RU2268504C9 - Method for recognition of speech patterns and device for realization of method - Google Patents

Abstract

FIELD: analysis and recognition of speech signals, can be used for recognition of speech patterns.

SUBSTANCE: device for realization of aforementioned speech phoneme recognition method has: computing system, including clock generator, controller, random-access memory device, central microprocessor unit, meant for forming bispectral signs and recognizing them on basis of speech phonemes, digital-analog converter, long-term memorizing device, video-controller and analog-digital converter, and also keyboard, display, headphones and a microphone.

EFFECT: increased precision of speech patterns recognition due to forming of phoneme signs for speech phonemes recognition based on application of bispectral analysis, based on transformation of digital code series, appropriate for speech signals, to bispectral zone, characterizing interaction between values of Fourier components at different frequencies within speech spectrum, and thus, to provide selection of an additional, significantly new information from speech signals, to increase precision of phoneme recognition.

2 cl, 5 dwg

Description

The invention relates to the field of analysis and recognition of speech signals and relates to a method for recognizing phonemes of speech based on the use of generated bispectral features of phonemes and devices for implementing the method.

A known method for recognizing speech words [1], in which words are stored in the dictionary in the form of a phonetic model, composed of a sequence of phonetic symbols corresponding to the phonemes of the word. For each phoneme from the word, characteristic parameters (attributes) are determined. To obtain phoneme features, a speech signal representation in the spectral region is used, i.e. the speech signal is subjected to spectral analysis, which is carried out using fast Fourier transform (FFT) algorithms.

The formation of phoneme signs in this method is based on the allocation of the spectrum features by evaluating the spectral power density in different frequency bands, finding the maximum energy of the spectrum (measuring the formant frequencies).

The disadvantage of this method is that the spectral features are not always stable and do not provide high accuracy of recognition of phonemes.

There is also a method of distinguishing such a sign of phonemes of voiced sounds, as the frequency of the fundamental tone [2]. The method is based on the decomposition of the speech signal into a sequence of Fourier spectra, finding the absolute maximum, estimating the average frequency of the fundamental tone of the studied speech signal, corresponding to the phoneme of voiced sound, by approximating the estimates of the fundamental frequency obtained using three types of parallel separators.

The disadvantage of this method is the not very high accuracy in determining the frequency of the fundamental tone, which is associated with a strong dependence of the circuit as a whole on the accuracy of each algorithm included in its composition.

Closest to the proposed method is a speech phoneme recognition method adopted as a prototype implemented in real-time speech recognition system based on computer technology [3], based on the representation of words in the form of sequences of characteristic segments, generally corresponding to speech phonemes. Each segment is described by a limited set of attributes, and each word is described by a sequence of such segments.

To distinguish features, the speech signal corresponding to the speech segment (phoneme) is converted in an analog-to-digital converter (ADC) into digital code sequences subjected to spectral analysis in the spectrum analyzer using the Fast Fourier Transform (FFT) algorithm.

By analyzing the obtained successive spectra in the data processing and control unit, the formation of such phoneme attributes as the positions of the formant frequencies (by finding the maximums of the spectrum), the dynamics of the spectrum energy in specially selected frequency bands are formed. The phoneme recognition procedure in this method is based on comparing the generated features of the speech segment with the signs of phoneme patterns or phoneme classes stored in the memory of the data processing and control unit, as a result of which the recognized phoneme is sequentially assigned to the phoneme class, and then a specific phoneme is identified.

Thus, the signs of phonemes in this method are formed, taking into account the properties of the spectra of the corresponding sections of speech.

The disadvantage of this method is the low accuracy of recognition of phonemes due to the use of unstable informative features of phonemes based on the use of spectral representation.

The objective of the invention in terms of the method is to increase the accuracy of recognition of phonemes of speech.

