FR2691829A1 - Speech recognition system using neural network and fuzzy logic processing - Google Patents

Abstract

The speech recognition system receives an input from a microphone (10) that is amplified (20) and received by an analyser (30). The analyser output is transferred via an interface (40) to a personal computer (60) complete with hard and floppy discs (50), keyboard (70) and display screen (80).The speech analyser divides the speech signal into predetermined frequency bands in which energy levels are defined. The personal computer received the defined energy signal levels and compares the signal levels of adjacent bands. A binary data value is then generated for identification. The processing is based upon a neural network and fuzzy logic theory.

Description

Speech recognition system.

  The present invention relates to a speech recognition system, and more particularly to a Korean monosyllabic speech recognition system. When features are extracted for speech recognition, the perception of an intrinsic characteristic used to distinguish spoken syllables is very difficult as a result of speech rate, accent, and pronunciation habits of a given speaker , environmental variations during the speech, the emotional state of the speaker, etc. Moreover, as a result of the variations of the articulated language, even a phoneme represents different phonetic characteristics which,

  are, in turn, affected by the above phonemes

  These factors cause difficulties

  to develop an algorithm to extract the characteristics

  particular characteristics of speech, and make expression

  -sion and the integration of knowledge obtained through

of the inconvenient algorithm.

  Research into the use of established systems has

  been largely done to solve the pre-existing problems

  In this way, we know different processes such as ana-

  formant lysis for placing vocal sounds according to a voice component, a dynamic time distortion (DTW) method in which a distortion between the individual articulations of a single word is decreased by means of a dynamic programming technique so that the most consistent speech can be selected for recognition, and a Hidden Markov Model (HMM) method for speech recognition

  by visualization of a signal composed phonetically.

  However, since most of the speech recognition systems heretofore realized using the aforementioned methods require a considerable amount of

  calculations to recognize a natural spoken human voice

  genuinely and accepting the different articulations of speech, their actual use and relevance is dubious and a recognition of speech in real time

  becomes difficult A network model -

  fuzzy set theory is proposed as

  process which can solve the general problems of recognition

  birth of configuration (for example, recognition

of speech).

  Unlike a Fbhn-Noimann calculator, the neural network model can not only learn the correct rules to resolve issues arising from ambiguous, incomplete and discordant data, but also parallelize a multitude of neurons, so that the neural network model can be used in a field requiring parallel processing, such as speech recognition The neural network model

  has three main advantages.

  First of all, it is highly adaptable

  to say that a human path is diversified according to the noise

  environment and the characteristics of a speaker who can

  In the second place, the learning procedure is reasonable While an algorithmic processing by correctly extracting an intangible characteristic from a variety of speech data is very difficult,

  the neural network model can extract itself the characteristic

  to learn through the appro-

study-by-example.

  Thirdly, the parallel processing

  above can be executed in the network model, the

  resulting from the many fundamental neurons

  rates that perform parallel processing, so that a huge amount of time is required for learning.

  weaving can be manipulated by parallel processing.

  In the established procedures, a specific reference configuration is determined beforehand, or an abundance of rules relating to speech data are programmed one by one while in the system using the neural network, since the externally revealed characteristics of information are learned configurations can be categorized without requiring specific instruction for feature variation information, and better performance can be

  be obtained with respect to a modified configuration.

  A representative neural network model that introduces

  actually the neural network as a de-recognizing process

  birth of speech to improve speech performance is a time-delay neural network (TDNN) that has shown performance in speech recognition by syllabic phoneme Also, an experiment in which a sub-network corresponding to a phonological group is formed as a module resulted in an extended range of recognized objects without

  lower the high recognition speed of the sub-

network.

  In order to actually apply and use suf-

  In addition to the characteristics of the neural network described above, a circuit realization is necessary but, unlike software simulation, this has many limitations. Therefore, an input multi-layered neural grating

ordered (IDMLP) is proposed.

  Also, since the frequency characteristic differs according to each speaker even though the same word is pronounced (i.e., due to speech rate diversity), the fuzzy set theory is introduced.

