JPH01112299A - Voice recognition equipment - Google Patents

Abstract

PURPOSE: To omit the execution of recognition processing for an unnecessary voice input by inputting a correct voice even after inputting unnecessary voice. CONSTITUTION: A buffer 3 is connected to the post stage of an input part 2 so that voice is temporarily stored in the buffer 3. When a candidate selection switch SW is depressed, only a syllable, paragraph or sentence most close to a current point of time out of syllables, paragraphs or sentences stored in the buffer 3 is recognized. Consequently only a necessary syllable, paragraph or sentence can be recognized after checking that a voice input is not noise, mispronunciation or an unnecessary chat.

Description

[Detailed description of the invention] [Table of contents] Overview Industrial field of application Prior art (Figures 25 and 26) Means for solving the problem to be solved by the invention (Figure 1) Working examples (Figures 2 to 24) (1) First embodiment (2) Second embodiment (3) Third embodiment (4) Fourth embodiment Effects of the invention C Summary] Regarding the speech recognition device, unnecessary The purpose is to avoid performing recognition processing on unnecessary voice input by inputting the correct voice even if there is voice input, and to analyze the input voice, extract characteristic parts, and compare it with a dictionary. A speech recognition device that performs recognition is provided with a storage means for temporarily storing input speech and a recognition instruction means for instructing recognition of the speech stored in the storage means, and a speech recognition device that inputs the speech to be recognized and performs the above-mentioned speech recognition. When the recognition instruction means is operated, the voice input portion immediately before the operation is extracted and recognized.

[Industrial application field]

The present invention relates to a voice recognition device, and particularly to a voice recognition device that can exclude undesired input such as coughing from the recognition target range when directly inputting a document by voice.

In a speech recognition device that inputs speech by dividing it into single syllables, words, phrases, or sentences, it recognizes it immediately after the speech input is divided, displays and outputs one of the most likely candidates of the recognition result, and then selects the candidate of the recognition result. Alternatively, you may need to choose between homophones, or you may need to prevent unnecessary sounds from entering your microphone, such as conversations with other people, coughs, or ambient noises. There is a need for a speech recognition device that can realize this.

[Conventional technology]

For example, in the speech recognition device of a speech input document creation device, in order to manually input the speech by dividing it into single syllables, words, phrases, or sentences, conventionally a manually operated switch was used as shown in Figure 25. .

In FIG. 25, 61 is an input section which receives audio input from a microphone, amplifies it to the required strength, and converts it into a digital signal.

Reference numeral 62 denotes a speech section detecting section, which detects a single syllable, word, phrase, or sentence unit based on the break in the speech input.

Reference numeral 63 denotes a recognition unit, which recognizes the audio input signal by referring to a dictionary (not shown). Reference numeral 64 denotes a candidate selection/homophone selection section, which selects another candidate when the first recognition result is a homophone and is not the desired word.

A display section 65 displays the recognition result of the recognition section 63 or other candidates selected by the candidate selection/homonym selection section. Switches SWI, SW2, and SW3 are manual switches operated by an operator. SWI is a voice input mode selector switch, which allows you to switch between a voice input mode that allows you to input voice manually, and a voice input mode that does not allow voice input, so that unnecessary sounds for document creation, such as conversations with other people, coughs, and ambient noise, do not enter the microphone. This is a switch for switching to non-input mode.

SW2 is a candidate selection/homonym selection switch, which is used to display other candidates by pressing this switch when the recognized result is not the desired one.
SW3 is for canceling an undesired input when an undesired input is made due to a mistake in speaking, coughing, or the like.

FIG. 26 is a flowchart for explaining the operation of the conventional human-powered voice document creation device shown in FIG.

As shown in the figure, when the first voice input is made, the voice section detection section 62 detects the section based on the delimitation, and the recognition section 63 performs recognition by comparing it with a dictionary. The recognition result is displayed on the display section 65. This recognition result is checked, and if it is correct, a second voice input is performed. Then, the previous recognition result is assumed to be correct, and the next second voice recognition and recognition process begins.

If the recognition of the result of the first voice input is incorrect, the candidate selection/homophone selection switch SW2 is depressed. As a result, a new candidate is displayed, and if this is the desired one, perform the next voice input.