A fundamentally new approach to solving the problem of increasing the accuracy of phoneme recognition is the use of bispectral analysis, based on a special type of transformation of digital sequences of codes corresponding to speech signals, into a bispectral region characterizing the interactions between the values of the Fourier components at different frequencies in the speech range Thus, the allocation of additional, essentially new information from recognized speech x signals.

The objective of the invention is achieved by the fact that in the method of recognition of phonemes of speech, sound vibrations corresponding to speech commands are received and converted into electrical signals, analog-to-digital conversion of electrical signals into digital sequences of codes that convert to the form of graphic functions in time, upon analysis of which on the display, the operator sets the boundaries of phonemes as part of speech commands, while simultaneously listening to the sound signals in the headphones corresponding to the selected sections of the speech out teams.

In the computing system, at intervals within the selected boundaries of the phonemes, the processing of the corresponding digital code sequences is carried out, which consists in the formation of bispectral features of phonemes. The formation of bispectral features is based on the conversion of digital code sequences into the bispectrum region, which, due to its properties, provides a more complete selection of information from the signal, which increases the accuracy of phoneme recognition.

Accumulating the formed bispectral attributes for phonemes identical in the alphabetic code from different speech commands, the phoneme standards (aggregate matrices of bispectral attributes) are formed. When recognizing phonemes, the operator, on the basis of audiovisual analysis, identifies the boundaries of the phoneme being recognized (its letter code is stored in the memory of the computer system). For digital sequences of codes corresponding to the selected interval, bispectral features are formed and compared with the standards of all phonemes to obtain a decision on the recognized phoneme.

The recognition accuracy of phonemes is determined by comparing the letter code of the recognized phoneme with the letter code of the recognized phoneme, extracted from the memory of the computer system.

The invention is illustrated by drawings.

Figure 1 shows the region of symmetry and the region of existence of the bispectrum module in the frequency plane f ₁ , f ₂ , where f ₁ , f ₂ are the frequency axes corresponding to the studied range of the speech signal (0-8 kHz).

The bispectrum module has inherent symmetry properties, which are reflected in figure 1 in the form of symmetry regions. For a complete description of the bispectrum module, it is sufficient to use only one region of symmetry, called the region of existence of the bispectrum module, limited by the shaded triangle OEX in Fig. 1.

Figure 2-4 shows examples of graphical images of arrays of values of the bispectrum modules of phonemes A, C, B, equivalent to the region bounded by the triangle OXY, including two symmetric regions OEX and OEY, shown in figure 1.

The arrays of values of the bispectrum module in these examples are calculated by the claimed method, where f ₁ , f ₂ - frequency axis (sampling step 125 Hz).

Figure 5 presents a functional diagram of a device for implementing the method.

The scheme includes the following elements:

1. Microphone - receives sound vibrations corresponding to speech commands, and converts them into electrical signals;

2. Analog-to-digital converter (ADC) - converts electrical signals into digital code sequences;

3. Controller - provides information management in a computer system;

4. Random access memory (RAM) - accumulates and stores digital information about phonemes used in online mode;

5. Processor - makes the necessary transformations and calculations according to the program;

6. Clock generator - synchronizes the operation of the ADC, processor, controller, RAM, digital-to-analog converter (DAC), long-term storage device (DZU);

7. Long-term storage device (DZU) - accumulates digital information about phonemes necessary for long-term storage;

8. Keyboard - designed for the operator to enter information on the analyzed and recognizable phonemes from speech commands;

9. Display - displays phoneme information for the operator;

10. Digital-to-analog converter (DAC) - converts digital sequences of codes corresponding to speech phonemes into electrical signals;

11. Headphones - convert electrical signals into audio signals;

12. Video controller - converts information into a video signal for display;

13. Computing system - includes blocks that implement: converting electrical signals corresponding to sound vibrations into digital sequences of codes, calculating the parameters necessary for the formation of signs and recognition of phonemes, transmitting information to other blocks of the device.