  One of the aims of the present invention is to

  to set up a system of speech recognition of mono-

  Korean syllables, using a neural network and an algo-

  rithm that adapts to a fuzzy configuration.

  To achieve the above object of the present invention, there is provided a speech recognition system comprising: speech input means for inputting a speech signal; speech analysis means for dividing the speech signal from the speech input means into predetermined frequency bands for expressing the divided speech signal as the energy level in each frequency band; and

  a host computer to enter the signal

  from the speech analysis means, digitizing the speech signal by comparing the energy level of each frequency band with respect to the energy level of a neighboring frequency band, and outputting externally

  a speech signal corresponding to digitized data.

  The aforementioned aims and advantages of this

  invention will be better apparent from the detailed description

  of a preferred embodiment thereof with reference to the accompanying drawings in which: Fig. 1 is a block diagram showing a circuit structure for performing speech analysis according to the present invention; Fig. 2 shows an embodiment an analog amplifier of Figure 1 according to the present invention; Figure 3 shows an embodiment of the speech analyzer of Figure 1 according to the present invention; Figure 4 shows a single-frame output signal of the speech analyzer relative to a spoken Korean vowel; Figure 5 is a flowchart showing speech analysis of the speech analyzer; Fig. 6 is a flowchart showing a step of extracting speech analysis data; Figure 7 shows the spectral response of a speech signal

  FIG. 8 represents a spectral response

  merited the speech signal of Figure 7; Figure 9 shows the steps of digitizing data made fuzzy;

  Figure 10 shows a structure of a

  modular neuronal bucket IDMLP; Figure 11 is an array of syllables to classify into monosyllabic consonants-vowels; and FIGS. 12A to 12E show a rate of

  respective classification with respect to each module.

  In Figure 1, a speech recognition system comprises: a microphone 10 for inputting speech; an analog amplifier 20 for amplifying the speech of the microphone 10; a speech analyzer 30 for analyzing a speech signal from the analog amplifier 20; an interface panel 40 to ensure

  the interface with a personal computer (PC); some

  hard disk drive and flexible drive mechanisms (HDD and FDD) for exchanging data with the computer a host computer 60 connected to the interface panel 40 and the HDD and FDD 50; a keyboard 70 which constitutes an input unit of the host computer 60; and a monitor 80 that

  constitutes an output unit of the host computer 60.

  The analog circuit for speech input as shown in Figure 2 is designed to adjust the volume of speech input, using a variable resistor. A differential amplifier 90 (eg, a TLO 72 chip) has a bandwidth. frequency above 10 KHz since normal human speech frequencies are up to 7 KHz. Since digital and analog signals are mixed in the analysis portion of

  speech, great attention must be paid to noise.

  In the present invention, such noise is eliminated by using a capacitor having a capacitance of about

  0.1 RF The output signal of the analog amplifier stage

  logic is used as an input signal for

  the speech analyzer 30 which delivers 8-bit data.

  Fig. 3 is a block diagram showing an embodiment of the speech analyzer 30 which is divided into two parts. One is an interface part comprising an 8-bit comparator 31 (for example a chip 74 L 5688). and an interface chip 32 (for example,

  8255 PPI), and the party performing the analysis of the

  For example, since chip 74 L 5688 is an open-collector type chip, a resistor 35 is

  mounted between the output (pin 19) and a power source

  (Vcc) Using the 8255 P Pl chip for

  kill the interface with the PC, the interface panel is designed and manufactured in a simple way The chips 8255 P Pl and p PD 7763 are controlled by means of the host computer

  , which will be described in more detail with reference to the software.

  The address and data signal input to the speech analyzer chip 33 uses data buses and address buses, and a 4 MHz clock is required to synchronize the operation of its internal circuit. which is solved by means of a crystal oscillator

at 4 M Hz 34.

  A RESTORE entry to the interface chip 32

  restores automatically when the calculator is

  tialized by connecting the bollard in one area

I / O of the host computer 60.

  The restoration of the speech analyzer chip 33 must be performed before setting an input mode, but the overall circuit becomes excessively cumbersome thereby operating. This problem is solved in the present invention by the use of the 8255P chip. Pl and the associated software In more detail, a value corresponding to the restoration signal is transferred along the data bus to the restoration terminal of the speech analyzer chip 33 via an output port of

the interface chip 32.