[Problem to be solved by the invention]

However, in such conventional methods, in the voice input mode, voice segment detection processing is always performed, so unnecessary chatter and surrounding noise are not allowed, causing excessive tension to the speaker. There is a problem with this.

In addition, since the candidate selection switch for the recognition result is pressed or not pressed each time a voice is uttered, the timing of the utterance is inconsistent, resulting in poor operability and usability from a man-machine interface perspective. There is. Also,
It is very difficult to clearly separate each unit of utterance such as a word or phrase into utterances, and if you continue to input voice one after another, the utterances will become faster and faster, and eventually two utterance units will become continuous. However, there are problems in that it may be difficult to use, and may cause misrecognition.

The present invention has been made in view of these points, and an object of the present invention is to provide a speech recognition device that requires less operation of switches and allows manual work without worrying about noise, mispronunciation, etc. shall be.

[Means to solve the problem]

FIG. 1 is a diagram showing the principle of the present invention. In the figure, 1 is a microphone, 2 is an input section, 3 is a buffer, 4 is a voice section detection section, 10 is a recognition/candidate selection section, 8 is a display section, and SW is a candidate selection switch.

The uttered voice is converted into an electrical signal at the microphone 1, analyzed at the input section 2, and stored in the buffer 3 in 1k-times. The buffer 3 must have a minimum capacity to store the maximum length of the voice that can be input manually. The voice section detecting section 4 refers to the data in the buffer at the time when the candidate selection switch SW is pressed, and detects the voice section closest to the current time. The recognition/candidate selection unit 10 recognizes the voice section closest to this current point,
Furthermore, the correct result is selected from the recognition result candidates. The recognition results are displayed on the display section 8, and the correct result is selected while viewing this display. The buffers constitute a ring buffer, and old data is sequentially replaced with new data.

[Effect]

In this invention, a buffer 3 is provided after the human power section 2, and the voice input is temporarily held in the sofa 3, and when the candidate selection switch SW is pressed down, the voice input is held in the buffer 3. Since only the syllable, phrase, or sentence that is closest to the current moment is recognized among the syllables, phrases, or sentences, the necessary syllables, phrases, or sentences are recognized after confirming that the voice input is not noise, mispronunciation, or unnecessary chatter. Alternatively, it is possible to recognize only sentences.

〔Example〕

(1) First Embodiment A first embodiment of the present invention will be explained based on FIGS. 2 to 6.

FIG. 2 is a configuration diagram of main parts of a document creation device using the present invention;
FIG. 3 is a configuration example of the voice section detection section, FIG. 4 is a voice power curve diagram, FIG. 5 is an explanatory diagram of the operation of the first embodiment, and FIG. 6 is a comparison diagram of the operation of the conventional example and the first embodiment. be.

In FIG. 2, the same parts as in the principle diagram of FIG. 1 are given the same numbers, so detailed explanations of those parts will be omitted. In this embodiment, in addition to separately providing a recognition section 6 and a candidate selection/homophone selection section 7, the speech section detection section 4 is provided with a power calculation section 31 as shown in FIG.
, a high detection section 32, a memory 33, and a determination section 34.

In FIG. 3, the power calculation unit 31 is a part that calculates the power of the audio stored in the buffer 3, and as its output, a audio power curve according to time is obtained as shown in FIG.

The high detection unit 32 detects the voice power curve shown in FIG.
This is a part that detects areas a, b, and c that are equal to or higher than a predetermined threshold value 21. These areas a, b, and c are called island areas and correspond to locations where some audio input has been made.

The memory 33 stores the start and end times of each island area detected by the island detection unit 32. for example,
Start SI and end El of island area a (hereinafter a (S
abbreviated as l, El), b(32, E2), c
(s3, Ei) are stored respectively.

The determination unit 34 determines a voice section such as a single syllable, a phrase, or a sentence, which is one input unit, from the voice input shown by the voice power curve, and further, according to the present invention, switches S
Determine the voice section closest to the moment when the field is pressed. for example,
It is assumed that the switch SW is pressed down at time t2. Island regions a, b, and c are spaced at intervals j! , , I12 are both larger than a predetermined threshold Th/, each island area a, b
, c are determined to be independent voice sections, and only the voice section C, which is the island region closest to the switch S-field depression time t2, is sent to the recognition unit 6 as the voice section to be recognized. l5.12, are both smaller than the dark value THE, the combined area of island areas a, b, and c (a+b+C) is determined to be one voice section, and this is the highest at the time t2 when the switch S field is pressed. Since it is clear that they are close, this voice section (a + b + C) is sent to the recognition unit 6 as the voice section to be recognized.

