JP2011180308A - Voice recognition device and recording medium - Google Patents

Abstract

A feature unique to speech is reliably and robustly extracted and used for speech recognition regardless of various situations in an actual use environment.
In a method based on a Hidden Markov Model (HMM) that is currently widely used as a general speech recognition method, all the analysis points (frames) in the entire speech section are compared when the input speech is compared with a standard pattern. Since the voice and noise are equally reflected in the score in order to uniformly evaluate the feature amount of the feature, the method uses the features of the maximum and minimum parts of the formant in which important features for identifying the voice exist. By using the amount intensively, it becomes possible to perform speech recognition robust to noise.
[Selection] Figure 2

Description

  The present invention relates to a voice recognition device and a recording medium.

  As a method for improving the recognition accuracy in a noisy environment in speech recognition, a method of calculating and subtracting the noise spectrum information of the silent part from the whole (spectrum subtraction method) or obtaining the average value of cepstrum over a long time In addition to a pre-processing method such as a method for normalizing (cepstrum average normalization method), a method for statistically modeling an assumed noise with a Markov model is known (see, for example, Patent Document 1). .

JP 2006-139033 A

However, in the above speech recognition method, when the input speech is compared with the standard pattern, the feature quantities at all the analysis points (frames) in the entire speech section are evaluated on the same scale, so that both speech and noise are scored equally. It is reflected in. For this reason, even in the case of speech recognition of small vocabulary, there is a problem that the recognition accuracy is greatly lowered due to the influence of noise.
In addition, if the noise characteristics are somewhat steady and predictable in the environment, it can be handled by the conventional method by modeling it statistically as in Patent Document 1, but in an actual application situation, Unexpected noise is unavoidable, and a new approach is required.

  Therefore, the present invention makes it possible to perform speech recognition that is robust against noise by intensively using the feature quantities of the maximum and minimum parts of the formant in which important features for identifying speech exist, and to achieve speech recognition. When constructing an evaluation pattern, the knowledge of phonetics and acoustic analysis is used to automatically generate maximum and minimum formant chain patterns from the dictionary notation text of the recognition target word. A speech recognition device that does not require statistical learning from such a large amount of data and can recognize arbitrary vocabulary written in letters, and a storage medium storing speech recognition feature data used for these devices It is intended to provide.

The speech recognition apparatus according to the present invention is a formant frequency, formant power level, fundamental frequency, autocorrelation coefficient, cepstrum peak value, spectral tilt, pitch channel power, or frequency high, which is obtained by frequency analysis of a speech signal uttered by a speaker. Means for extracting the maximum value and the minimum value in the trajectory of the change pattern in the time axis direction of the parameter of the region power, and the time interval in the vicinity thereof,
Means for generating a feature amount pattern of an audio signal by a chain pattern in which feature parameters in a section are connected;
Means for generating a standard feature pattern corresponding to each word from the character notation of each word of the vocabulary to be subjected to speech recognition;
Means for determining a word suitable for the voice signal by comparing the feature pattern of the voice signal uttered by the speaker with the standard feature pattern.
The speech recognition apparatus according to the present invention is a voiced sound in which temporal partial sections of a speech signal are classified by a speech generation method from feature parameters and feature amount patterns obtained by frequency analysis of a speech signal uttered by a speaker. Means for generating a sequence of articulatory description categories of segments, unvoiced segments, silence segments, or voiced consonant segments;
Means for deriving a standard symbol string of articulatory description classification corresponding to each word from the character notation of each word of the vocabulary targeted for speech recognition;
Means for generating a correspondence table between a column of articulatory description categories corresponding to a speech signal uttered by a speaker and a standard symbol string of articulatory description categories;
Means for determining a word suitable for the audio signal by performing pattern matching using the correspondence table and the feature amount pattern together.
The speech recognition apparatus according to the present invention includes means for generating a plurality of standard symbol sequences for one word according to a deformation rule that systematizes knowledge about speech fluctuations in a standard symbol sequence of articulatory description category. .
The speech recognition apparatus preferably further includes means for outputting the speech recognition result of the speaker.
The recording medium according to the present invention includes, for each speech language system, a dictionary in which the pronunciation contents of all words to be recognized are represented by symbols, a dictionary in which pronunciation deformation patterns of the words are described by symbols, and a standard for each phoneme At least one of a typical formant feature value and a temporal change pattern of a standard formant feature value of each word is stored.
According to the present invention, it is possible to intensively use the feature quantities of the maximum and minimum parts of formants in which important features for identifying speech exist. Therefore, according to the present invention, robustness against noise can be achieved. Voice recognition is possible.
According to this configuration, when constructing an evaluation pattern for speech recognition, the formant maximum / minimum chain pattern is automatically extracted from the text of the dictionary notation of the recognition target word using the knowledge of phonetics and acoustic analysis. It becomes possible to generate any word written in letters without requiring statistical learning from a large amount of data.

