JPS6228480B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention captures a command based on an input voice signal, that is, a time series of physical quantities extracted from a speaker's voice wave as a characteristic pattern, and compares this with a pre-registered pattern to recognize a command based on a voice signal. The present invention relates to a speech recognition device using a pattern matching method, and more particularly to a speech recognition device in which a means for sampling a speech signal and a means for determining similarity between patterns are provided with means for preventing erroneous recognition of speech. In general, there are two methods for speech recognition: one is to directly calculate the similarity between a feature (input) pattern obtained after extracting some feature from the speech signal and a registered pattern that has been registered in advance, and the other is to directly calculate the similarity between the feature (input) pattern obtained after extracting some feature from the speech signal and a registered pattern that has been registered in advance. There are two main types of methods: a method in which this is later replaced with a phoneme sequence, and this is compared with a pre-registered word dictionary (pattern) to calculate degree of similarity. Of these two methods, the latter is superior in speech recognition when there are many words because it identifies phoneme units. However, when the number of words is not so large, the former pattern matching recognition generally provides a higher recognition rate. An example of a speech recognition system using pattern matching, in which the number of words to be recognized is about several dozen, is used in consumer equipment, for example, when a television receiver is controlled by voice.
In other words, for controlling the television receiver's power supply, volume control, channel switching, etc., the voice of the words representing the control content is registered in the voice recognition device in advance, and the voice is stored in the response device as a recognition response. This is a case in which the voice command is compared with the registered control contents, and if they match, a voice response is sent to indicate that the control contents have been recognized, and a predetermined control is performed. For example, when selecting channel 1 in channel switching control, if you have previously memorized the voice ``1 channel'' as a registered pattern, then turn to the microphone that receives voice commands and issue the voice command ``1 channel.'' A voice response says "OK" and channel 1 is selected. However, the problem here is that when the voice command ``1 channel'' is given, the voice command ``8 channel'', which has a similar sound, is registered as a control pattern (registered pattern). That is, the sounds of both "ichi" and "hachi" are similar, and the problem is how to prevent the mistaken speech recognition of "ichi" and "hachi".
This is because in the words "ichi" and "hachi", the vocal energy of the pronunciation part of "chi" is large, making it difficult to distinguish between "i" and "ha". Generally, when there is accented speech in one word, the speech energy is concentrated in that part, making it difficult to recognize other parts of the speech. Therefore, when performing voice recognition, it is necessary to compare the characteristic (input) pattern and the registered pattern without losing information on parts of the voice command other than the strong sounds. Further, the voice used when registering the control content as a registered pattern by voice does not necessarily match the rate of voice emitted as a voice command. This means that after a certain word is registered, even if the word is uttered in the same way again, the word length will be different. Therefore, when evaluating the similarity between the input pattern and the registered pattern, erroneous recognition will occur unless the time axis is also taken into account. The present invention solves the above-mentioned problems and provides a control means for setting the speech sampling period to a period longer than the pitch interval of normal speech. The purpose of the present invention is to provide a speech recognition device that prevents erroneous recognition. Hereinafter, typical embodiments of the present invention will be described with reference to the drawings. In speech recognition using the so-called pattern matching method, speech recognition is performed by determining the degree of similarity between an input pattern indicating the characteristics of the input speech and a registered pattern in which predetermined words are registered. The false recognition rate depends on the method of extraction. Therefore, in the present invention, when extracting features of input patterns, amplitude normalization and time axis normalization for input audio are performed to prevent misrecognition, and furthermore, similarity calculation between both patterns is simplified. become FIG. 1 is a block diagram showing a speech recognition device according to the present invention, in which a speech input device has a function of converting sound pressure vibrations caused by vocalization into electrical signals using a microphone and flattening the frequency distribution of the speech. Part 1,
A feature extraction unit 2 extracts the features from the audio signal converted into an electrical signal obtained by the audio input unit 1, and a feature extraction unit 2 stores the features extracted by the feature extraction unit 2 and calculates a comparison between them and the input pattern. It has a recognition processing section 3 that performs processing and discriminates voice control commands, and a voice response section 4 that responds to the voice to indicate that the control command has been recognized is added as necessary. This voice response section includes a memory section 401 that stores words to be responded to as patterns;
Second I/O (input/output) port 402, control unit 4
03, D/A converter 404, low pass filter 4
05, and the voice circuit of the controlled device, such as the television receiver 406, responds with voice to indicate that the speaker's voice command has been recognized. In the audio input section 1, input audio is converted into FM waves by a wireless microphone 11, received by an FM receiver 12, and inputted to a preamplifier 13, and inputted by a microphone 14 provided before the preamplifier. It can be incorporated into the system in either form. In either of these forms, the SN ratio, which is the ratio between the audio signal power necessary for recognition and other acoustic signals, is mainly influenced by the directivity of the microphones, so the microphones 11 and 14 are Use a directional one. The audio signal obtained by the preamplifier 13 and converted into an electric signal has its high frequency range emphasized by the pre-emphasis circuit 15 in order to improve the intelligibility of monosyllables. The output of the voice input section 1 obtained in this manner is supplied to the feature extraction section 2, where the feature data necessary for forming the input pattern is extracted. That is, it captures the frequency of the speaker's voice waves in time series, analyzes the frequency of the voice, samples this data at regular time intervals, and converts the sampled analog data into digital quantities using an A/D converter. . In other words, the feature extraction unit 2
Switched capacitor bandpass filters ( _hereinafter referred to as
It is called BPF. ) are connected. This 16(1)
The center frequency of the BPF of ~16 ₍₁₅₎ is determined by the applied clock, and each filter characteristic is a 6th-order Tievisiev characteristic and has an attenuation characteristic of approximately -36 dB/OCT. The band of approximately 200Hz to 6.4KHz is divided into 15 bands at 1/3 octave intervals by the BPFs 16 _{(1) to (15)} . 16 that passes the audio signal of the band components of these 15 bands.