Furthermore, when 12 is smaller than the threshold Thl and 11 is larger than the threshold Thz, the voice section is divided into island area a and island area (b+c).
It is determined that the area closest to time t2 (b
10c) is determined as the voice section to be recognized, and its address information is sent to the recognition unit 6. Based on this voice section information, a necessary area is read out from the buffer 3 and recognized.

Next, the operation of the present invention will be explained with reference to the operation flowchart of FIG.

In this invention, immediately after the voice is input from the microphone 1, the voice input is sent to the input section 2 without displaying the most likely recognition result candidate.
The data is analyzed and stored in the buffer buffer 3, and only the completion of storage in the buffer 3 is displayed on the display unit 8. That is, the operation will be explained with reference to the operation flowchart of FIG. 5. Voice input is performed (step 1), and when it is completed, a message to that effect is displayed (step 2). This display may be, for example, a blinking * mark on the display section 8.

Next, when the switch SW is pressed down, the voice section detection section 4
At step 4, the latest voice section is detected (step 4), but if the switch SW is not pressed down and there is voice input again, the process does not proceed to step 4 and returns to the beginning step 1. Therefore, for example, if you pronounce the pronunciation incorrectly at the beginning or cough, etc., without pressing the switch SW, wait enough time to determine the vocal section, and then try again with the correct pronunciation. All you have to do is enter.

In FIG. 4, if correct speech is input after incorrectly input speech sections a and b, this becomes speech section C and is stored in the buffer 3, so switch SW
When pressed, the voice section detection section 4 sends C, which is the latest voice section at this point, to the recognition section 6.

The recognition unit 6 reads the speech section from the buffer 3 based on this speech section information and performs recognition. It goes without saying that if the switch SW is pressed down immediately after only a single voice is input, that single voice is sent to the recognition unit 6 and recognized. Note that when the buffer 3 becomes full, it is sufficient to rewrite the input audio data one after another, starting from the oldest input audio data.

The speech sent to the recognition unit 6 is compared with a dictionary, and the highest priority candidates are output (step 5). The results are then displayed on the display section 8 (step 6). If the displayed result is correct, the next voice input is performed (step 7), and the process returns to step 1. If the result is a homophone but a different meaning, and is not the desired one, the switch SW is pressed down. Then, the next candidate will be displayed (Step 8, Step 6).

Repeat this operation, and when you get the result you want, perform the next voice input without pressing the switch SW, and you will return to step 1 and move on to the next voice input process. .

According to this invention, recognition is performed immediately after voice input, and without displaying the recognition result, the voice input is inputted to the bath sofa once, and only the latest voice section at the time of the switch operation is sent to the recognition unit. Since the system transmits and recognizes the correct voice, if the user makes a mistake in utterance, the correct voice can be input after a predetermined period of time has elapsed, and only the correct voice can be recognized by pressing a switch. Therefore, input can be made without worrying about not only erroneous pronunciations but also noise, coughing, etc.

In addition, it is possible to give a certain timing to the input by operating a switch for each vocal unit such as a word or phrase, making it easier to clearly separate each vocal unit and utter, so two vocal units will not be combined. , this eliminates misrecognition.

In addition, the user only has to operate the switch immediately after uttering the correct voice, ignoring noise and mispronunciation, and the operation is easy.

As shown in FIG. 6, which compares the operations of the present invention and the conventional example, when utterance 1 and utterance 2 are input, there may be noise or
It can be seen that when there is a speech error or a cough, the switch operation of the present invention can be minimized. Therefore, if a document is created using such a voice input recognition device, the voice input document can be created accurately.

(2) Second Embodiment A second embodiment of the present invention will be explained with reference to FIGS. 7 and 8. FIG. 7 is a diagram explaining the principle of the second embodiment, and FIG. 8 is a diagram showing the configuration of the embodiment. In FIGS. 7 and 8, the same parts as in FIGS. 1 and 2 are given the same symbols.

SWI is a switch for instructing voice section detection, SW2 is a switch for selecting candidates, and SW3 is a switch for deletion.

