognition result. Such a speech recognition apparatus has a discriminating power even in a noisy environment, and has a merit of recognizing a natural language smoothly.

Description

Speech recognition device and method

The present invention relates to a speech recognition apparatus and method, and more particularly, to a speech recognition apparatus and method having a high identification ability for speech signals.

Referring to FIG. 1, the speech recognition apparatus according to the related art includes an analog digital converter 110, a spectrum analyzer 120, a pattern matching unit 140, and a spectral template / model 130. The analog digital converter 110 converts a sampled analog type voice signal into a digital signal and outputs the digital signal. The digital speech signal is analyzed by the spectrum analyzer 120 to generate a spectral pattern based on the spectral information contained therein. The spectral template / model 130 stores spectral patterns for each speech signal that should be distinguished differently. The pattern matching unit 140 outputs voice data corresponding to a spectral pattern that matches the spectral pattern output from the spectrum analyzer 120 among the spectral patterns stored in the spectral template / model 130 as a recognition result. Since the speech recognition apparatus performs speech recognition using only the spectrum analysis result, the speech recognition apparatus can fully achieve the function when distinguishing speech signals having different frequency forms, but the speech signals to be recognized differently have different frequency forms. In a similar case, there is a problem in that the recognition rate cannot be improved because it cannot be distinguished. That is, as shown in FIGS. 2A and 2B, when the spectral shape shows a large difference, it is possible to distinguish two different voice signals corresponding to FIGS. 2A and 2B by simply spectral size information. On the other hand, as shown in Figures 2c and 2d, speech signals that are to be recognized differently, but similar in the form of frequency, there is a limit in distinguishing only the spectral information. For example, since oh and right have similar frequency forms, it is difficult to distinguish them by simply spectral information. Moreover, when the speech signal performs voice recognition by analyzing voices generated from various voice sources instead of a single voice source (that is, when voice recognition is performed by analyzing voices of several people instead of one voice). Since the distinction of the spectral magnitude information of the speech signals to be distinguished from each other is inferior, it becomes more difficult to distinguish the speech signals only by the spectral information. In addition, even when the speech signal to be recognized is not an isolated word but a natural sound that does not have a complete pronunciation form, it is impossible to perform speech recognition using only the spectral size information because the distinction of the spectral size information is blurred.

Accordingly, it is an object of the present invention to provide a speech recognition apparatus and method in which the recognition rate does not drop even in the case of speech signals generated from various speech sources.

It is still another object of the present invention to provide a speech recognition apparatus and method having a high recognition rate even in the case of natural sounds instead of isolated words.

1 is a block diagram of a speech recognition apparatus according to the prior art.

2A to 2D are diagrams for describing characteristics of speech recognition.

3 is a block diagram of a speech recognition apparatus according to the present invention.

FIG. 4 is a detailed block diagram of the spectrum analyzer 120 and the phase analyzer 310 of FIG. 3.

5A and 5B are graphs showing that the frequency signals have similar shapes in frequency, and FIG. 5C is a graph showing that their normalized phase sizes are shown differently even when the frequency sizes are the same. .

Explanation of symbols for the main parts of the drawings

Analog digital converter 120 Spectrum analyzer

130 ... Spectral template / model 140 ... Pattern matching

310 ... Phase Analyzer 320 ... Phase Templates / Models

330 ... pattern matching part

In order to achieve the above objects, the speech recognition apparatus according to the present invention includes an analog digital converter, a spectrum analyzer, a phase analyzer, a spectral template / model, a phased template / model, a first pattern matching unit and a second pattern matching unit. do. The sampled analog speech signal is converted into a digital signal by an analog digital converter and then applied to a spectrum analyzer and a phase analyzer. The spectrum analyzer and the phase analyzer perform spectral analysis and phase analysis from the digital speech signal to generate spectral patterns and phase patterns, respectively. The spectral patterns for speech signals to be recognized differently are stored in the spectral template / model, and the phase patterns are stored in the phased template / model by grouping those having the same spectrum. The first pattern matching unit outputs speech data corresponding to the output of the spectrum analyzer among the spectral patterns stored in the spectral template / model as first speech recognition data and is applied to the second pattern matching unit. The second pattern matching unit selects one group from each group of the phased template / model according to the first speech recognition data, and stores the speech corresponding to the output of the phase analyzer among the phase patterns belonging to the selected group. The data is output as the second speech recognition data which is the final recognition result.