Sample and hold circuits 17 (1) _{to (15 ) that} sample and hold signals at approximately 20mSec intervals are connected to each of the BPFs (1) to (15). Voice features are extracted. In this way, sample and hold circuit 17 _{(1
) to (15)} are analog quantities, for example, an 8-bit A/D converter (analog -
It is converted into a digital quantity by a digital converter) 18. At this time, switching control between the sample and hold circuits 17 _{(1) to (15)} and the A/D converter 18 is performed by a multiplexer 19. Therefore, the A/D converter 18 obtains a digitized amount of the time-frequency-level characteristics shown in FIG. 2 extracted from the audio signal. and,
The voice feature data extracted by the A/D converter 18 is supplied to the recognition processing section 3 via a first I/O (input/output) port 20 . The recognition processing unit 3 includes a registration pattern memory for storing and registering the voice characteristics extracted from the command voice when instructing control contents, such as specifying a channel to receive or controlling power on/off, by voice. 21. Input pattern memory 22 for storing the characteristics of the instruction voice as an input pattern when the speaker utters the desired control content;
A system program memory 23 that stores a program for determining whether the contents of this input pattern memory are similar to any registered pattern stored in the registered pattern memory 21, a CPU that executes the contents of this system program ( (Central processing unit) 24. And this CPU2
4 is an 8-bit microprocessor, for example, and the system program memory 23 is
It consists of a ROM with a capacity of 2K bytes, and includes the input pattern memory 22 and the registered pattern memory 21.
is composed of RAM with a capacity of 10K bytes. Of this 10 Kbyte RAM, 1.75 Kbyte is used as the input pattern memory 22 and approximately 7.5 Kbyte is used as the registered pattern memory 21. The data extracted by the feature extraction section 2 is sent to the recognition processing section 3 having such a configuration as input pattern data and registered pattern data. First, a case where registered pattern data is sent will be described. When the registered pattern data is sent to the registered pattern memory 21 of the recognition processing section 3, as described above, the speaker registers several desired control contents in the speech recognition device by uttering the desired control contents. Here, now 1
Regarding the case where channel selection is stored as control content in the registered pattern memory 21,
The audio characteristic of “1 channel” is the A/D described above.
The data is extracted as digital data by the converter 18. This data is then transferred to the first I/O port 2.
0 to the registered pattern memory 21, but at this time, the input pattern memory 22 has the following information:
It is temporarily stored in the form of the matrix shown below. m: Number of samples. n: Number of filters. a _ij ; digitized sample value; Here, the number of rows of the determinant is the number of samples, that is,
Indicates the number of times the output of the switched capacitor bandpass filter 16 is sampled in response to sample pulses at intervals of approximately 20 mSec, and the number of columns is
The number of BPFs 16 is shown, and each component is a digitized sample value of each BPF. In this way, the characteristics of the speaker's voice extracted have not yet been normalized to the amplitude information of the voice, and the information on soft sounds may recede depending on the position of the speaker's accent or strong sounds. Since no processing is performed on the voice, it cannot be said that the characteristics of the speaker's voice are sufficiently expressed. Therefore, the components of each row of the determinant are weighted. That is, the CPU 24 performs the following calculation stored in the system program 23 on the data once stored in the input pattern memory 22, represented by 〓 above, and the determinant 〓 of the calculation result is stored in the registered pattern memory 21. is stored as a registered pattern. In this way, the amplitude information of the audio information is normalized. This amplitude normalization is performed on all sounds uttered by the speaker as control content, and then the content (determinant) is stored in the registered pattern memory 21. In this way, when the speaker registers the desired control contents in the registered pattern memory 21 by utterance, the setting of the control contents for the speech recognition device is completed, and registration patterns of types equal to the number of control contents (〓 ₁ , 〓 ₂ ...〓
_o ) is stored in the registered pattern memory 21. As mentioned above, the calculation for normalizing the amplitude with respect to the pattern 〓 representing the characteristics of the voice is executed by the CPU 24 according to the program contents stored in the system program 23. This will be explained schematically. That is, the A/D converter 18 in FIG.
