JPS61246797A - Voice response switch - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field 1] The present invention relates to a voice response switch, and more particularly to a sound 73f response switch that operates by recognizing human voice.

[Background Art] As shown in FIG. 6, a conventional upstream voice response switch includes a filter circuit 11 that passes an input signal in a frequency band corresponding to voice, and a level detection circuit 12 that detects the output level of the filter circuit 11. The level detection circuit 1 compares the output of the level detection circuit 12 with a preset reference value.
It is composed of a control circuit 13 that outputs a control signal when the output of 2 is above a reference value, and a switch element 14 that is opened and closed by the control signal, and the input level to the control circuit 13 is above the reference value. A device is provided that sometimes determines that the input signal to the filter circuit 11 is an audio signal.

In this circuit configuration, it is determined whether or not it is a voice only by determining the level of a specific frequency band. If it is an input signal, the switch element 14 will be activated even if it is a noise that is not a voice, resulting in a problem of malfunction. Also, even if audio is input, it will be switched to switch element 1.
Since the switch element 14 is actuated regardless of whether the sound is emitted for the purpose of actuating the switch element 4, there is an inconvenience that the switch element 14 may be actuated when the switch element 14 is not desired to be actuated. - For this reason, as shown in FIG. 7, the voice recognition device 15 is used to compare the control voice stored in advance in the storage unit 16 with the input signal, and when the two match, the switch element 3 is opened or closed. However, when targeting unspecified speakers, there is a problem that a long time is required for arithmetic processing for voice recognition, making it difficult to control the switch element 14 in real time. Moreover, with the current level of technology, there is a problem that the recognition rate is generally low and malfunctions are likely to occur. Furthermore, since increasing the recognition rate requires a large amount of information and calculation, there is a drawback that the processing time is further delayed. On the other hand, when targeting a specific speaker, it is necessary to register the user's own voice before use, and the work required to use it is troublesome.

[Object of the Invention] The present invention has been made in view of the above-mentioned points, and its main purpose is to eliminate multiple formants that are dominant frequency components that characterize vowels in speech. By extracting the information from An object of the present invention is to provide a voice response switch that can be used for a specific speaker.

[Disclosure of the Invention J Figure 5 shows an example of a vowel spectrum,
The dominant frequency component that characterizes a vowel, that is, the frequency component at the peak of the spectrum, is called a formant. Generally, a vowel has a plurality of formants, and they are called, in descending order of frequency, a first formant F1, a second formant F2, a third formant F3%, and so on. Among these formants, the contribution rate of the 17th orman F and the second formant F2 is high, and by using the @1 formant F and the second formant F2, vowels can be determined with a fairly high degree of accuracy.

Here, the Japanese vowel /a//i//u//e//o/ is shown on the F, -F2 vector plane with the first formant F1 on the horizontal axis and the second formant F2 on the vertical axis. , each vowel is represented by the range shown by the broken line in FIG. Formants vary considerably depending on the vocal tract length of each individual, and are expressed with a certain degree of spread on the F and -F2 planes, with the ranges representing each vowel overlapping each other at the edges. However, in general, the formants of five vowels uttered by the same person in the same environment are Fl-F.
It becomes approximately pentagonal on two planes, and the environment changes,
It is known that even if the speaker changes, the relative positional relationship of the five vowels, that is, the pentagonal shape, remains unchanged and moves in parallel. However, the relative position when the vowel changes, that is, the change vector, remains approximately constant regardless of the environment or speaker. In other words, if the vector component of the vowel /a/ is (800Hz, 1800H2) and the vector component of the vowel 10/ is (500Hz, 1000Hz), the component of the change vector from /a/ to 10/ is (-30
0Hz, -800Hz), and the components of the change vector remain approximately constant even if the environment or speaker is different.

Therefore, in the present invention, a plurality of vowels are made up in succession to form a control voice, and it is determined whether the input signal matches a preset control voice by monitoring the change vector between each vowel. , discloses a voice responsive switch that opens and closes a switch element when an input signal matches a control voice. In addition, in the following explanation, the first formant F
Speech recognition is performed using F1 and second formant F2, but in order to further increase the recognition rate, the third formant F may be used as the third component of the vector. -F, if each vowel is represented on a vector space, the overlapping portion between each vowel will be removed, so the detection accuracy will be further improved.

(Example) As shown in FIG. 1, an audio signal is input to a formant extraction circuit 1, and a first formant F1 and a second formant F2 are extracted. The output of the formant extraction circuit 1 is input to a control voice discrimination circuit 2, and when it is determined that the input signal matches a preset control voice, a control signal is output. The output of the control voice discrimination circuit 2 is input to the switch element 3, and when a control signal is input to the switch element 3, the switch element 3 is opened or closed.