The inventive method is as follows:

reception of sound vibrations corresponding to the speech commands delivered by the operator, and their conversion into electrical signals is carried out by microphone 1, then the signal is fed to an analog-to-digital converter (ADC) 2, where electrical signals are converted into digital sequences of codes that are controlled by controller 3 of the computing system 13 are transferred to random access memory (RAM) 4 for storage and use during subsequent processing in the processor 5,

the synchronization of the ADC 2, controller 3, RAM 4, long-term storage device (DZU) 7, digital-to-analog converter (DAC) 10 and processor 5 is carried out by a clock generator 6,

To train the phoneme recognition system, phoneme standards are created, for this purpose digital code sequences corresponding to the commands spoken by the operator are transferred from RAM 4 to processor 5, where they are converted using the program to the form of a graphic function in time transmitted through video controller 12 to display 9 ,

the operator, analyzing the video information presented on the display screen 9 corresponding to the spoken command, performs manual marking, which consists in setting the boundaries of the phoneme as part of the team,

at the same time, the operator listens with the aid of the headphones 11 an audio signal corresponding to a dedicated phoneme obtained by processing the digital sequences of codes in the processor 5, transferring them to the DAC 10, where they are converted into an electrical signal that is converted into the audio signal in the headphones 11,

the operator enters using the keyboard 8 the composition of the team and the letter corresponding to the selected phoneme,

in RAM 4, the numbers of discrete samples of the beginning and end of the selected phonemes and the corresponding letter codes are accumulated, as well as the commands containing the borders of the phonemes,

the accumulated information is transmitted for storage in DZU 7,

for the formation of phoneme standards by the processor 5 according to the program, based on the phoneme's letter code, the corresponding digital code sequences are selected from the DZU 7,

further, in processor 5, using the FFT algorithm, an array of values of the spectrum modulus | F (jf) |,

where f is the frequency corresponding to the speech range of 0-8 kHz,

the array of values of the bispectrum module is calculated as the product of three Fourier transforms (spectrum modules) [4, 5] in the following form:

| S (f ₁ , f ₂ ) | = | F [-j (f ₁ + f ₂ )] | · | F (jf ₁ ) | · | F (jf ₂ ) |

where f ₁ , f ₂ are the frequencies corresponding to the studied range of the speech signal (0-8 kHz with a sampling step of 125 Hz),

moreover, as the values of the spectrum moduli | F [-j (f ₁ + f ₂ )] |, | F (jf ₁ ) |, | F (jf ₂ ) | values from a previously computed array of spectrum modulus values are used.

Figure 2, figure 3, figure 4 shows examples of graphic images of arrays of values of the module of the bispectrum of phonemes A, C, B, calculated by the present method; the presented examples show different maximum distributions in the array of values of the bispectrum modulus for different phonemes. An analysis of phonemes identical in letter code shows the proximity of the areas of the presence of maxima of the bispectrum modulus. In this case, each of the phonemes is characterized by an individual arrangement of the regions where the maximum values of the high-intensity bispectrum modulus are present, in particular, as can be seen from FIG. 2, FIG. 3, and FIG. low frequencies (up to 2 kHz), and for the phoneme C in the high frequency region (above 4 kHz). These properties were used in the claimed method for recognition of phonemes.

In the resulting array of values of the bispectrum modulus, maxima are selected and their positions in the frequency plane f ₁ , f ₂ are determined,

for phonemes identical in alphabetic code, the positions of the maxima of the bispectrum module (IMB) are accumulated in the form of the aggregate IMB matrix in the frequency plane f ₁ , f ₂ , which is the phoneme standard, phoneme standards are accumulated in DZU 7,

before recognition of phonemes, the operator enters the threshold value for the probability of the presence of IMB, which is used in the recognition process and stored in RAM 4,

when phonemes are recognized, the operator using the keyboard 8 dials a command, in which it is necessary to recognize the phoneme, while the graphic implementation of the selected command is extracted from the DZU 7 and transmitted via the video controller 12 to the display 9,

the operator selects the phoneme that needs to be recognized, and enters its alphabetic code stored in RAM 4,