The spectrum analyzer includes a fast Fourier transformer (FFT) for fast Fourier transforming the output of the analog digital converter; And a spectrum extractor for extracting the spectrum from the output of the fast Fourier transformer and outputting a spectral pattern.

The phase analyzer includes a fast Fourier transformer (FFT) for fast Fourier transforming the output of the analog digital converter; And a phase extractor for extracting phase information from the output of the fast Fourier transformer and outputting a phase pattern.

The first and second pattern matching units perform a Hidden Markov Model Algorithm or a Dynamic Time Warping Algorithm.

In order to achieve the above objects, a speech recognition method according to the present invention includes a first step of converting a sampled analog type speech signal into a digital signal; A second step of performing spectral analysis on the digitally converted speech signal to generate a spectral pattern; A third step of performing a phase analysis on the digitally converted voice signal to generate a phase pattern; A fourth step of generating, as first speech recognition data, speech data corresponding to matching with the spectral pattern obtained in the second step among the spectral patterns for speech signals to be recognized differently; And selecting one group from among phase patterns for voice signals to be recognized differently according to the first voice recognition data, and matching the phase pattern generated in the third step among the phase patterns belonging to the selected group. And a fifth step of generating the voice data as second voice recognition data.

The second step is a fast Fourier transform of the digital speech signal and the spectrum is extracted against it to generate a spectral pattern. The third step is a fast Fourier transform of the digital speech signal and extracts phase information for it to generate a phase pattern.

The fourth and fifth steps are performed by a Hidden Markov Model Algorithm, or by a Dynamic Time Warping Algorithm.

Next, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

3 is a block diagram illustrating a speech recognition apparatus according to the present invention, wherein an analog digital converter 110, a spectrum analyzer 120, a phase analyzer 310, a spectral template / model 130, a phased template / model ( 320, the pattern matching unit 140 and the pattern matching unit 330 are configured.

The analog digital converter 110 converts the sampled voice signal into a digital signal and outputs the digital signal. The digital voice signal DV output from the analog digital converter 110 is applied to the spectrum analyzer 120 and the phase analyzer 310. The spectrum analyzer 120 performs spectral analysis on the digital voice signal DV, extracts spectral information included therein, and outputs the spectral pattern SI accordingly.

The phase analyzer 310 performs phase analysis on the digital voice signal DV, extracts phase information included therein, and outputs a phase pattern PI according thereto. The phase information is composed of a gate that represents individual differences, that is, a gate that represents the characteristic phase of the characteristic sound and characteristic sounds that appear because of different sizes and lengths of articulation organs. Therefore, when the speech recognition apparatus enables the speaker to recognize the speaker, a phase pattern is generated according to the voice letter indicating the individual difference, and when the speech recognition device is for the simple distinction of the sound, the phase pattern according to the intrinsic phase voice of the sound is generated. To occur.

The spectral template / model 130 stores spectral patterns for each voice signal to be recognized differently. The pattern matching unit 140 finds a spectral pattern that matches the spectral pattern (SI) output from the spectrum analyzer 120 among the spectral patterns stored in the spectral template / model 130 and searches the voice data corresponding thereto. Output as first speech recognition data. The phase type template / model 320 stores phase patterns for voice signals to be recognized differently, and each phase pattern is classified and stored in groups having the same spectrum. The pattern matching unit 330 selects one group from each group of the phased template / model 320 based on the first speech recognition data output from the pattern matching unit 140. Then, among the phase patterns belonging to the selected group, one matching with the phase pattern output from the phase analyzer 310 is found, and the voice data corresponding thereto is output as the second voice recognition data. Here, the second speech recognition data is the recognition result of the speech recognition apparatus according to the present invention.

The spectral pattern and phase patterns stored in the spectral template / model 130 and the phased template / model 320 are obtained from speech signals that have a certain length and must be recognized differently. In addition, in order to increase the recognition rate of the sound other than speaker recognition, it is necessary to consider the variation of the phase information according to the individual difference by analyzing the voice signals acquired from various voice sources, making spectral patterns and phase patterns, and storing them as averages. have.

Here, when the phase pattern PI output from the phase analyzer 310 includes only information related to a voiceprint indicating an individual difference for speaker recognition, the pattern matching unit 140 performs speech recognition using a spectral pattern. The pattern matching unit 330 performs speaker recognition according to the phase pattern.