FIG. 2 shows an example of the formant extraction circuit 1.

The formant extraction circuit 1 includes a bandpass filter group consisting of a large number of bandpass filters 111 to 11n each having a band of 200 Hz and different pass frequencies, and an analog/band filter group that converts the output signal of each bandpass filter 111 to Iin into a digital signal. It is composed of a differential conversion circuit 12 and an arithmetic circuit 13 including a microprocessor or the like that detects formants from the output level values of the bandpass filters 111 to lln.

Bandpass filters 111 to lln each have a frequency of 0 to 200Hz.
, 200-4008Z, 400-600 Hz
.

......, 2200-2400 Hz, ...
* and pass frequency bands are different from each other, and are set so that all frequencies in the audio band can be passed by combining the full band filters 111 to fin. The arithmetic circuit 13 is the 17th Orman) F, the 1127th Orman) F2. ! : is detected and outputs a phoneme change signal that determines whether the input voice has changed. Although formant extraction is performed in a /% manner depending on the circuit configuration, it may also be performed using a software method such as a linear prediction method.

FIG. 3 shows an example of the control speech discrimination circuit 2, which, when a phoneme change signal is input, generates a vector having the first formant F and the second formant F2 as components. A first vector holding circuit 22 for storing, and a first vector holding circuit 2 for storing a phoneme change signal.
A change vector is calculated by subtracting the vector stored in the second vector holding circuit 23 from the vector stored in the first vector holding circuit 22. a change vector calculation circuit 24 for calculating the change vector, a storage section 25 in which change vectors between adjacent phonemes in the control speech to drive the switch element 3 are stored in a predetermined order, and an output value of the change vector calculation circuit 24 and the storage section 25. The change vector of the input audio signal is compared with the set value stored in the storage unit 2.
a comparison determination circuit 26 that outputs a match signal when the change vector is within the set range of the change vector stored in 5; and a control signal generation circuit that outputs a control signal for opening and closing the switch element 3 when the match signal is input. It consists of 27. In the storage unit 25, change vectors between adjacent phonemes of the set control voice are stored in a form that allows a certain degree of error.

In other words, the permissible error range of the change vector is set in consideration of errors in the change vector due to individual differences and environmental differences.For example, the range of the change vector from /a/ to 10/ is (300± α1Hz.

800±ff2Hz), and α1
, a2 are set appropriately to adjust the sensitivity. Therefore, in the control speech discriminating circuit 2, each time a phoneme change signal is input to the control speech discriminating circuit 2, the change vector of the input speech signal is checked to see if it is within the tolerance range of the change vector stored in the storage section 25. If it is determined that the change vector between each phoneme of the input voice signal is within the set range of the change vector of the control voice stored in the storage unit 25, a match signal is output from the comparison determination circuit 26. It is output. Note that the parts of the control voice discrimination circuit 2 other than the storage section 25 can be constructed using the microprocessor 20.

(Operation) The operation will be explained below. First, when some audio signal is input to the formant extraction circuit 1, the formant extraction circuit 1 extracts the vector components of each input signal on the FIF2 plane, and generates a phoneme change signal at each point in time when the phoneme changes. . In the control voice discrimination circuit 2, when the 1#1 voice is input, first, the vector component of the first voice is stored in the first vector holding circuit 22. Next, when the second voice is input and a phoneme change signal is obtained, the vector component of the first voice stored in the first vector holding circuit 22 is input to the WS2 vector holding circuit 23, and the first vector holding circuit 22 stores the vector component of the fJ&2 voice. At this time, the change vector calculation circuit 24 calculates the component of the change vector from the amount of change between the vector component stored in the second vector holding circuit 23 and the vector component stored in the first vector holding circuit 22. Here, the setting range stored in the storage unit 25 and the component of the change vector as an output value of the change vector calculation circuit 24 are compared, and it is determined whether the change vector is within the setting range stored in the storage unit 25. The zero-order tJIa
When 3 voices are input, the vector components of the 2nd voice stored in the 11 vector holding circuit 22 are input to the second vector holding circuit 23, and t! The vector component of the J3 voice is stored in the first vector holding circuit 22, and the change vector calculation circuit 24 calculates the change from the 12 voices stored in the second vector holding circuit 23 to the plS3 voice stored in the first vector holding circuit 22. The components of the vector are calculated. This change vector is compared with the set range of the second change vector stored in the storage unit 25 in the comparison/judgment circuit 26, and the output value of the change vector calculation circuit 24 is within the set range of the change vector stored in the storage unit 25. It is determined whether or not.