The uttered voice is converted into an electrical signal at the microphone 1, digitally converted at the input section 2, and then temporarily stored in the buffer 3. The buffer 3 needs to have a minimum capacity to store the maximum length of input audio. The voice section detecting section 4 refers to the data in the buffer at the time when the switch SWI is pressed, and detects the voice section closest to the current time. The recognition/candidate selection unit 10 recognizes the voice segment closest to the current time, and selects the correct result from the candidates of the recognition results.

The recognition result is displayed on the display section 8, and the correct result is selected while viewing this display.

Here, the switch SWI is used to instruct the extraction of a speech section immediately after utterance, the switch SW2 is used to select candidates or homophones for recognition results, and the switch SW3 is used to delete incorrect recognition results. It is for the purpose of

(3) Third Embodiment In the third embodiment of the present invention shown in FIG. 9, the switches SW1 and SW2 in the second embodiment shown in FIG. It simplifies operation. i.e. after only a single voice is input,
Immediately when the switch SW1 is pressed down, the voice is in the buffer 3, so the voice section detecting section 4 detects the voice section and starts recognition.At this time, there are no recognition result candidates yet, so the recognition candidate selection process is performed. will not be carried out.

Further, when there is no voice uttered and only recognition result candidates are selected using the switch SWI, since there is no voice in the buffer 3, the voice section detecting section 4 does not operate and only the candidates are selected. For this reason, switches SWI and SW
2 can be made common.

(4) Fourth Embodiment The fourth embodiment of the present invention shown in FIG. 10 is almost the same as the third embodiment shown in FIG. The difference is that a function to temporarily stop the operation of section 2' is added.

What differs among each component in each embodiment is the input section. The input section 2 of the first to third embodiments is the same, and its internal configuration is shown in FIG. 11. The input section 2' of the fourth embodiment is different from the others, and its internal configuration is shown in FIG.

In FIG. 11, an audio signal input from a microphone is input to an analog filter 20. In FIG. analog filter 2
0 is a low-pass filter having a cutoff frequency slightly lower than half the sampling frequency of the sample hold 21 in the next stage. The sample hold 21 quantizes the time axis of the audio signal that has passed through the analog filter 20 in accordance with the clock supplied from the AD conversion at the next stage. The AD conversion 22 quantizes the amplitude of the time-axis quantized audio signal, outputs the time series Dj of the audio digital signal together with its clock ckl to the next stage, and also outputs the lock required for sample and hold to the sample and hold 21. do. The clock 23 is necessary for the AD conversion 22, and a lock is generated using a crystal oscillator or the like.

In FIG. 12, the components 20.21.22.23 are the same as in FIG. 11. However, 24.25, and 2
With the signal from switch SW1,
The clock input to the AD conversion 22 is stopped for a certain period of time. 24 is a trigger circuit, which is implemented as a one-shot trigger circuit. 25 is a NOT circuit. 26 is an AND circuit, which serves as a gate that supplies an AD conversion clock to the AD conversion 22 only when the output of the NOT circuit 25 is high.

FIG. 13 shows a signal timing diagram of the circuit diagram of FIG. 12. First, when the signal X from the switch SWI is input to the trigger 24, the trigger 24 generates a pulse signal y with a width of several seconds. The operation of the input section will stop during these few seconds. The negation 2 of this several seconds wide pulse is the NOT circuit 2
5 and is supplied to an AND circuit 26 which is a gate circuit.

FIG. 13 shows the relationship between the gated AD conversion click W and each signal.

FIG. 14 is a diagram illustrating the configuration of the buffer.

Dj from the input section is transferred as is to the voice section detection section and is also supplied as write data to the memory section 301. Further, the clk from the input section is transferred as is to the voice section detection section and is also supplied as a count-up clock to the counter 300. A counter 300 counts write addresses in the memory section 301. On the other hand, the address control unit 302 receives iss, fee, 5t from the recognition unit.
I got b2. Immediately after the 5tb2 signal becomes 1, the address control unit 302 sequentially generates addresses from iss to ice together with the clock clkd. The generated address is used as a read address of the memory section 301. Data Dk and clkd read from the memory section 301 are sent to the recognition section 6 and used for recognition.

FIG. 15 is a diagram illustrating the voice section detection section 4, which is common to each embodiment. First, the power calculation unit 40 calculates the power of the digital audio signal read out from the buffer every few milliseconds, and temporarily stores the obtained power time series. The island detection section 41 reads out the power time series from the power calculation section 40 in accordance with the voice section detection instruction from the switch SWI, and detects islands. The determination unit 42 determines the interval between the obtained masterpieces and determines the final voice section.

