JPH04136900A - Voice recognition processing method for voice input/ output device - Google Patents

Abstract

PURPOSE:To select a recognized word matching an input voice without re- inputting a voice by adding +1 to the contents of a misrecognition frequency memory corresponding to a recognized code number when the recognized number is selected and determined as a 2nd or succeeding candidate. CONSTITUTION:A voice input person confirms that his or her voice is displayed on a display 8 as a 2nd candidate, and then presses a cursor key 9b once. A CPU 1 reads the contents of a register R2 and moves a cursor to below the 2nd candidate. When a confirmation key 9a is depressed after the movement of the cursor to the 2nd candidate is confirmed, +1 is added to the contents of the misrecognition frequency memory M1 corresponding to the recognized code number. When there is no matching word among displayed recognized words, the recognized words are displayed on the display 8 in the decreasing order of frequencies of misrecognition. Consequently, the recognized word matching the input voice can be selected without re-inputting a voice.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a voice input/output device, and particularly to a voice recognition processing method.

(Prior Art) In general, a voice input/output device can be roughly divided into a voice recognition processing means for recognizing an input voice signal, an output processing means for outputting the result of the recognition process, a voice recognition processing means, and an output processing means. It consists of a control processing means for controlling. The speech recognition processing means converts the input speech signal into a speech pattern using a known technique such as linear predictive coding analysis method, and converts the input speech signal into a speech pattern using a known technique and N standard speech patterns registered in a standard speech pattern memory in advance. Matching is performed using a certain dynamic whale programming method, etc., and standard speech and speech with a high degree of similarity are
Select the recognition code number corresponding to the mother turn. The control processing means transmits data to an output processing means, such as a word processor, a speech printer, a speech synthesis translation device, guidance, control of electrical equipment, etc., based on the selected recognition code number.
Outputs a control signal. The output processing means input data,
Displaying characters, printing, producing sounds, operating machine tools, and turning on electricity based on control signals.

By the way, since speech includes unstable elements, the speech recognition processing means does not always select the correct recognition code number for the input speech signal, and may select an incorrect recognition code number. Therefore, the voice input person must check whether the selected recognition code number is correct. The checking method is to select n (N) recognition code numbers in descending order of similarity to the input audio signal, display the recognition words corresponding to these recognition code numbers on the display means, and then It was left to the person who entered the information to decide.

(Problems to be Solved by the Invention) In the conventional speech recognition processing method for a speech input/output device,
If the recognition word for the input voice is not selected as a candidate among the n recognition words selected in order of similarity, the voice input must be performed again and the process must be repeated until it is selected as a candidate. The problem was that it was difficult for people to use it.

The present invention provides a voice that allows you to select a recognized word that matches the inputted voice without re-inputting the voice when the recognized word for the inputted voice is not selected among the n recognition word candidates as a result of voice recognition processing. The purpose of this invention is to provide a voice recognition processing method for an input/output device.

(Means for Solving the Problems) In order to achieve the above object, in the speech recognition processing method for a speech input/output device of the present invention, N standard speech patterns are made to correspond to recognition code numbers assigned to them. A memory for the number of times of recognition errors is provided, and when the recognition code number of the standard voice pattern corresponding to the input voice signal is not determined as the first candidate among n, but is selected and determined as the remaining second and subsequent candidates, Or exclude selected n recognition code numbers (N
-n) When recognition code numbers are selected and determined as candidates in order of the number of misrecognitions from among the recognition code numbers, +1 is added to the content of the number of misrecognitions memory corresponding to the input audio signal. This is what I did.

(Function) In the speech recognition processing method of the speech input/output device configured as described above, the input speech is converted into a speech pattern, and when it is sequentially compared with N standard speech patterns, Recognition code numbers corresponding to n standard voice patterns are sequentially selected as candidates. Then, n recognized words are displayed on the display means. The voice inputter looks at the displayed n recognition words and selects the second recognition word that matches the input voice.
If it is below the candidates, a selection process is performed to determine it. Upon selection and decision, +1 is added to the contents of the erroneous recognition count memory corresponding to the input voice recognition code number. or,
If there is no match among the n recognized words displayed, recognition is performed in order of the number of misrecognitions from among the (N-n) recognized words, excluding the n recognized words already selected as candidates. Words are selected and displayed as candidates. By selection and decision, +1 is added to the contents of the erroneous recognition count memory for the recognition code number of the input voice.

Therefore, when the recognition word corresponding to the input speech is not selected from among the n recognition words, the recognition word matching the input speech can be selected without re-inputting the speech.

(Example) An example of the present invention will be described with reference to the drawings. Note that elements common to each drawing are given the same reference numerals.