  When a FRAME frame terminal of the chip 33

  the speech analyzer delivers a logical "1" that represents

  At the completion of the analysis of a frame, the calculator reads the value of a series of 16-channel filters inside the chip of the speech analyzer itself. By using the interface chip 32 as the interface of this part, the pretreatment part requiring the most time to recognize speech is constituted

  by means of circuits thus constituting a system of recognition

  birth of speech in real time as a whole.

  So, although a significant amount of time is wasted to analyze speech signals in

  using a general software simulation, in the pre-

  In this invention, the output signal of a pass filter is

  16-channel band is obtained in the circuit, so that the time required is reduced to thus form a system that can be used for a recognition of

speech in real time.

  Table 1 below represents the characteristics of

  frequency ticks of the 16 bandpass filters in the

  of the speech analyzer 30 of Figure 1.

TABLE 1

  Figure 5 shows the output signals of a frame from the speech analyzer 30 (FIG.

  1) in relation to a Korean vowel, allowing observation of

  of the vowel trainer Thus, the system of analysis

  Speech lysis designed and manufactured according to this

  invention does not present disadvantages such as the requirement

  a pre-treatment panel.

  FILTER # FREQUENCY FREQUENCY FREQUENCY OF

CENTRAL CUT-OFF

(BASS) (HIGH)

  GMP 1 40 Hz 250 Hz 640 Hz GMP 2 120 Hz 450 Hz 820 Hz GMP 3 250 Hz 650 Hz 1060 Hz GMP 4 380 Hz 850 Hz 1240 Hz GMP 5 660 Hz 1050 Hz 1400 Hz GMP 6 850 Hz 1250 Hz 1650 Hz GMP 7 1080 Hz 1450 Hz 1940 Hz GMP 8 1210 Hz 1650 Hz 2140 Hz GMP 9 1450 Hz 1900 Hz 2500 Hz GMP 10 1510 Hz 2150 Hz 2800 Hz GMP 11 1810 Hz 2500 Hz 3280 Hz GMP 12 2200 Hz 2900 Hz 3720 Hz GMP I 3 2560 Hz 3300 Hz 4280 Hz BPF 14 2900 Hz 3800 Hz 4740 Hz BPF 15 3100 Hz 4500 Hz 6440 Hz BPF 16 3720 Hz 5400 Hz 7400 Hz The constitution of the software is twofold First of all, a control program allows a user to easily use the speech analysis panel designed and manufactured as in the foregoing Such a control program sets the operating modes of the interface chip 32 and the speech analyzer chip 33, and paces the input and / or the data output Second, another program detects a speech interval from the data interpreted as being in a region of specific equations after traversing the speech analyzer chip 33, normalizes differently articulated lengths, and finally digitizes data

  used as input to the IDMLP neural network.

  The determination of the mode of operation of the interface chip 32 is to designate a connection point

  which makes the entry and exit in an operating mode

  In the present invention, "PAO" and "P Bl" are set as output ports, and the "PBI" port is also designated as a port.

  The above operating mode is determin-

  born using C language software commands Output port b (Ox 307, Ox 82); / * Port A & B output, Port B input * / In the circuit structure, the interface chip 32 is used to initialize the port. speech analyzer chip 33 and check the completion of the speech analysis Therefore, the first program above

  to control the interface chip 32 is very simple.

  Then the second program above to compare

  it is necessary to call the speech analyzer chip 33.

  At this moment, the time required to analyze the data

  and read the analyzed data must be calculated carefully

  to ensure that the entire system is stable.

  The speech analysis chip must be initialized

  using the interface chip 32 by executing the commands

  the following C language: Output port b (O x OO); / * Restore signal p PD 7763 * / Delay (1); / * Restore Signal Time * / Output port b (O xll); * Free Restore Signal * / Since the restore signal initializes the

  system, the command above delay (int) becomes necessary.

  The input / output controller of the analyzer chip

  Speech Manager 33 controls the data operations of

  external trea / output, using the bus terminals

DBO-DB 7.