using digital code sequences corresponding to the selected phoneme, operations similar to those described above are performed: calculation of an array of values of the spectrum module; calculation of an array of bispectrum module values; the selection of IMB and the determination of their positions on the frequency plane f ₁ , f ₂ ,

further, the IMB positions of the recognized phoneme are compared with the IMB positions of each of the phoneme standards, while the probability of the presence of IMB in the standards is higher or equal to the threshold value determined previously by the operator and stored in RAM 4,

a decision to recognize a phoneme is made in relation to the number of identifiable phonemes coinciding in the IMB position with the IMB of each of the phoneme patterns and the total number of IMB of each of the phoneme patterns. That phoneme, for the standard of which the maximum of this ratio is reached, is considered recognized, in accordance with it, an alphabetic code is formed and compared with the alphabetic code of the phoneme previously entered by the operator and extracted from RAM 4; if they coincide, the phoneme is considered correctly recognized; the display 9 displays a recognizable and recognized phoneme and a decision on the accuracy of recognition of the phoneme.

2. A device for implementing the method.

To implement the method of recognition of phonemes of speech, the device shown in Fig.5.

The prototype of this device is a device [3], which is characterized by the presence

a microphone for receiving sound vibrations and converting them into electrical signals,

analog-to-digital Converter, converting electrical signals into a digital sequence of codes,

a data processing and control unit (BODU) receiving a digital sequence of codes,

spectrum analyzer, which is part of the BODU, implemented on an integrated circuit that performs digital spectral analysis of speech signals using the FFT algorithm,

in BODU produced:

logical analysis of successive spectra to form features of speech segments,

assignment of a speech segment to a specific class of phonemes based on a comparison of the characteristics of a recognizable speech segment with signs of all classes of phonemes in the BODU memory,

identification of a specific phoneme related to a previously defined class of phonemes based on a comparison of the spectral features of a recognizable phoneme with spectral features that uniquely characterize a phoneme based on statistics of voice announcers.

This device by technical nature is the closest analogue of the invention.

The disadvantage of this device is the low accuracy of recognition of phonemes, due to the fact that they use the spectrum of the signal corresponding to the recognizable phoneme, which are not sufficiently informative and stable features due to the limited dimensionality of the frequency space.

The objective of the proposed device is to increase the accuracy of recognition of phonemes of speech.

This object is achieved by the fact that in the device for implementing the method containing a microphone and an ADC, a computer system is additionally introduced, including a clock generator, controller, random access memory (RAM), processor, digital-to-analog converter (DAC), long-term memory (DZU), a video controller and ADC, while the ADC is made with an additional input connected to a clock generator and with sixteen outputs, as well as a keyboard, display, and headphones.

Figure 5 presents a functional diagram of a device for implementing the method of recognition of phonemes of speech, necessary to perform the proposed method as a process of performing actions on a material object using material means necessary for the implementation of the claimed invention, where

1. Microphone.

2. An analog-to-digital converter (ADC).

3. The controller.

4. Random access memory (RAM).

5. The processor.

6. The clock generator.

7. Long-term storage device (DZU).

8. The keyboard.

9. The display.

10. Digital-to-analog converter (DAC).

11. Headphones.

12 Video controller.

13. Computing system.

An explanation of the operation of the device according to figure 5.

The device comprises a computing system 13, including an ADC 2, a clock 6, a controller 3, RAM 4, a processor 5, a DAC 10, a DZU 7, a video controller 12, as well as a microphone 1, a keyboard 8, a display 9, headphones 11, and

the microphone output 1 is connected to the first input of the ADC 2,

the output of the clock generator 6 is connected to the second input of the ADC 2, with the first input of the controller 3, with the first input of the processor 5, with the first input of RAM 4, with the first input of DAC 10, with the first input of DZU 7,