The same operation as described above is repeated until the input signal stops, and when the change vector of the speech signal for all the input phonemes is within the setting range stored in the storage section 25, the determination circuit 26 determines that a match is found. A signal is output, and the control signal generating circuit 27 receives the coincidence signal and outputs a control signal.

The control signal is input to the switch element 3 to open and close the switch element 3. Needless to say, when the input signal is different from the setting range set in the storage section 25, the switch element 3 maintains its previous state.

By the way, the control voice is /i//e//a/ 10//u
The vowel string consists of a vowel string that starts with an arbitrary sound in a vowel loop that circulates in the order of //i/, and that goes around the vowel loop at least once. In other words, an example of the configuration of the control voice is / l
tegll 嘗O! Ut! //a+oyU fee: 5etl
l//otue! ! Possible examples include e1atOtuy I/. In this way, since the control voice is composed of a vowel string that goes around the vowel loop at least once,
/i/ to /e/ and /u/ to/. Since there is no change in which the change vectors become approximately equal like the change between / and all change vectors are different from each other, there is no misrecognition of the input audio signal, and the recognition rate is low. This will improve the results. Also, in normal conversation, 5
Words with consecutive vowels include ``Blue House'' and ``Aiueo,'' but it is rare for five vowels to occur consecutively along a vowel loop. It is something that will not malfunction depending on the words used in the conversation.

In the above embodiment, a vector in a two-dimensional space is used with the first formant F and the second formant F2 as vector components to detect a vowel, but higher formants higher than the third formant F1 can also be used as vector components. By using this, vowels may be determined using vectors in a multidimensional space of three or more dimensions. Further, in the above embodiment, the formant extraction circuit 1 and the control voice discrimination circuit 2 each have a microprocessor, but the microprocessor of both circuits 1 and 2 may be shared and integrated into one. Furthermore, the vowel loop is F, −
It forms a pentagon on the F2 vector plane, and since it is detected that the input audio signal changes along this vowel loop, the judgment is made using the fact that it forms a pentagon on the F, -F2 vector plane. This makes it easy to determine whether the audio signal matches the preset control audio.

[Effect 1 of the Invention] As described above, the present invention includes a formant extraction circuit that extracts at least a first formant and a second formant from an input speech signal, and a formant extraction circuit that extracts at least a first formant and a second formant from an input speech signal, and a formant extraction circuit that extracts at least a first formant and a second formant from an input speech signal, and a It consists of a control voice discrimination circuit that outputs a control signal when the formant changes are in a predetermined order and the amount of change is within a predetermined range, and a switch element that is opened and closed by the control signal, and the control voice is /i/ /
Since the structure is such that the starting sound is any sound in the vowel loop that circulates in the order of e//a//o//u//r/, and the vowel loop is configured to go around at least once, the vowel in the voice By extracting the formants, which are the dominant frequency components that characterize the vowel, we determined whether to activate the switch element by moving the speech vector in a vector space formed by multiple formants. It requires only a change to be detected, the amount of calculation is small, and speech recognition can be performed reliably, and it has the advantage of being able to operate in real time and having a high recognition rate. Furthermore, since speech is recognized based on changes in formants, it has the advantage that it can be used for any type of speaker.

In addition, since the control voice is composed of a vowel string that goes around the vowel loop at least once, /i/ to /e/
Since changes such as the change from /u/ to 10/ in which the change vectors are almost the same do not occur, and all change vectors are different from each other, misidentification of the input signal does not occur. This will further improve the recognition rate.

[Brief explanation of drawings]

@ Figure 1 is a block diagram showing an embodiment of the present invention, Figure 2 is a block diagram showing a formant extraction circuit used in the above, Figure 3 is a block diagram showing a control speech discrimination circuit used in the same, and Figure 4 is a block diagram showing a control speech discrimination circuit used in the same. The figure is an operation explanatory diagram showing an example of F+F- diagram,
FIG. 5 is an operation explanatory diagram showing an example of vowel frequency characteristics, FIG. 6 is a block diagram showing a conventional example, and FIG. 7 is a block diagram showing another conventional example. 1 is a formant extraction circuit, 2 is a control speech discrimination circuit, 3
is a switch element.

Claims (1)

[Claims] (1) A formant extraction circuit that extracts at least a first formant and a second formant from an input speech signal; It consists of a control voice discrimination circuit that outputs a control signal when the amount of change is within a predetermined range, and a switch element that is opened and closed by the control signal, and the control voice is /i//e//a//o//
A voice response switch characterized in that the voice response switch is configured to have an arbitrary sound in a vowel loop that circulates in the order of u//i/ as the starting sound, and to make at least one round of the vowel loop.