FIG. 16 is a diagram illustrating the internal configuration of the power calculation section 40, which is common to all embodiments.

The power calculation unit accumulates the square values of the n pieces of audio digital data read from the buffer, and uses the accumulated value as the audio power. The audio digital data Dj obtained from the input section 2 or 2' and stored in the buffer 3 is stored in a square ROM.
(400) is input into the address field. Since the square value of the address value is stored in each address of the square ROM, the square of the audio digital data is obtained as the output data of the ROM (400). The adder 402 and the selector 403 constitute an accumulator, and the square ROM (400
) The squared values of the audio digital data obtained in step 2 are accumulated. The accumulated value is supplied to the address of the logarithmic value ROM (4'04). The logarithmic value of the cumulative value is obtained as data in the ROM (404). Logarithmic data is stored in temporary memory (406
) are stored sequentially. -The stored data Pi is read out by the island detection section 41 with address i specified. The clock ckl obtained from the buffer 3 is input to a clock frequency dividing section 401, and its frequency is divided into 1/n.

FIG. 17 shows the relationship between the clock ckl from the buffer 3 and the frequency-divided clock ck2. Divided clock C
k2 is first used to clear the accumulator. That is, it is supplied as a signal for the selector 403 to select the set value 0 without selecting the cumulative value of the output of the adder 402. Further, the clock ck2 is used as a clock for a counter that determines the address of the - hour memory, and is also used as a write signal for the - hour memory.

Next, the function and configuration of the island detection section 41 will be explained using FIG. 18, FIG. 19, and FIG. 20.

FIG. 18 is a diagram explaining the principle of island detection, and shows the contents of the temporary memory 406 in the power calculation section 40.

In FIG. 18, the horizontal axis shows address i, and the vertical axis shows data Pi.

The address l corresponds to the time axis of the audio.The island detection unit 41 detects a portion (island) where the data Pi is continuously large using the following method.Threshold value pth1 and Pt
h2(<P thl) is given in advance.

First, a portion (■■■) larger than P th2 is set as a temporary island. By doing this, the part A is removed as a noise part. Search is performed from the temporary island ■■■ until just before it drops below pth2 on both sides. As a result of the search, parts A and 2 are obtained as islands. The method described above is not suitable for hardware because the contents (Pi) of the -time memory are accessed randomly.

- An equivalent method for sequentially accessing the contents of the memory (Pi) will now be described.

First, the event Pi≦P th2 is α, the event Pth2<Pi≦Pt old is β, and pthi<p
The event i is defined as T1.

Next, as shown in FIG. 19, the four states 5o1S1, B
2. Consider B3. In this method, access to Pi is performed sequentially from the larger i to the smaller i. In FIG. 19, at the start time, the state SO is entered. Sequentially decrease i and repeat the state transition every time events α, β, and γ regarding Pi occur. If processing contents are assigned to state transition arcs, those processings are performed at the same time. The state transition diagram will be explained below using the example of power data in FIG. 17.

Processing is performed from the point marked * in FIG. 18 forward. In the present invention, the switch SWI is considered to have been pressed at this point. The state first goes to SO. At the point marked *, Pi is P th
Since it is smaller than 2, the event in this case is α. In other words, the state remains SO.

As l decreases, event β occurs, and the state changes from SO to 8.
Transition to 1. The i at this time is temporarily stored in an internal variable called STMP. Since the period β continues for a while, the state remains at Sl. Next, event γ occurs and the state transitions to B3. At this time, the contents of the previously stored STMP are stored in the internal storage SR. After that, the γ interval continues for a while (■
part of the temporary island), the state remains at B3. Next, event β
occurs, and the state transitions to B2. Next, γ occurs and the state returns to B3 (temporary island part of ■). After that, event β occurs again, and the state transitions to 82, and then event α occurs, and the state returns to SO. Here, the value of i is set as internal variable EH
to be memorized. At this point, the SR and ER have the addresses at both ends of the island (A). As we proceed further, event β will occur,
The state transitions to S1 and the value of i is stored in STMP (
part A). However, since the next occurrence of event α causes the state to return to SO, it is not possible to find the island (A). Processing continues in the same way for the temporary island ■ and island (tsu) parts.