from the first to sixteenth outputs of the ADC 2 are connected to the second to seventeenth inputs - the outputs of the processor 5, which is designed to form bispectral signs and recognition of phonemes of speech on their basis, from the second to seventeenth inputs - the outputs of RAM 4, from the second to seventeenth inputs - the outputs of the DZU 7, from the second to seventeenth inputs - the outputs of the controller 3, from the second to seventeenth inputs of the DAC 10, from the first to sixteenth inputs of the video controller 12,

the output of the DAC 10 is connected to the input of the headphones 11,

from the first to fifth outputs of the keyboard 8 are connected to the eighteenth to twenty second inputs of the controller 3,

the first to fifteenth outputs of the video controller 12 are connected to the first to fifteenth inputs of the display 9.

When a voice command is delivered by the operator, sound vibrations are generated, which are fed to the input of the microphone 1, where they are transformed into piezoelectric signals into electrical signals.

From the output of the microphone 1, the electrical signals are fed to the first input of the ADC 2, where they are converted into digital sequences of sixteen bit binary codes.

The synchronization of the ADC 2, processor 5, controller 3, RAM 4, DAC 10, DZU 7 is carried out by the clock generator 6. In accordance with the pulses of the clock generator 6 received at the second input of the ADC 2, from the second to the seventeenth outputs of the ADC 2 when controlling the controller 3 digital sequences of codes are transferred to the second ... seventeenth inputs - outputs of RAM 4 for their online storage and to the second ... seventeenth inputs - outputs of DZU 7 for long-term storage.

To train the phoneme recognition system, phoneme standards are created. For this, the digital code sequences corresponding to the speech commands spoken by the operator are transmitted from the second to the seventeenth inputs - the outputs of RAM 4 to the second ... seventeenth inputs - the outputs of the processor 5, where with the help of the corresponding program they are converted to the form of a graphic function in time, which transmitted through the video controller 12 to the display 9.

The operator, analyzing the video information presented on the display screen 9, corresponding to the spoken command, carries out manual marking, which consists in setting the boundaries of the phoneme as part of the team.

When fixing the boundaries of phonemes, information about the number of samples goes to the eighteenth ... twenty-second inputs of the controller 3 and through the second ... seventeenth inputs - the outputs of the controller 3 goes to the second ... seventeenth inputs - the outputs of the processor 5, where it is processed, and then transmitted on the second ... seventeenth inputs of the DAC 10, where they are converted into an electrical signal from the output of the DAC 10 to the input of the headphones 11, where the electrical signal is converted into an audio signal.

The operator listens with a headphone 11 for an audio signal corresponding to the highlighted phoneme, and uses the keyboard 8 to enter the composition of the command and the letter corresponding to the highlighted phoneme. Through controller 3, to the second ... seventeenth inputs - outputs of RAM 4, the numbers of discrete samples of the beginning and end of phonemes and the corresponding letter codes, as well as commands containing phoneme borders are accumulated and accumulated.

The accumulated information for long-term storage is transmitted to the second ... seventeenth inputs - outputs of the DZU 7.

For the formation of phoneme standards in the processor 5, the program selects from the DZU 7 digital sequences of codes based on information about the phoneme's letter code. This information goes to the second ... seventeenth inputs - the outputs of the processor 5, where according to the program using the FFT algorithm, an array of values of the spectrum modulus is calculated.

Using the array of values of the spectrum module as an intermediate operation, the processor 5 calculates an array of values of the bispectrum module, examples of graphic images of which for phonemes A, C, B are shown in FIG. 2, FIG. 3, FIG. 4, (see the method of the invention), isolated maxima bispectrum modulus (IMB) and define their positions in the plane of the frequency f _1, f _2, form etalons phonemes that represent cumulative bispectral matrix signs, and transmitted from the second to seventeenth inputs - outputs of the CPU 5 for long-term storage the second ... the seventeenth Inputs - Outputs 7 DZU.

Information about the end of the formation of phoneme standards is received at the first ... fifteenth inputs of the display 9.