1st I/O port 20, system program 2
3. The operation of the CPU 24 can correspond to the functional operation shown in FIG. 3 below. In other words, latch circuits 30(1) _{to (15)} in FIG.
(It actually corresponds to the input pattern memory 22.)
, data corresponding to the determinant 〓 is latched, and the latched contents are supplied to an adder 31 and a multiplier 32, respectively. And this adder 31
The output of the level judgment circuit 33 and the divider 34(1) _~
(15) is supplied. The adder 31 adds the elements of each row component of the determinant,

【formula】

【formula】

The row component elements latched in the latch circuits 30(1) _{to (15)} are divided by the respective sum values in the dividers 34(1) _{to (15)} . Here, multipliers 32(1) _{to (15)} are provided before the dividers 34(1) _{to (15)} to perform N multiplications, but this is because the result of the division is evaluated in the integer part. This may be omitted in some cases. Further, the data whose amplitude has been normalized by being divided by the dividers 34(1) _{to (15)} is stored in the registered pattern memory 21 as a registered pattern through the bus line. Further, a predetermined level threshold is set in the level determiner 33, and when the level of the output of the adder 31 is below the set threshold, the latch circuits 30(1) _{to (15)} and The latched contents of 35(1) _{to 35(15)} are cleared, and the two latch circuits are not controlled at other times. In this way, the latch circuit 3
0(1) _{to (15)} and 35(1) _{to (15)} are the adder 31
By latching only when the output is above a certain value, malfunctions due to noise can be prevented when the detected voice is small. As can be seen from the explanation of FIG. 3 above, in the process of registering the control content desired by the speaker in the registered pattern memory 21, the characteristic pattern before the amplitude is normalized is partially composed of RAM. The characteristic pattern stored in the input pattern memory and then having its amplitude normalized is stored in the registered pattern memory. Next, a case will be described in which the speaker instructs the desired control content by voice with respect to the registered control content. When the speaker utters a desired control content among the registered control content and gives a voice command, the characteristic pattern of the voice is normalized in amplitude and recorded in the input pattern memory 22 in the same way as the registered pattern. Here, it is assumed that an input pattern obtained by normalizing the amplitude of the content of the voice command given by the speaker is expressed by the following determinant. This amplitude is normalized and input pattern memory 22
The input pattern 〓 stored in is referenced with the registered pattern already registered in the registered pattern memory 21 as the control content. By calculating the similarity between both patterns using this reference operation, it is determined that the control content corresponding to the pattern with the closest similarity is the control content commanded by the speaker. The degree of similarity between the input pattern and the registered pattern is determined by calculating the distance between the patterns shown below. That is,
The determinant obtained by taking the absolute value of the difference between the registration pattern whose amplitude has been normalized, the input pattern, and each element _kij , _ij is defined as the determinant distance pattern, which represents the distance between both patterns. , the similarity is calculated based on the total value of each element of this determinant 〓. To further explain this, the distance pattern 〓 is expressed by the following equation, and the similarity d is expressed as follows. The above calculation of the degree of similarity d is performed for all registered patterns, or in other words, patterns representing all control contents, and the pattern with the smallest value of degree of similarity d is the pattern commanded by the speaker by voice. It is determined that Speech recognition is performed in this way,
By normalizing the audio amplitude as described above, the misrecognition rate is significantly reduced. In this way, the speech recognition of the speaker's utterance is performed by calculating the degree of similarity between the registered pattern and the input pattern by the CPU 24 executing calculations instructed by the similarity calculation program set in the system program memory 23. It becomes possible to control devices using voice recognition. However, in speech recognition using the speech pattern matching method described above, the speech misrecognition rate is reduced by normalizing the amplitude, but even if a speaker utters the same word, the utterance time is not always the same. It does not match. To solve this problem, it is necessary to normalize the time axis as well. The time axis is normalized by always dividing the time taken between the pronunciation start time and the pronunciation end time of a word pronounced by the speaker by a constant constant n. This constant n