  Table 2 below is given to explain operations according to five control signals CS, WR, RD,

At O and A 1.

TABLE 2

  CS RD WR A O A 1 Operation O 1 O O PC CSR-0 (gain & frame duration) 0 1 O 1 PC CSR-1 (low cutoff frequency & on / off)

0 O 1 1 O CSR-0 -> PC

0 O 1 1 1 CSR-1 _> PC

  0 O 1 O First in-first out -> PC After the recovery signal is released, an operating mode must be set within 378 gs The operating mode of the analyzer chip

  33 is fixed by writing data on a re-

  control / status register (CSR) within the chip via the data bus, wherein the types of operating mode that can be controlled by the computer are: (1) the duration of a scanned frame , (2) the preamplifier gain, (3) the on / off function of an equalizer, and (4) the cutoff frequency of a

low pass filter.

  These four basic information are obtained

  naked by writing data on the CSR via the data bus inside the calculator Here, using the C language again, they can be obtained as follows: Output port b (Ox 304, Ox 4 c ) / * Od B, 16 ms * / Output port b (Ox 304, Ox 02); / * 25 Hz, OFF EQ * / Figure 5 represents a global flowchart

  from the speech input to the reading of the results

state-analyzed.

  Figure 6 shows a detection step

  the range of speech and the obtaining of nu-

  merited to use as an input of the IDMLP neural network, which is executed after storing the output signal of the speech analyzer in a memory of the

  computer via the interface chip 32.

  In the present invention, a frame length is set at 16 ms. The speech interval must be detected

  in completely analyzed data In the present invention

  when the energy level of a frame is higher than

  at a predetermined threshold voltage, the frame is determined

  equal to the length of the speech interval When a person repeats the same word several times, the lengths of the spoken words are rarely the same, so that a time axis normalization must be performed When a monosyllabic word is repeated, it generally lasts from 8 to 26 frames Thus, 15 frames are fixed as reference, the time axis normalization is performed, and the data obtained above

are digitized.

  In accordance with the present invention, in order to be used as input to an IDMLP neural network, the step from the input of the microphone to the output of the 16-channel filter series is performed in circuit, thereby save time collecting input data. The step of extracting the input data is to enter the data described above finally standardized and digitized, and is performed as below

  1 The speech signal is received through the

microphone.

  2 Using the circuit of FIG. 2, the received speech signal is amplified to a correct volume to be used as input to the speech analyzer.

word 30.

  3 The result of the analysis is read from the analyzer

of speech.

  4 The speech interval is detected, using

  the predetermined threshold value.

  The speech interval is normalized with respect to the time axis corresponding to the reference frame. 6 The output of each bandpass filter is compared to that of its adjacent filter, by digitizing

  thus their relative energy levels.

  Table 3 below represents the data resulting from the speech interval detection step above. 0 ú 1 Ol ú J 9 o Oz LZ LZ LL 91 IL O 6 11 0 O Il It 9 if 91 zi IL 8 V LI HI 8 O UL 09 811 OO 9 9 r 1V ZI LV Lú I 1 1 LL LIú Ol L 6 9 V VZ LE IL 9 V 8 VL Lz LI Seen L 98 911 0 t I Oz 81 Lt 61 LZ 6 Z 99 V 9 El 'ti Ol VI LL 6 9 Iú 0 ú 8 8 S LZ 9 z 91 01 9 V 8 V 8 U 9 L IL Voi t II Vi 9 t II UL O ú 1 Vz LZ oz IV zi Zú V 6 LIJI L LZ Iz if 91 zz PZ THE LE Il 1 VI iz 6 V 01 Z Il O ZI VZ Z 9 1 91 Vz Vg 88 97 Z oz LI 8 V tzzl II s I 8 9 z 9 z 1 Vt 9 z 91 Vi Z 8 ú 9 9 V 6 LOV 09 LZ 011 L zz 81 zz Lt 61 CI VV Bone Z 9 19 811 901 LJI 861 Z 91 1 it 61 Oz OL 9 V LC úZ 9 ú 9 V 19 IL 98 LO ZZI L 81 OLU 9 V 1 98 81 It 61 6 V t V OV çz 6 LZ: EL LAW M 6 99 V 81 91 Z OLI 8 L LI zi t I 1 z LL LL LL 8 Z LV 99 6 V Vt 011 O Vi 901 ZOI 91 ZI 91 8 Z 6 81 ZL O Lg OZ O i Z 69 69 L u z 8 9 I 0 z VI ZI 9 ú 1 01 LI OL Vz 61 LI LZ LL 6 L 1 gt I z OZ 6 61 91 8 OI LI oz 61 Lz 91 9 9 z OV Uu 91 O l z O 8 zz 8 i O 8 9 Il o O l I it lú ZI 81 ZI O º ZI LI LT 61 LV Ll 1 91 z Figure 7 shows the spectral response of the system, and is shown in Figure 7.