Before recognition of phonemes, the operator enters using the keypad 8 the threshold value for the probability of the presence of IMB, which, passing through controller 3, arrives for storage at the second ... seventeenth inputs - outputs of RAM 4.

When recognizing phonemes, the operator uses the keypad 8 to enter a command, in which it is necessary to recognize the phoneme. This information goes to the eighteenth ... twenty-second inputs of the controller 3 and then to the second ... seventeenth inputs - the outputs of the DZU 7, from which the graphic implementation of the selected command is received, coming through the video controller 12 to the first ... fifteenth inputs of the display 9.

The operator for the graphic implementation of the speech command selects the phoneme that needs to be recognized, and uses its keyboard 8 to enter its alphabetic code, which through the controller 3 is transmitted to the second ... seventeenth inputs - the outputs of RAM 4 for storage. The digital code sequences corresponding to the phoneme selected by the operator are processed, including the recognition procedure, in the processor 5 according to the program (see the method of the invention).

The letter code of the recognized phoneme in the processor 5 is compared with the letter code of the recognized phoneme coming from RAM 4, in order to evaluate the accuracy of recognition of phonemes. Alphabetic codes of recognizable and recognized phonemes, as well as a decision on recognition accuracy, are transmitted from the second to seventeenth inputs - the outputs of the processor 5 to the first ... sixteenth inputs of the video controller 12, where they are converted into a video signal, which is transmitted from the first to the fifteenth outputs of the video controller 12 for display on first ... fifteenth display inputs 9.

To perform the inventive device uses the following standard elements:

microphone 1 - compatible according to AC'97 standard;

analog-to-digital converter (ADC) 2 - implemented as a chip according to the AC'97 standard, known from [6, p. 50];

controller 3 - implemented in the form of an EVA-X1630C microcircuit manufactured by ADVANTECH, known from [6, p. 49];

random access memory (RAM) 4 - is implemented as an SDDIMM chip [6, p. 50];

processor 5 - standard chip INTEL CELERON 400 [6, p. 50];

clock generator 6 - is part of the controller 3 ;.

long-term storage device (DZU) 7 - made in the Compact Flash standard, the memory capacity is at least 64 MB [6, p. 205];

Keyboard 8 - PS / 2 compatible

display 9 - compatible with the SVGA standard;

digital-to-analog converter (DAC) 10 - implemented as part of a microchip according to the AC'97 standard [6, p. 50];

earphones 11 - compatible by AC'97 standard;

video controller 12 - microchip SMI 721 [6, p. 51];

computing system 13 - the SOM-ETX4400 module in the form of a printed circuit board [6, p. 50].

Using the invention will improve the accuracy of recognition of phonemes of speech.

Information sources

1. EPP patent No. 420825 class. G 10 L 5/06, published in 1991

2. Patent RU No. 2184399 class. G 10 L 15/00 // G 10 L 101: 02, published in 2002

3. US patent No. 4852170 class. 381/41, published in 1989

4. Lohmann A.V., Wirnitzer B.V. The correlation function of the third order // TIIER, 1984, T. 72, No. 7.

5. V.V. Latyshev, I. S. Ryzhak. Application of moments, cumulants and high-order spectra in modern signal processing methods. - M .: MAI. 1998.

6. ProSoft 9.0. Short product catalog 2003/2004.

Claims (2)