  As described above, according to the present invention, a speech recognition apparatus, a speech recognition method, and speech that can perform speech recognition robust to noise without requiring learning from a large amount of data for an arbitrary word. A storage medium storing a recognition program and various information used in the speech recognition is provided.

The figure which shows the structure of the computer which operate | moves as a speech recognition apparatus concerning embodiment of this invention. Diagram showing the configuration and processing flow of a speech recognition system The figure which shows the example which extracted the feature-value related to formant from an audio signal Diagram showing examples of extraction of local maximum and minimum values of formant frequency trajectory Diagram showing maximum / minimum chain pattern of formant frequency in a chain of two vowels A diagram showing a column of speech articulatory description segments extracted from a speech signal sample of a certain word "password"

DESCRIPTION OF EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a diagram showing a configuration of a computer that operates as a speech recognition apparatus according to an embodiment of the present invention.
As shown in FIG. 1, the computer that operates as the speech recognition device 10 includes a central processing unit (CPU) 12a, a main storage device (memory) 12b, an auxiliary storage device 12c, an input / output interface 12e, a microphone 12f, and a monitor 12d or A signal output port 12g is provided.
The CPU 12a, the memory 12b, the monitor 12d, and the input / output interface 12e are connected to each other via a system bus 12h, and the microphone 12f and the signal output port 12g are connected to the system bus 12h via the input / output interface 12e. Yes.

  Hereinafter, the flow of speech recognition by the speech recognition apparatus 10 will be schematically described with reference to FIG. In the speech recognition system, a set of words to be subjected to speech recognition is determined in advance. From this set, the user utters one of the words that are the elements toward the microphone 12f. The sound at this time is collected by the microphone 12f and converted into digital data by the input / output interface 12e as an analog sound signal. Hereinafter, this digital data is referred to as “voice signal” or “voice waveform data” in the sense that the waveform of an analog signal is converted into digital data. A program for voice recognition is stored in an auxiliary storage device 12c such as an HDD or a flash memory, and is executed by the CPU 12a or the memory 12b. The result of recognizing the input audio signal is displayed on the screen by the monitor 12d or output as a signal for controlling another device from the signal output port 12g.

  Hereinafter, a voice recognition system for operating a computer as the voice recognition device 10 will be described. FIG. 2 is a diagram illustrating a processing procedure in the speech recognition system. A speech recognition system 20 shown in FIG. 2 includes a speech signal analysis unit, an articulatory description section extraction unit, an articulatory description category determination unit, a feature amount series generation unit, a recognition target word set pronunciation dictionary database (DB), and an articulatory pronunciation description. Generator, articulatory pronunciation description dictionary database (DB), pronunciation modification rule database (DB), pronunciation pattern generator, pronunciation pattern dictionary database (DB), phoneme standard feature database (DB), standard feature sequence generator, A feature amount pattern matching unit is provided.

  A voice signal to be subjected to voice recognition processing is input to a voice signal analysis unit. The voice signal analysis unit obtains a basic feature amount by using a signal processing technique centered on frequency analysis. The feature amount is obtained every fixed time (for example, ms) so that a temporal change pattern can be expressed. Examples of suitable feature values are shown in Table 1. Table 1 is a table showing examples of feature amounts obtained by frequency analysis of audio signals.

  Here, the pitch is a fundamental frequency corresponding to the pitch of the voice, and the extraction method includes a method of estimating periodicity based on the autocorrelation function of the speech waveform and a cepstrum that is the Fourier transform of the logarithmic power spectrum. There is a method to estimate using. In a preferred implementation, both are used together and selectively used according to noise characteristics. Formant is an important feature quantity for recognizing speech at a frequency that resonates strongly when the inside of the mouth (the vocal tract) becomes an acoustic tube when generating speech. As a formant extraction method, a method is generally known in which the resonance characteristics of the vocal tract are approximated by an AR model and its pole is obtained from a linear prediction coefficient (LPC). An example of formant extraction is shown in FIG.