speech signal.

  Figure 8 shows a frequency spectrum

  Here, the energy level of each filter passes

  The read band of the speech analyzer is compared to the output of the filter whose band is immediately below, so that a logical "1" is set as the output of the filter when its output value is greater than that of the filter. of lower frequency (when Ef Nl 1 C Ef N) and a logic "" is fixed in the case of a lower value (when Ef N1> Ef N> Table 4 below represents the digitized speech output signal obtained using the process

  described above (Here, the number of input bits corre-

laying at one frame is fifteen).

  o I T O O O O O O O O I 1 O O O O I I O O T T T O O 1

I O O O T O O T O O I O O I 111

O O I I O O O O I I I O O O ZI It

O O O T I O O I I I O O O IIL

I I O O O I O O O I I O O O O L

O O O I O O O I I T IO O 61

  _ O O O I T O O I T I O O O 8 L 1 I O I O O O O O I O O I I Li

O O I I O O O I O O O I _ 9 L

I O O T O I O O I I O O I I I

O I T O O O O T I O O I I O IT

O O O O I O O O I O I _ O O 1

I t O O I I I O O I O O I ZL

O _ I O O O O O O O O O O TL

  | _I_ | bl Ia | Z If | _Ia I | 01 Jl 6 _ | 8 J_ LJ | 9 I_ t_ afa J _L ___ -

  Figure 4 shows the output signal of

  digitized speech obtained using the method described above.

  (Here, the number of input bits corresponding to

a frame is fifteen).

  In the present invention, as a step

  antecedent of Korean speech recognition monosyl-

  labics, fuzzy set theory and the ability

  previous adaptation of the IDMLP neural network are

  to recognize pronounced numbers from "O" to "9" (which are monosyllabic in Korean) Also, in the recognition using the IDMLP network, each syllable of the speech data is blurred with the learned result, so that the experiments recognition of the learned result by means of a syllable of speech data, and the possibility of combining the neural network and the

  fuzzy set theory is measured.

  Figure 9 represents the scanning step

data made fuzzy.

  Each syllable of the speech data obtained by pronouncing the numbers from "O" to "9" ten times is doubled

  and blurred, and the blurred data is

  digitized to an appropriate threshold value.

  Table 5 below represents the result

  from the above digitization of the data made fuzzy.

I O O I O O I O O O O O O SIL

O O O O I O I O O O O O O 17

  O O O I O I O O O O O O () I ú 11

  O O O O O O O O O O O I ú 11 0 0 0 t 0 t 0 0 0 0 t 0 0 0 l Et O O O O O O O O O O O 1 til

0 O O I O I O O O I O O I I 011

O O O I O I O O O I O O I I 6 1

0 O O I O I O O O I O O I I 81

0 O O I O I O O O I O O I I Li

0 O O I O I I O O I O O I I 91

  0 0 0 t 0 t t 0 0 I t 0 0 t I 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 I 0 0 0 I 0 0 0 I 0 0 0 0 0 Pl úId | bt 1 Zlà | J 14 01 J 6 J | 84 | L-4 | 9-q | d 5 q új | Zj | Evidence on Co O. Since the number of input nodes is regular during the learning of the network, the lengths of the sounds

  pronounced differently must be standardized

  In the present invention, using the time axis

  reading 15 frames as a reference, if the frame of a

  input configuration is longer than the frame

  priority, normalization along the time axis is achieved

  to progressively establish an appropriate frame interval.