On the other hand, when the phase pattern PI output from the phase analyzer 310 is for enhancing the speech recognition rate and includes only information related to the intrinsic phase voice, the pattern matching unit 140 performs rough speech recognition. In the pattern matching unit 330, more detailed speech recognition is performed. For example, OH and WU can be classified into the same group as having almost similar frequency magnitude forms. In this case, the pattern matching unit 140 recognizes the groups to which they belong, and the pattern matching unit 330 determines whether the pattern matching unit 330 is false or right based on the phase pattern PI generated from the phase analyzer 310. . In addition, the spectral template / model 130 stores a common spectral pattern that O and U belong to and the group to which the O and U belong. Correspondingly stored, the phase pattern for the right is stored corresponding to the voice data called the right. Thus, the pattern matching unit 330 outputs an error if the phase pattern output from the phase analyzer 310 matches the phase pattern corresponding to the error, and outputs the right if the pattern matches the phase pattern corresponding to the right. do. As shown in FIGS. 5A and 5B, speech signals having similar spectral information are classified into the same group, and representative spectral patterns are stored in the spectral template / model 130 together with the group information. 5C is a graph showing that their phase information is different even for voice signals belonging to the same group. In other words, it can be seen that even in the case of speech signals that cannot be sufficiently identified only by the spectral pattern, when the phase pattern is used, the speech signal can be sufficiently identified.

In another case, the phase analyzer 310 may generate a phase pattern based on two or more phase information, such as a voiceprint indicating individual differences and a voice inherent phase voiceprint for fine speech recognition. In such a case, the phase pattern is in the form of digital data such that certain bits in the bit format indicate information about individual differences, and the remaining bits indicate information about negative intrinsic phases. In this case, the phased template / model 320 is configured accordingly. For example, the phase type template / model 320 is a portion in which the phase patterns of voice signals having the same spectrum are classified and stored in the same group, and the phase patterns of the gates representing individual differences are stored. Can be divided. More specifically, in the case where the K bits of the N-bit phase patterns generated by the phase analyzer 310 are information representing negative intrinsic voiceprints and the remaining NK bits are voiceprint information indicating individual differences, the phased template / model ( 320 may be configured to store the storage by classifying the K-bit phase patterns of the speech signals into groups having the same spectral pattern and storing the phase pattern of the NK bit according to the gates indicating individual differences. . In this case, the pattern matching unit 330 performs the algorithm on the K-bit phase pattern and the N-K bit phase pattern in parallel. Therefore, the recognition result includes not only information on the recognized sound but also information on the speaker.

The spectral patterns and the phase patterns stored in the spectral template / model 130 and the phased template / model 320 are analyzed for the speech signal obtained by making a large number of people pronounce the spectral pattern and phase pattern. You need to create and save it.

The phase pattern may be normalized so that a phase value of a specific frequency becomes zero among phase information, and another frequency may be obtained by applying a normalization coefficient.

The pattern matching unit 140 and the pattern matching unit 330 extract a matching spectral pattern and a phase pattern, respectively, and perform a Hidden Markov Model Algorithm or a dynamic time warping algorithm. ).

FIG. 4 is a detailed block diagram of the phase analyzer 310 and the spectrum analyzer 120 shown in FIG. 3. The digital voice signal output from the analog digital converter 110 is a fast Fourier transform (FFT). After conversion at 410, the signal is applied to the spectrum information extractor 420 and the phase information extractor 430, respectively.

The speech recognition method according to the present invention first converts a sampled analog type speech signal into a digital signal. Then, spectrum analysis and phase analysis are performed on the digitally converted speech signal to generate spectral patterns and phase patterns. Recognition is performed using the generated spectral pattern and phase pattern. The first recognition uses a spectral pattern, in which spectral patterns for voice signals to be distinguished from each other are stored in advance, and voice data corresponding to a match among prestored spectral patterns is generated as first speech recognition data. do. Then, the second recognition is performed by using the first speech recognition data and the phase pattern generated by the phase analysis. In the second recognition, a group is determined according to the first voice recognition data, and voice data corresponding to a match among phase patterns belonging to the group is output as second voice recognition data. Here, the second speech recognition data is the final recognition result. In the generating of the spectral pattern and the phase pattern, a fast Fourier transform of the digital speech signal is performed, and the spectral and phase information are extracted for the spectral pattern and the phase pattern. In the first and second recognitions, recognition is performed using a Hidden Markov Model Algorithm or a Dynamic Time Warping Algorithm, respectively.