  Next, the obtained time series pattern of the feature amount is input to the articulatory description section extraction unit. In the articulatory description section extraction unit, as shown in FIG. 3, from the time change pattern of the formant frequency, as shown in FIG. The peripheral section ({p, v} section) of the part taking the value is obtained. The time width of the peripheral section is preferably set by giving a threshold value for the change amount of the feature amount. Since the formant is a feature amount including information that is fundamental to the generation of voice, it is considered that important information for distinguishing voice is concentrated in this section. At this time, even if there is noise in a section other than the p and v sections, it is not affected, so robustness against noise can be expected. Hereinafter, this section is referred to as an articulatory description section. Depending on the type of consonant, such as unvoiced consonants, features other than formants are important, and features that are suitable for the type of phoneme can be adopted. To do.

  As a typical example of the formant change pattern, FIG. 5 shows a p-v pattern when two vowels are chained in a voice. For example, when linking from / i / to / a /, F1 (first formant) takes a pattern of v → p, and F2 (second formant) takes a pattern of p → v. It is an advantage of this method to capture the global characteristics of such a pattern.

  Next, in order to obtain a more detailed feature amount pattern, phonemes constituting the speech are roughly classified, and the global structure of the speech is described by a sequence in which the phonemes are arranged. This is called articulatory description division. For example, as a preferred implementation, voices are classified based on the presence or absence of vocal cord vibrations and described as four categories of {voiced sound, unvoiced sound, voiced consonant, silence}. Here, the silence includes a silent section in the word accompanying the closing of the plosive or the prompt sound.

The articulatory description category is determined by using the value of the p-v pattern obtained by the articulatory description segment extracting unit in the previous stage and other feature amounts in the articulatory description category determining unit of FIG. As a suitable example, Table 2 shows parameters used for determining the four articulatory description categories in the previous section.
Table 2 is a diagram showing feature parameters used to determine the articulatory description category.

  For example, the voiced consonant section (Vconso) is determined from the valley (v) of the autocorrelation peak value and the valley (v) of the global power parameter in the voiced sound section. An example of the articulatory description division obtained in this way is shown in FIG. In this example, an audio signal uttered as "PIN code" is analyzed to determine the articulatory description category and displayed as a rectangle with different shades.

Next, a standard pattern generation method will be described. A set of words targeted for speech recognition is determined, and a set of words described in symbols with phonemes as units is prepared as a recognition target word set pronunciation dictionary DB in FIG.
The articulatory pronunciation description dictionary DB is obtained by converting the recognition target word set pronunciation dictionary DB into an articulatory description corresponding to each phoneme in the articulatory pronunciation description generator. As the articulatory description, for example, when the voiced sound is described by (V), the voiceless sound is (L), the voiced consonant is (C), and the soundless is (S) by classification based on the presence or absence of vocal cord vibration as described above, The typical pronunciation description dictionary DB is described by four types of symbols {V, L, C, S}.

In general, since human speech is acoustically diverse even for the same word, it is desirable that the pronunciation dictionary includes not only one standard pronunciation but also a plurality of pronunciation variations. For this purpose, the pronunciation deformation pattern generation unit automatically generates a pronunciation dictionary DB including various phonetic deformation patterns. For this purpose, the pronunciation modification rule DB is used. The pronunciation deformation rules include empirical rules that observe the results of acoustic analysis of many audio data samples, as well as deformation rules based on phonetic knowledge such as so-called devoicing and omission.
Table 3 shows examples of pronunciation modification rules based on the above-mentioned articulatory pronunciation description method.

For example, the word “adapter” is described as VCVSLVSVLV in a sequence obtained directly from the dictionary, but by applying the rewrite rules x and ii, a sequence such as VCVSLSV is also obtained.
From the pronunciation pattern dictionary obtained in this way, the standard feature value series generation unit calculates the average value of formants (F1 to F3) of Japanese vowels obtained in advance for the vowel phonemes corresponding to the V section. By associating, a series of feature vectors having a formant frequency as a feature quantity is obtained as a standard pattern.

On the other hand, with respect to the input speech signal, the pronunciation description sequence obtained by the articulatory description category determination unit is associated with the pronunciation pattern dictionary DB reflecting the pronunciation modification rule, and the corresponding articulatory description section. The feature amount pattern corresponding to the input voice signal can be obtained by selectively adopting the feature amount in and configuring the time series. This processing is performed by the feature amount series generation unit in FIG.
By comparing the feature amount pattern obtained in the previous section with the standard pattern obtained in the previous section and calculating the degree of matching based on a suitable distance measure, the most suitable word is obtained as the recognition result. .