  Here, 200 syllables of speech data uttered by a person are used as the study data of the neural network IDMLP, and the recognition experiment is performed by means of 100 syllables of speech data to be recognized which are extracted during three periods of

  different times (morning, noon and evening), respec-

  to determine the adaptability of the speech recognition system in accordance with this

  invention with respect to the diversity of speech data.

  After learning, in relation to the study data, both the digitized data and the data rendered

  blurs have a recognition rate of 100%.

  In addition, with respect to the test data, learning using digitized data and fuzzy data both has a high recognition rate (over 94%). These experimental results are

  represented on the following tables.

  When the IDMLP neural network is educated using digitized data, the recognition rate for the recognition experiment is 94% in the morning (Table 6A), 99% at noon (Table 6B) and 96% in the evening (Table 6A). 6 C),

  for an overall recognition rate of 96.3%.

TABLE 6 A

  X 1 "O" 1 "1" i "2" | "3 g" 1 "4" 1 51 "'" 6 "<<" 7 "" 8' "911 recognition

___ (%)

"O" 10 100

"1" 9 i 90

"2" 10 100

"3" 10 100

"4" '10 100

I "5" 10 100

1 "6" 2 7 1 70

"7" '2 8 80

"8" 10 100

"9" 10 100

TABLE 6 B

* i "O" "1" 1 "21 '3" 1 "411" 5 "161" 711 | 7 8 "9" ire onnaissance

"1" 10 100

"2" <10 100

"2" 10 100

"31 10 1000

"14" 1 0 100

the 5 "10 100

"6" 9 I 90

1 "7" q 10 100 g "8" l 10 100

19 "10 100

TABLE 6C

  On the other hand, when the IDMLP neural network is educated using fuzzy data, the recognition rate is 97% in the morning (Table 6D), 99% at noon (Table 6E) and 98% in the evening (Table 6 F), or an overall recognition rate of 98%. j f O "s" 1 "'211 tg 3" 1411 5115 "6" 7 "11" 8 11 9 Reconnaissance

: 0 "10 100" 1 "X 9 1 90

"2" 10 100

1 "3" 10 100

"4" 10 100

"5" 10 100

"6" 1 2 7 1 70

lt "7" 10 100

"8" 10 100

"9 19-10 100

TABLE 6 D

  | O "1" | 1 "" 2 "|" 3 "|" 4 "5" 1 "6" -Q 7 "|" 8 "1" 9 "recognition

"10 I O 100

t 111 "9 1 90

"2" 1 10 100

3 "1 10 100

11 "4" l 10 100

11 "511 10 100

11 "6" 1 8 1 80

11 "711 10 100

11811 10 100

"19" 1 '10 100

TABLE 6 E

  "011 111 11211 3 t" 3 11411 11511 11611 11711 "8" "119 l acknowledgment

___ (%)

t "O" 10 100 "1" k 10 100

211 10 100

3 "10 100

14 "11 10 100

t 5 10 100

1 "6" I 9 90

11711 10 100

11811 10 100

"19" 10 100

TABLE 6 F

  As can be understood from the tables

  above, the recognition rate at noon is the highest among the three moments of experiments, and the data with the worst recognition rate is the pronounced number "6" (in Korean). The learning of the IDMLP neural network is completed in one time with respect to the learning, either by the digitized data or by the data made fuzzy Although the structural characteristic of the network

  neuronal IDMLP has not been verified as a result of recognition

  birth does not make a big difference when learning is done using data made fuzzy, while the rate of recognition of numbers

pronounced is slightly increased.

  "1 | O" 1 l "1" "12" 3 "# 41" 5 "'" 611 "7 8" "9" | recognition

11 "O10 100

"Iit 9 1 90

"2" '10 100

11311 10 100

I 14 "1 10 100

1 "5" {10 100

11 "6" 11 9 90

"7" 10 100

1 "8" 10 100

"9" 1 -10 100

  For the purpose of applying the IDMLP neural network

  to the recognition of Korean monosyllabic words, a

  type classification neural bucket to classify the words into five vowel-oriented modules is first realized, and the total network consists of six modules so as to be finally recognized by each

  subnet by classified types Each module consists of

  killed by an IDMLP neural network, Figure 10 represents the

  modular IDMLP neural network structure.