FIG. 1 is a block diagram of an embodiment of the present invention. A main memory 2 (hereinafter referred to as memory 2) and a false recognition number processing means 3 are connected to the central processing unit 1 (hereinafter referred to as CPU 1) through path lines 12 and 13, respectively. Also, C
A speech recognition LSI 6, a standard speech A' turn memory 7, a display 8, an operation section 9, a speech synthesis processing section 10, and a speech synthesis data memory 11 are connected to the PU 1 via a pass line 14. The misrecognition number processing means 3 processes M1 to MNN1.
6, pieces. It consists of a memory for the number of misrecognitions and an adder 4, and each memory for the number of misrecognitions and the adder 4 are connected by a path line 15. Also, each misrecognition number memory and adder 4 are C
It is connected to PU 1 by the aforementioned pass line 13. The voice recognition LSI 6 and the microphone 5 are connected by a line 16. The memory 2 stores a control program for controlling the voice input/output device and display data of N recognized words corresponding to (1 to N) recognition code numbers.

For example, if 16 bytes are used per word and 64 words of display data are to be stored, a data storage area of 16×64=1024 bytes−1 kB is allocated. Then, the display data of the recognized words "Coffee '° Kocha Kan Coco Coa 6 Cola" Beer "Whisky" is stored in the data storage area in association with the recognition code numbers 1 to 64. An address is given by adding a fixed value X to the code numbers 1 to 64. Therefore, for example, the misrecognition number memory M1 stores the number of misrecognitions of the recognized word "°Koko" at the address "X+1". The mark voice/main turn memory 7 stores 64 recognized words in the voice of a voice input person as a standard voice pattern. The order of entering recognition words is “coffee” kocha “
The inputs are in the following order: Cocoa, "Axis Cola,"I' Beer, and II Whiskey. At this time, the CPU I assigns recognition code numbers to the standard voices and turns in order and registers them in the standard voice pattern memory 7.Voice The synthesis data memory 11 stores, for example, speech synthesis data of an English conversation sentence in correspondence with the recognition code number.The operation unit 9 has a confirmation key 9a and a cursor key 9b.

Next, the operation will be explained with reference to FIG. 2.

FIG. 2 is an operational flowchart of one embodiment.

Now, in order to order coffee from a waiter at a cafe in a foreign country, the voice input user inputs "coffee" through the microphone 5 in step Sl.Speech recognition processing is performed in step S2.The voice recognition LSI 6 inputs "coffee". The speech signal is converted into a speech pattern, and compared with the 64 standard speech patterns stored in the standard speech pattern memory 7, n words, for example, 3 recognized words "coffee" and "cola" are selected in descending order of similarity. Select the recognition code numbers "1", "4", and "2#" for the "cochia". In step S3, the recognition words "coffee", "cola", and "cocha" are displayed on the display 8. That is, the CPU 1 reads out the recognition code number with the highest similarity among the selected recognition code numbers from the voice recognition LSI 6, and stores it in the memory. The recognition code numbers 1 to 64 stored in 2 are compared, and the display data of the matching recognition word "coffee" is displayed on the display 8 as the first candidate.Similarly, pakola "kocha" is displayed as the second candidate. Displayed as the 3rd Marquis.

At this time, CPU J receives the recognition code number "1'" of each candidate.
"4" and "2" are stored at predetermined addresses. or,
The position data under the first character of each recognized word displayed is set as the cursor display position and is stored in register R1+ in CPU l.
Store it in R2+R3. That is, the cursor display position data under the U-shape of the first candidate "coffee" is stored in register R1, and the cursor display position data under each U-shape of "Cola" and "Kocha" is stored in registers R2+R3.
to be memorized. Also, the cursor position data is stored in register R4.
to be memorized.

In step S4, the voice inputter utters the voice "coffee".
is displayed on the display 8 as the first candidate, and presses the confirmation key 9a of the operation section 9 in step S5.

CPU J compares the contents of register R4 and register R1, detects that the first candidate has been selected since they match, and reads the recognition code of the first candidate stored at a predetermined address in memory 2. The number ``'1'' is sent to the speech synthesis processing unit 1o.
Output to.

Here, the predetermined address of memory 2 and register R1゜R2
+ To describe the relationship between R3+ and R4, the contents of register R4 are compared with the contents of registers R, R2, and R3 in order, and the contents of register R4 are compared to registers R1+ R2+
When the respective contents of R3 match, the recognition code numbers of the first candidate, second candidate, and third candidate are read from a predetermined address in the memory 2. In step S6+87, the speech synthesis processing unit 1° reads out the speech synthesis data for the input recognition code number '°1' from the speech synthesis data memory 11 and synthesizes the speech, for example, "A cup of co".
Output ffee please from Sugi Ichiriki, etc. By the way, in the speech recognition process of step S2, for the input voice "coffee", the recognition code numbers "4", "1" are given in descending order of similarity.
, "2" is selected, the display 8 displays "cola", "coffee", and "cocha"f as the first, second, and third candidates in step S3.

In step S4, the voice inputter utters the voice "coffee".
is displayed on the display 8 as the second candidate, and in step S8, the cursor key 9b of the operation section is pressed once. Read the contents of CPU 11d Renostar R2,
Move to register R4 and move the cursor under the second candidate "coffee". Step S9
The voice input user confirms that the cursor has moved to the second candidate and presses the confirmation key 9a in step SIO. Step S
1, the CPU 1 processes the number of misrecognitions. That is, C.P.
Since the contents of register R4 and register R2 match, U1 reads the recognition code number ``1'' from a predetermined register in the memory 2, and adds a certain value ``X'' to the address. Then, adder 4 adds the contents of erroneous recognition count memo IJM1 corresponding to address °'X+1''.
, add +1 and record the number of misrecognitions again.
Write in +. Finally, the recognition code number Pa1''' is output to the speech synthesis processing section 10.The process is then completed through steps S6 and S7.When the input speech is subjected to speech recognition processing as the third candidate, the above-mentioned step contents are repeated. S1□
, S13 are added.