1. The method of recognition of phonemes of speech, which consists in receiving audio signals corresponding to speech commands, and converting them into electrical signals in a microphone, in converting electrical signals into digital code sequences in an analog-to-digital converter (ADC), characterized in that the digital code sequences obtained by analog-to-digital conversion of electrical signals are transmitted in accordance with the clock pulses coming from the clock generator into random access memory your (RAM) and long-term storage device (ROM) computer system, the conversion of digital sequences of codes corresponding to speech commands into graphic functions in time is carried out in the processor, after which they are transferred to a video controller to create phoneme standards for conversion into a video signal and displayed on the display, in a digital-to-analog converter (DAC), digital sequences of codes are converted into electrical signals, in headphones, electrical signals are converted into sound such signals, the operator performs video analysis of the information on the display, and audio analysis of the audio signal in the headphones with the aim of manually highlighting the boundaries of the phoneme as part of the speech command, entering the composition of the speech command and the letter code of the selected phoneme using the keyboard and transferring this information for storage in RAM and DZU, in the processor, the program processes digital sequences of codes corresponding to the selected phoneme intervals, including the calculation of the array of values of the spectrum module, determined by the fast cerned Fourier transform converts the digital code sequence corresponding to the selected interval of the phoneme in the array range of values of the modulus | F (jf) |, where f - frequency corresponding to the investigated range of speech further conducted calculation array bispectrum values of the modulus | S (f _1, f ₂ ) | as the product of three values of the spectrum modulus by the formula | S (f ₁ , f ₂ ) | = | F [-j (f ₁ + f ₂ )] | · | F (jf ₁ ) | · | F (jf ₂ ) |, where f ₁ , f ₂ are the frequencies corresponding to the studied speech range, and as the values of the spectrum modulus | F [-j (f ₁ + f ₂ )] |, | F (jf ₁ ) |, | F (jf ₂ ) | the values from the previously calculated array of values of the spectrum module are used, when processing the array of values of the bispectrum module, the maxima of the bispectrum module (IMB) are extracted, the positions of the IMB are fixed, then, repeating the operations of calculating the array of values of the spectrum module, calculating the array of values of the bispectrum module, extracting the IMB and the determination of their positions on the frequency plane f ₁ and f ₂ for all phonemes identical in alphabetic code from the total volume of speech commands, the phoneme standards are formed, which are aggregate IMB matrix at the points of the frequency plane, after which before recognition by the operator, the threshold value for the probability of MMB used in recognition of phonemes is entered into the RAM, when the operator recognizes the speech command spoken by the operator, the phoneme being recognized is selected, the alphabetic code of the recognized phoneme is stored in RAM, the processor calculates the array of values of the spectrum module of the recognizable phoneme, calculates the array of values of the bispectrum module for the recognizable phoneme , highlighting in the resulting IMB array and fixing the positions of IMBs, deciding on a recognized phoneme to maximize the ratio of the number of recognizable phonemes coinciding in the IMB position with the IMB of each of the phoneme patterns to the total number of IMBs of each of the phoneme patterns, while the probability of the presence of IMBs in the patterns is higher or equal to the threshold value previously determined by the operator, comparison in the processor of the alphabetic code of the recognized phoneme and the alphabetic code of the recognizable phoneme extracted from RAM, the decision on the accuracy of recognition of the background m of speech and transmission through a video controller to display alphabetic codes of recognizable and recognized phonemes on the display, as well as decisions about the accuracy of recognition of phonemes of speech. 2. A speech phoneme recognition device containing a microphone and an ADC, the microphone output being connected to the first ADC input, characterized in that it additionally includes a computer system including a clock generator, controller, RAM, and a processor for generating bispectral features and recognition based on them, phonemes of speech, DAC, DZU, video controller and ADC, which is made with sixteen outputs, and the keyboard, display, headphones are also included in the device, and the output of the clock generator is connected to the second input the ADC house, with the first input of the controller, with the first input of the processor, with the first input of the RAM, with the first input of the DAC, with the first input of the DZU, the first to sixteenth outputs of the ADC are connected to the second to seventeenth inputs - the outputs of the processor, from the second to seventeenth inputs - RAM outputs, with the second to seventeenth inputs - the outputs of the ROM, with the second to seventeenth inputs - the outputs of the controller, with the second to seventeenth inputs of the DAC, with the first to sixteenth inputs of the video controller, the DAC output is connected to the headphone input, from the first to the fifth keyboard outputs ry connected to the eighteenth to twenty second inputs of the controller, from the first to fifteenth outputs of the video controller are connected to the first to fifteenth inputs of the display.

Images