  During the classification of type, using

  five vowel sounds as a reference, a sound of

  is classified into five groups The objective phonemes to be classified are 70 Korean monosyllables of a consonant-vowel structure (C-V) formed together with five

  Korean vowels, as shown in Figure 11.

  The data used to instruct the Y-neu-

  type classification which classifies the sound of

  in five groups are obtained by extracting only the vowel portion of the input sound, thus performing the learning In each case, the data used are of monosyllabic structure C-V and therefore, the vowels can be extracted simply Because a

  vowel is positioned at the end of the spoken word, the extrac-

  of several frames present at the end of the reference frames (for example 15 frames) is sufficient to be used to classify the types. However, if we consider the progressive extension of the objective words to be recognized, that is, say by considering a word having a monosyllabic consonant-vowel-consonant structure

  (C-V-C), the algorithm above is not suitable.

  Therefore, in the present invention, only the central portion of the entire frame is extracted for use as neural network type classification data.

  neural instructed using the data extracted as

  above, the type classification is tested in the morning, at noon and in the evening, as a recognition experiment.

pronounced digital birth.

  The study data uses data entered by extracting the vowel parts from 350 syllables of speech data (70 syllables formed together with the five

  vowels and each pronounced five times by one person).

  The test data uses about 420 syllables of data.

  Speech count for each of five modules, and a total of 2500 syllables, thus performing the module analysis test. Figures 12A-F represent a rate of analysis

  compared to speech data of a mono-

  syllabic C-V, representing, on average, a success rate

  type classification of 98.4% Classifi-

  respective type are measured at 97.6% for module "a", 98.6% for module "e", 98.1% for module

  "i", 98.3% for the "o" module, and 98.1% for the "u" module.

  Here, although the learning was done by extracting the data from the C-V monosyllabic speech data, a classification rate of greater than 90% is

  obtained by experimenting with a type classification

  port to monosyllabic C-V-C speech data.

  As a result, the speech recognition system according to the present invention can obtain a

  satisfactory result for the recognition of monosylla-

  Korean needs, as well as for a word recognition beyond this monosyllabic word, that is to say a word with multiple syllables, thanks to such an experience

  In addition, a new system of recognition

  birth of the word can be realized.

  Although the present invention has been shown

  and described in particular with reference to embodiments of

  of the latter, the person skilled in the art understood

  that different variations in shape and detail may

  be carried out without departing from the spirit and

scope of the invention.

Claims (2)

R E V E N D I C A T IO N S   1 speech recognition system characterized in that it includes -   speech input means (10, 20) for en-   generating a speech signal, speech analysis means (30) for dividing said speech signal from said speech input means into predetermined frequency bands for expressing the divided speech signal as a level of energy in each frequency band; and a host computer (60) for inputting the signal from said speech analyzing means, digitizing said speech signal by comparing the energy level   each frequency band in relation to the energy level   of an adjacent frequency band, and to deliver   a speech signal corresponding to the digitized data. -2 speech recognition system according to claim 1, characterized in that it further comprises interface means (40) for providing the interface between said speech analysis means (30) and said host computer (60).   A method of recognizing speech in a speech recognition system comprising speech input means (10) for inputting a speech signal   speech analysis means (30) for dividing speech   said speech signal from said means for   speech in predetermined frequency bands to express the divided signal as a level   of energy in each frequency band, and a calculation   host (60) to input the signal from said speech analyzing means, digitizing said input signal by comparing the energy level of each band   frequencies relative to the energy level of an adjacent frequency band, and externally   speech program corresponding to the digitized data, the   said method being characterized in that it comprises: a first step of inputting speech analysis data into said host computer; a second detection step in the speech interval, using data from said first step;   a third stage of execution of a standardization   along the time axis, using data from   from said second step; a fourth step of digitizing data that has been normalized along the time axis; and a fifth step of storing the result of said fourth step in a memory of said host computer.