In this invention, is a constant larger than the pitch interval time of the sounds produced by the speaker. This can be done by sampling the input audio at a period longer than the pitch interval to avoid missing the peak value detection during the sampling period, which not only normalizes the time axis but also reduces the normalization error for the amplitude value. This means that it can be suppressed. As mentioned above, extraction of the characteristics of the speaker's voice in both input patterns and registered patterns is performed by both the BPF 16(1) _{to (15)} and the sample/hold circuits 17(1) _{to (15).} Although dependent, both circuits have a time constant element in their operation. In particular, the peak detection method of the sample-and-hold circuit greatly influences the correct detection of the end time of the speaker's utterance. Therefore, the peak detection method and sampling timing in the sample-and-hold circuit that constitutes the feature extraction unit 2 are important points in normalizing the time axis after accurately capturing the speaker's utterance length. . Next, another embodiment of the feature extracting section 2 suitable for making correction of the time axis will be described. Generally, when a speaker utters a certain single sound (speech waveform shown in Figure 4a), the output of the BPF 16(1) _{to (15)} has a pitch between peak values of P as shown in Figure 4b. The wave number can be obtained. For example, this pitch P is approximately 8 mSec when a single sound such as "a" is uttered, but in normal speech, this pitch is 5 to 5 mSec.
Enter within 15mSec. The outputs of the BPFs 16(1) _{to (15)} shown in FIG. 4B having such a pitch P are each subjected to peak detection as shown in FIG. Depending on the constant, the end time of the utterance may be detected incorrectly as shown in FIGS. 4d and 4e. In other words, when the time constant is increased to reduce ripples caused by peak detection, the detection output changes at time t ₂ even though vocalization actually ends at time t ₁ , as shown in Figure 4 (d).
It is recognized that the audio continues until the end. On the other hand, if the time constant is made small, ripples will occur in the detected waveform, making it impossible to expect accurate feature pattern extraction. This may affect the normalization of the time axis and the extraction of feature patterns, leading to incorrect speech recognition. Therefore, in this embodiment, this problem is solved by performing peak value detection at a cycle longer than the pitch cycle. The present embodiment will be described below with reference to the drawings. FIG. 5 is a circuit diagram showing another embodiment of the feature extraction section 3 shown in _FIG .
The audio signal from the audio input section 1 (not shown) is input to
It is supplied to BPF41(1) _{to (o)} . And this
Each output of BPF41(1) _{to (o)} is a diode D(1)
_~(o) , a sample/hold circuit 42(1) having a peak detection function, a MOS transistor Q(1) ~ _(o) constituting _~(o) , and a capacitor C(1) _~(o) that holds the peak value. ₎ is used to detect the peak. The peak value detected by the peak detection, that is, the audio amplitude data is held in the capacitors C(1) _{to (o)} , and these amplitude data are sent to the binary-decimal decoder 43 and the MOS transistor Q'( 1) is supplied to the A/D converter 45 via a multiplexer 44 consisting of _(o) . Here, the MOS transistor Q(1) _~(
configure said multiplexer when _o) is on;
The MOS transistors Q'(1) _{to (o)} are in an off state, and are controlled so that when one transistor group is on, the other transistor group is off. Therefore, the MOS transistors Q(1) _{to (o)}
When the MOS transistor Q is on, the audio amplitude data held in the capacitors C(1) _{to (o)} is transferred to the MOS transistor Q.
(1) When _~(o) is off, the MOS transistor
The signals are supplied to the A/D converter 45 via Q'(1) _{to (o)} and converted into digital quantities. The sampling of the peak value is performed for a time T longer than the time of the pitch P mentioned above, and after the peak value is held for the time T, the transistors T(1) _~(o) , T'(1) _{~( o)} ,
The capacitors C(1) _{to (o)} are reset by a reset circuit 46 composed of resistors R(1) to _(o) and R'(1) to _(o).
) is discharged. After this discharge time, peak detection starts again and is repeated until the end of the speaker's utterance. To explain this using FIG. 6 _{, FIG} . The peak value is held as it is detected, and the charges in the capacitors C(1) _{to (o)} are discharged by the reset pulse shown in c of the same figure, so the input of the A/D converter 45 is The waveform shown is input. As can be seen in Figure 6, the peak value of the voice is held for a time T longer than the pitch P mentioned above, and since the discharge is during the reset pulse period, the amplitude data of the detection output that is erroneous due to the discharge is A/D converted. There is no need to send it to vessel 45. Next, the fifth section regarding the means for generating a sampling pulse for holding the peak value and the means for generating a reset pulse for the time T mentioned above.