The present invention is not limited to the above embodiments, and many variations are possible by those skilled in the art within the spirit of the present invention.

As described above, the present invention can clearly recognize natural sound acquired in a real environment in which various noises are generated. That is, the speech recognition apparatus according to the related art can perform speech recognition only for clearly pronounced sounds in a relatively quiet environment such as a laboratory, and does not overcome the noise situation, natural utterance, etc. of the actual recognition environment, and thus the recognition rate is increased. On the contrary, the speech recognition device and method according to the present invention have a discriminating power even in a noisy environment, and the transition process between the sound and the sound is more prominent without the characteristic of achieving a sufficient sound value. Recognition can be performed smoothly even in natural language with various prosody characteristics. In addition, if necessary, by using the phase information, speaker recognition can be performed in parallel with speech recognition, thereby improving the performance of the system employing it.

Claims (14)

An analog digital converter for converting a sampled analog type voice signal into a digital signal; A spectrum analyzer configured to perform spectrum analysis on the output of the analog digital converter to output a spectrum pattern; A phase analyzer configured to output a phase pattern by performing phase analysis on the output of the analog digital converter; A spectral template / model that stores spectral patterns for speech signals to be recognized differently; Phase patterns for voice signals to be recognized differently, each phase pattern being a phased template / model in which those having the same spectrum are divided and stored in a group; A first pattern matching unit configured to compare the output of the spectrum analyzer with the spectral patterns stored in the spectral template / model and output voice data stored corresponding to the matching spectral pattern as first speech recognition data; And Select one group of each group of the phase type template / model according to the first speech recognition data, compare the phase patterns belonging to the selected group with the output of the phase analyzer, and store the corresponding phase patterns And a second pattern matching section for outputting the voice data as the second voice recognition data. The method of claim 1, wherein the spectrum analyzer A fast Fourier transformer (FFT) for fast Fourier transforming the output of the analog digital converter; And And a spectrum extractor for extracting a spectrum from an output of the fast Fourier transformer and outputting a spectral pattern. The method of claim 1, wherein the phase analyzer A fast Fourier transformer (FFT) for fast Fourier transforming the output of the analog digital converter; And And a phase extractor for extracting phase information from the output of the fast Fourier transformer and outputting a phase pattern. The apparatus of claim 1, wherein the first pattern matching unit performs a Hidden Markov Model Algorithm. The speech recognition apparatus of claim 1, wherein the first pattern matching unit performs a dynamic time warping algorithm. The apparatus of claim 1, wherein the second pattern matching unit performs a Hidden Markov Model Algorithm. The apparatus of claim 1, wherein the second pattern matching unit performs a dynamic time warping algorithm. A first step of converting the sampled analog form voice signal into a digital signal; A second step of performing spectral analysis on the digitally converted speech signal to generate a spectral pattern; A third step of performing a phase analysis on the digitally converted voice signal to generate a phase pattern; A fourth step of generating, as first speech recognition data, speech data corresponding to matching with the spectral pattern obtained in the second step among the spectral patterns for speech signals to be recognized differently; And Selecting one group of phase patterns for voice signals to be recognized differently and matching the phase pattern generated in the third step among the phase patterns belonging to the selected group And a fifth step of generating corresponding speech data as second speech recognition data. The method of claim 8, wherein the second step Fast Fourier transforming the digital speech signal; And Extracting a spectrum from the fast Fourier transformed digital speech signal to generate a spectral pattern. The method of claim 8, wherein the third step Fast Fourier transforming the digital speech signal; And Extracting phase information from the fast Fourier transformed digital speech signal to generate a phase pattern. The speech recognition method of claim 8, wherein the fourth step is performed by a Hidden Markov Model Algorithm. 9. The speech recognition method of claim 8, wherein the fourth step is performed by a dynamic time warping algorithm. 9. The speech recognition method of claim 8, wherein the fifth step is performed by a Hidden Markov Model Algorithm. The speech recognition method of claim 8, wherein the fifth step is performed by a dynamic time warping algorithm.