FIG. 20 is a hardware configuration diagram that realizes the island detection method described above.

In FIG. 20, 4111 is a clock generator;
A clock is generated from the moment the switch SW1 is pressed. A counter 412 is loaded with the value i' of a counter 405 inside the power calculation unit 4o as an initial value at the moment SWl is pressed, and is counted down according to the clock of the clock generator 4111. counter 4112
The value indicates the value of i on the horizontal axis in FIG. 18, which starts from the point marked * and gradually decreases. Based on this value i, the power calculation unit 4
The contents Pi of the temporary memory 406 of 0 are sequentially read out and supplied to BO1 of the comparator 4100 and B1 of 41.01. A threshold value P th2 is supplied to AO of the comparator 4100, and a comparison with Pi is performed.

A threshold value P thl is supplied to AI of the comparator 4101, and P
A comparison with i is made. The output of comparator 4100 where BO≦AO corresponds to event α. Bl>A of comparator 4101
The output of l corresponds to the event γ.

In the AND circuit 4102, BO>AO of the comparator 4100
The AND of the output of B1≦AI of the comparator 4101 is calculated, that is, an output corresponding to the event β is obtained.

Here, α, β, and γ never become 1 at the same time.

4103 and 4104 are flip-flops, which are used to store the states SO and S3 as shown in Table 1.

Table 1 Relationship between states and flip-flops 4105.4106.4107.4108.4109,
The state transition shown in FIG. 19 is realized by the elements 4110 and 4110.

4103 and 4104 are first reset when a pulse is received from switch SWI (not shown in the figure),
The status becomes SO. According to the state transition diagram, when event α occurs, there is always a transition to SO from any state, so α is O
It is connected to the reset input of 4103 through an R circuit 4108, and also to the reset input of 4104.

From the state transition diagram, when γ is 1, S is always present from any state.
3, γ is connected to the set input of 4104 and also connected to the set input of 4103 through OR circuit 4107.

Also, when β becomes 1 in state SO, the state transitions to Sl, so the AND circuit 4109 first detects the current state SO, and then the AND circuit 4105 detects the logical product of β and the output of 4109, and then OR with the output of circuit 4105
4103 is set through circuit 4107. This realizes the transition from SO to 81. Also, state S
3, when β becomes 1, the state transitions to state S2. Therefore, the AND circuit 4110 detects the current state S3, and the AND circuit 4106 detects the logical product of β and the output of 4110. OR circuit 4108 at output
4103 through. This results in state S3
A transition from state S2 to state S2 is realized.

4113.4114.4117.4119 and 4121
is a three-man-powered AND circuit, each of which detects the transitions from ■ to ■ in the state transition diagram of FIG.

The AND circuit 4113 detects the transition ■, and the AND circuit 41
14 detects the transition ■. The OR circuit 4115 detects the transition of ■ or ■. If a transition of ■ or ■ is detected, the value of i is stored in the register 4116 (ER).

The AND circuit 4117 detects the transition ■. If ■ is detected, the value of i is sent to register 4118 (STMP).
to be memorized.

The AND circuit 4119 detects the transition ■. If ■ is detected, the selector 4120 selects the register (STMP
) and store it in register 4123 (SR). The AND circuit 4121 detects the transition ■. ■
If detected, selector 4120 selects the value of i and stores it in register 4123 (SR).

OR circuit 4122 is AND circuit 4119 or 4121
is supplied to the flip-flop 4123. Flip-flop 4123 is reset by the signal from switch SWI, and is set by the output of OR circuit 4I22. The output of 4123 is connected to one shot trigger 4124. Switch SW by 4123 and 4124
Only one output from 4122 immediately before I is pressed becomes a write signal for register 4125.

The signals clk and stb of each section and the values ER3 and SR of the registers are supplied to the determination section 42 at the next stage.

FIG. 21 shows the internal configuration of the determination section. 420 is a counter, which is counted up by the clk signal of the island detection section 41 and cleared by the logical sum of the stb signal and the ie multiplied signal. s
The OR circuit 424 calculates the logical sum of the tb signal and the ie multiplied signal. The counter 420 counts the length (βφ in FIG. 18) from the detection of the end point of one island to the start point of the next island. When this length exceeds THE, the output of comparator 421 becomes 1. However, a flip-flop 425 and an AND circuit 426 are provided to prevent unnecessary output from the comparator 421 from being output as a 5tbl signal (strobe signal) to the recognition unit 6 before the end point of one island is detected. . The flip-flop 425 is reset by the signal from the switch SWI and set by the ie multiplied signal island detection signal). In other words, the output of the flip-flop 425 is small (one island is detected).
The output of the comparator 421 is gated by the output of the comparator 421.