This will be explained using Figures 7 and 8. Said capacitor C(1)
A sampling pulse is obtained by a frequency divider 47 and a NAND gate 48 in order to sample and hold the peak value of the voice at _~(o) . That is, the clock pulse indicated by CK in FIG. 7 is applied to the clock terminal _CK of the frequency divider 48, and the clock pulse indicated by CK in _FIG . Figure 7 a
The sampling pulse shown is obtained. This sampling pulse is applied to the MOS transistor Q(1) _~(
The conduction of _o) is controlled as described above. Further, the first monomulti 49 detects the falling edge of the sampling pulse a, generates a pulse shown in FIG. 7b, and inverts the output of the flip-flop 50 (see FIG. 7c). Then, clock pulse CK' is applied to m-bit counter 53 via NAND gate 51 and inverter 52, and this clock pulse
The m-bit counter 53 starts counting CK' and sequentially switches the binary-decimal decoder constituting the multiplexer 44, and when all scanning is completed, the m-bit counter 53
A reset pulse is supplied from the output Q of the flip-flop 50 to the flip-flop 50 through the inverter 54, and the state of the flip-flop 50 is inverted again. and,
At the same time, the second monomulti 55 makes the transistors T(1) _{to (o)} conductive and the capacitors C(1) _{to (o)}
A reset pulse (corresponding to d in FIG. 7 and c in FIG. 6) is generated to discharge the charged charge. An initialization circuit 57 connected to the frequency divider 47 is used to reset the frequency divider 47 when the power is turned on. Further, the timing of reading data into the A/D converter 45 is determined by generating the pulse shown in FIG. 8f as follows. As mentioned above, at the falling edge of the sampling pulse (Figure 7a), the first monomulti 49 generates a pulse (Figures 7 and 8b). This pulse inverts the state of the flip-flop 50 (FIGS. 7 and 8c), and the m-bit counter 53 receives the clock pulse.
CK' (Fig. 8e) is applied. The falling edge of this clock pulse (Fig. 8e) is detected by the third monomulti 53, and the output of this third monomulti 53 has the signal shown in Fig. 8e. A pulse is generated.
This pulse is then used as a data read timing pulse for the A/D converter 45. In this way, in this embodiment, the time T, which is larger than the above-mentioned pitch P observed when uttering a single sound, is used as the sample time for extracting speech features, and erroneous feature extraction during speech recognition due to ripples during peak detection is prevented. do. Furthermore, when determining the end time of a speaker's utterance, the error range can be set to a range that is approximately smaller than the pitch length P, so that erroneous recognition can be reduced when normalizing with respect to the time axis. In other words, even if the time required for a speaker to utter the same word differs each time the speaker utters the same word, the rate of speech misidentification caused by this can be reduced. As is clear from the above invention, according to the present invention, it is possible to provide a speech recognition device in which malfunctions in speech recognition using the pattern matching method are significantly reduced. Note that the device to be controlled by the voice recognition device according to the present invention is not limited to television receivers, but can be applied to general systems that require remote control.

[Brief explanation of the drawing]

FIG. 1 is a circuit block diagram showing an embodiment of the speech recognition device according to the present invention, FIGS. 2 and 3 are characteristic diagrams and circuit block diagrams for explaining the present invention, and FIG. FIG. 5 is a circuit block diagram showing another embodiment of the present invention, and FIG. 6 is a waveform diagram explaining the operation of another embodiment of the present invention. 7, and 8 are timing charts for explaining the operation of other embodiments of the present invention. 16(1) _~(o) , 41(1) _~(o) ...Filter, 17
...Peak value detection means, D(1) _~(o) , 42(1) _~(o
) , 44, 46...Peak value detection means, 3...Recognition processing unit.

Claims (1)

[Claims] 1. A plurality of filters for extracting predetermined frequency band components from the speech generated by a speaker, and a sample pulse having a predetermined period longer than a normal speech pitch interval for each output of the filters. a plurality of peak value detecting means for detecting peak values at each peak value; an A/D converter for converting the peak values detected by the peak value detecting means into digital values; and the A/D converter and the peak value detecting means. and a reset means for controlling the peak value detecting means so that the output of the peak value detecting means is once reset during the non-sampling period and then sequentially detecting the peak value during each of the sample periods. and a timing control circuit that drives the peak value detection means every sample pulse of a predetermined cycle and drives the reset circuit and the A/D converter during a non-sampling period; 1. A speech recognition device comprising at least a recognition processing unit that recognizes features of speech in response.