By the way, as a result of the speech recognition process for the input voice "Pacoffee", if "Cola""Cocoa" are displayed as candidates in the order text play 8 in step S4, the cursor key 9b is pressed three times. The CPU 1 completes steps S8 and S9 when the cursor key 9b is pressed for the first time, passes through step S12 r sta when the cursor key 9b is pressed for the second time, and proceeds to step SI5 when the cursor key is pressed for the third time. Step SI
In step 5, the recognized words are displayed on the display 8 in the order of the number of misrecognitions that is the highest among the recognized words excluding the recognized word °'cola""cocoa". That is, the contents of the memory for the number of misrecognitions are read out and Sorting is performed in descending order of the number of misrecognitions using Create an array.

Next, transfer the largest misrecognition count value to register R1,
Compare with the contents of the erroneous recognition count memos IJM1 to IJM64. At this time, register R2 is used as a counter, and the initial value "
0" is stored. Then, +1 is added each time a comparison process is performed. This counter value corresponds to the recognition code number. In this comparison process, the contents of register R1 and the number of misrecognitions memo IJ Ml-M64 are stored. Find the recognition code number when the contents match.The found recognition code number and memory 2
The recognition code numbers "4", "2",
The recognition code numbers are compared with "3" and the recognition code numbers excluding those that match are stored in the data area of the memory 2. Similarly, from the misrecognition number array, set the next largest misrecognition number value to register R.
1, and obtain the recognition code number corresponding to the number of misrecognitions. The obtained recognition code number and the recognition code number stored at a predetermined address in memory 2 are
II and II3'I, excluding those that match, the remaining recognition code numbers are transferred to the memory 2 and stored following the recognition code number corresponding to the largest number of misrecognitions. Thereafter, in the same manner, a recognition code number array corresponding to the erroneous recognition frequency value array is created in the memory 2. The CPU 1 reads the display data of the recognition word corresponding to the recognition code number having the smallest sequence number in the recognition code number array from the memory 2 and displays it on the display 8. If the answer is correct in step S16, the process moves to step SIO. If the answer is not correct in step S16, the process returns to step S14. Step S14
When cursor key 9b is pressed in step 815, the CPU
1 reads the display data of the recognition word corresponding to the recognition code number with the next smallest array number from the recognition code number array from the memory 2 and displays it on the display 8. The process is then completed through the steps described above.

In this embodiment, a speech synthesis processing section is provided as an output processing means after speech recognition processing, but a word processor, a speech printer, a control section for electrical equipment, etc. may also be provided. Further, in this embodiment, a memory for the number of false recognitions and an adder are provided as the means for processing the number of false recognitions, but the CPU may be used as the adder and the data area of the main memory may be used as the memory for the number of false recognitions. .

(Effects of the Invention) Since the present invention is configured as described above, it produces the effects described below.

A memory for the number of incorrect recognitions is provided in correspondence with the recognition code numbers assigned to N standard voice patterns, and the recognition code number of the standard voice pattern corresponding to the input voice signal is the first candidate among the n pieces. When the recognition code number is not determined and is selected and determined after the remaining second candidate, or when the selected n recognition code numbers are excluded ((N=n) recognition code numbers have the highest number of misrecognitions) When recognition code numbers are sequentially selected and determined as candidates, +1 is added to the contents of the incorrect recognition count memory corresponding to the input audio signal, so that the number of recognition words for the input audio is n recognition words. If it is not selected in the list, you can select a recognized word that matches the input voice without having to re-enter the voice.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the configuration of an embodiment of the present invention, and FIG. 2 is an operational flowchart of the embodiment. DESCRIPTION OF SYMBOLS 1... CPU, 2... Memory, 3... Misrecognition number processing means, 4... Adder, 6... LSI for speech recognition
, 7...Standard audio pattern memory, 8...Display, 9...Operation unit, 9a...Confirmation key 9b...
Cursor key, 10...Speech synthesis processing section, 1no...
・Speech synthesis data memory.

Claims (1)

[Claims] The input speech signal is converted into speech patterns, and the recognition code numbers assigned to the standard speech patterns are n( In a voice recognition processing method for a voice input/output device that selects (n<N) standard voice patterns and determines one among them, a memory for the number of false recognitions is provided corresponding to the recognition code numbers assigned to the N standard voice patterns. , when the recognition code number of the standard voice pattern corresponding to the input voice signal is not determined as the first candidate among the above n, but is selected and determined as the remaining second or subsequent candidates, or when the selected n When the recognition code numbers are selected and determined as candidates in order of the number of misrecognitions from among (N-n) recognition code numbers excluding the recognition code numbers, the input audio signal is A speech recognition processing method for a speech input/output device, characterized by adding +1 to the contents of a recognition number memory.