In the following, we describe the results of building a prototype of a speech recognition system using this method and evaluating its effectiveness.
CENSREC is widely used as common data for evaluating the speech recognition performance of Japanese words in a noisy environment. Among them, CENSREC-3, which supports word speech recognition, contains speech samples recorded in a car with various driving environments and a plurality of types of microphones. CENSREC-3 assumes a general hidden Markov model (HMM) as a speech recognition method, and prepares sentence speech prepared for learning and speech of 50 words as shown in Table 4 prepared for evaluation. However, since this method does not perform statistical learning, sentence speech is not used. Table 4 is a table showing a recognition target vocabulary list used in the speech recognition experiment.

  Table 5 shows the results recognized by this method using the word sound for evaluation of CENSREC-3.

  Since the experimental conditions such as the standard model generation method are different, it cannot be directly compared. However, compared with the baseline recognition performance of CENSREC-3, this method provides high recognition accuracy with a small amount of features in a noisy environment. The possibility to be obtained was shown.

  In the above, the entire speech recognition program shown in FIG. 2, the recognition target word set pronunciation dictionary DB, the articulatory pronunciation description dictionary DB, the pronunciation modification rule DB, the pronunciation pattern dictionary DB, and the phoneme standard feature DB are such as CD-ROM. It can be recorded on a medium and used in the speech recognition apparatus 10. In other words, each recognition target word set can be recorded in one CD-ROM and used in accordance with the application scene.

According to the voice recognition device 10 described above, the following effects are produced. In other words, by using the speech recognition device 10, it is possible to realize speech recognition that operates robustly against various noise environments and various pronunciations. Can do.
Moreover, according to the speech recognition apparatus 10, compared with the method based on the existing general hidden Markov model (HMM), since the standard model is simple without requiring statistical learning from a large amount of data, the recognition is performed. Light processing and compact program implementation.
Further, according to the voice recognition device 10, since the feature amount corresponding to the state of articulation is adopted, it is considered that there is an advantage in application to voice training and the like.

10 Voice recognition device 12a Central processing unit (CPU)
12b Main memory (memory)
12c Auxiliary storage device 12d Monitor 12e Input / output interface 12f Microphone 12g Signal output port 12h System bus 20 Voice recognition system

Claims (5)

A frequency analysis of the speech signal uttered by the speaker, formant frequency, formant power level, fundamental frequency, autocorrelation coefficient, cepstrum peak value, spectral tilt, pitch channel power, or frequency high-frequency power parameters in the time axis direction Means for extracting local maximum values and local minimum values in the locus of the change pattern and temporal intervals in the vicinity thereof;
Means for generating a feature amount pattern of the audio signal by a chain pattern in which feature parameters in the section are connected;
Means for generating a standard feature amount pattern corresponding to each word from the character notation of each word of the vocabulary to be subjected to speech recognition;
A speech recognition apparatus comprising: means for determining a word that matches a speech signal by comparing the feature amount pattern of the speech signal uttered by the speaker with the standard feature amount pattern. From the feature parameter and feature amount pattern obtained by frequency analysis of the speech signal uttered by the speaker, the temporal partial section of the speech signal is classified by the speech generation method, voiced sound classification, unvoiced sound classification, silent classification, Or means for generating a sequence of articulatory description categories of voiced consonant categories;
Means for deriving a standard symbol string of the articulatory description category corresponding to each word from the character notation of each word of the vocabulary targeted for speech recognition;
Means for generating a correspondence table between the column of the articulatory description section corresponding to the speech signal uttered by the speaker and the standard symbol string of the articulatory description section;
A speech recognition apparatus comprising: means for determining a word that matches a speech signal by performing pattern matching using the correspondence table and the feature amount pattern together.   The standard symbol string of the articulatory description section comprises means for generating a plurality of standard symbol strings for one word according to a deformation rule that systematizes knowledge about speech fluctuations. The speech recognition apparatus according to 2.   The speech recognition apparatus according to claim 2, further comprising means for outputting a speech recognition result of the speaker.   For each spoken language system, a dictionary that describes the pronunciation of all the words to be recognized in symbols, a dictionary that describes the pronunciation deformation patterns of the words in symbols, and the standard formant feature values for each phoneme A recording medium that stores at least one of the temporal change patterns of the standard formant feature of each word.