The multiplier 422 converts the address of the temporary memory 406 into the address of the buffer 3 by multiplying the value of SR by 0 to return it to the address before thinning out, and sends it to the recognition unit 6 as isr. Similarly, the multiplier 423 multiplies the ERO value by 0 to set the address of the temporary memory 406 to the buffer 3.
, and sends it to the recognition unit 6 as an ier. ier is the start point address of the audio within the buffer 3, and isr is the end address of the audio within the buffer 3.

The recognition unit 6 recognizes isr and ie when 5tbl becomes 1.
Import r and start recognition.

FIG. 22 is a diagram showing the internal configuration of the recognition unit 6.

isr, ier, and 5tb1 from the voice section detection section 4 are transferred as they are to the buffer section 3 as iee, iss, and 5tb2, respectively. The audio data Dk and clock clkd read out from the buffer unit 3 using the iee, iss, and 5tb2 signals are transferred to the audio recognition unit 600. The speech recognition unit 600 recognizes the speech data Dk. The speech recognition unit 600 refers to the speech template stored in the speech Tenbrecht memory 601 during recognition. The recognition results obtained by the speech recognition unit 600 are obtained as first to several candidates. The recognition result candidates are transferred to the candidate selection/homophone selection section 7.

Next, the operation of the second embodiment of the present invention will be explained with reference to the operation flowchart of FIG. 23.

In this invention, immediately after the voice is input from the microphone 1, the voice input is digitally converted in the input unit 2 without displaying the most recent recognition result candidates, and then it is stored in the bass sofa 3 for a while. Then, only the completion of accumulation in the buffer 3 is displayed on the display section 8. That is, the operation will be explained with reference to the operation flowchart of FIG. 5. Voice input is performed (step 1), and when it is completed, a message to that effect is displayed (step 2). This display may be, for example, a blinking * mark on the display section 8.

Next time the switch SWI is pressed down, the voice section detection unit 4 will detect the latest voice section (step 4), but if there is a voice input again without SW1 being pressed down, Do not proceed to step 4 and return to step 1. Therefore, for example, if you make an incorrect pronunciation at the beginning or cough, you can wait enough time T to determine the voice section without pressing switch SW1.
Just open HN and then input the audio with the correct pronunciation. In FIG. 18, if the correct voice is input after the incorrectly input voice section (1), this becomes the voice section (A), so if the switch SW1 is pressed down, the voice section detecting section 4 detects this point. (A), which is the latest voice section at , is sent to the recognition unit 6 (in the case of βφ>THβ). In addition, immediately after only a single voice is input,
Needless to say, when the switch SW1 is pressed down, that single voice is sent to the recognition section 6 and recognized. Incidentally, when the buffer 3 and the -time memo IJ 4.06 become full, it is sufficient to rewrite them one after another in order from the oldest input audio data.

The speech sent to the recognition unit 6 is compared with a dictionary, and the highest priority candidates are output (step 5). The results are then displayed on the display section 8 (step 6). If the displayed result is correct, the next voice input is performed (step 7), and the process returns to step 1. If the result is a homophone but a different meaning, and is not the desired one, the switch SW2 is pressed down. Then, the next candidate will be displayed (step 8,
Step 6).

Repeat this operation, and when you get the result you want, input the next voice without pressing switch SW2, and you will return to step 1 and move on to the next voice input process. .

As shown in FIG. 24, which compares the operations of the present invention and the conventional example, when inputting utterances 1 and 2, if there is noise, erroneous pronunciation, or coughing, the present invention It can be seen that the number of switch operations is extremely small. FIG. 24 shows an example in which when a customer applied for a ticket for ABC Airlines Co., Ltd.'s flight XXXX bound for Osaka, the operator mistakenly applied for a ticket for EFG Airlines Co., Ltd. By not operating the switch, unnecessary words such as ``no'', ``departure'', ``yuki'', rEFGJ, ``excuse me'', and ``it's a flight'' are removed.
``Please tell me your name.''

”, “It’s Mr.”, “Please wait for a while.

” -・- is shown as an example of operation so as not to be recognized and processed.

〔Effect of the invention〕

According to this invention, recognition is performed immediately after voice input, without displaying the recognition result, the voice input is input into the buffer, and only the latest voice section at the time of switch operation is sent to the recognition unit. Since the system transmits and recognizes the correct voice, if the user makes a mistake in utterance, the correct voice can be input after a predetermined period of time has elapsed, and only the correct voice can be recognized by pressing a switch. Therefore, input can be made without worrying about not only erroneous pronunciations but also noise, coughing, etc.

In addition, it is possible to give a certain timing to the input by operating a switch for each unit of utterance, such as a word or phrase, making it easier to clearly separate each unit of utterance and utter it (this prevents two units of utterance from being combined. (This eliminates misrecognition.

In addition, the user only has to operate the switch immediately after uttering the correct voice, ignoring noise and mispronunciation, and the operation is easy.

[Brief explanation of the drawing]

Fig. 1 is a diagram explaining the principle of the present invention, Fig. 2 is a configuration diagram of the first embodiment of the present invention, Fig. 3 is an example of the configuration of the voice section detection section in the first embodiment, and Fig. 4 is a curve of speech power. Fig. 5 is an operation flowchart of the first embodiment, Fig. 6 is a comparison diagram of the operation of the conventional example and the present invention, Fig. 7 is a principle explanatory diagram of the second embodiment of the present invention, and Fig. 8 is a diagram of the present invention. 9 is a block diagram of the third embodiment of the present invention, FIG. 10 is a block diagram of the fourth embodiment of the present invention, and FIG. 11 is a block diagram of the first embodiment of the present invention.
Example - Example of configuration of the input section of the third embodiment, FIG. 12 is an example of the configuration of the input section of the fourth embodiment, FIG. 13 is a timing diagram of the input section of the fourth embodiment, and FIG. 14 is a diagram of the hash sofa section. 15 is an example of the configuration of the voice section detection section, FIG. 16 is an example of the configuration of the power calculation section, FIG. 17 is an illustration of the clock of the power calculation section, FIG. 18 is an illustration of the island detection state, Figure 19 is a state transition diagram of the island detection unit, Figure 20 is a configuration example of the island detection unit, Figure 21 is a configuration example of the determination unit, Figure 22 is a configuration example of the recognition unit, and Figure 23 is the second embodiment. FIG. 24 is an operation comparison diagram between the second to fourth embodiments and the conventional example, FIG. 25 is a configuration diagram of the conventional example, and FIG. 26 is an operation flowchart of the conventional example. ■-Microphone 2-Human power section 3-Buffer 4-Speech section detection section 6-Recognition section 7-Candidate selection/homophone selection section 8-Display section Patent applicant Fujitsu Ltd. Representative patent attorney Akira Yamatani\,7 ~・, "Iy Kuuku◇ \-I~\Male 1'Y'-"

Claims (6)

[Claims] (1) In a speech recognition device that analyzes input speech, extracts characteristic parts, and performs recognition by comparing with a dictionary, there is a storage means (3) for temporarily storing input speech, and a storage means (3) for storing the input speech. A recognition instruction means (SW) is provided which instructs recognition of the voice to be recognized, and when the recognition instruction means (SW) is operated by inputting the voice to be recognized, the voice input portion immediately before the operation is extracted and recognized. A voice recognition device characterized by: (2) A speech section detection means (4) is provided for extracting an island region from the speech stored in the storage means (3), and when the recognition instruction means (SW) is operated, the island region immediately before the operation is extracted. 2. The speech recognition device according to claim 1, wherein the speech recognition device is configured to perform recognition using a speech recognition method. (3) A display means (8) is provided, and the recognition instruction means (SW
) to display the recognition result and confirm the recognition of the correct input voice.
The voice recognition device described. (4) The voice input document creation device according to claim 1 or 3, wherein the document is created based on the result confirmed by the display means (8). (5) The speech recognition apparatus according to any one of claims 1, 2, and 3, characterized in that the recognition instruction means and the recognition candidate selection means are common. (6) The speech recognition device according to any one of claims 1, 2, 3, and 5, characterized in that the operation of the input unit is temporarily stopped when a recognition candidate is selected.