JPS6152698A - Voice recognition equipment - 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] [Industrial Application Field] The present invention relates to an apparatus for inputting a human voice signal and recognizing data expressed by the voice, and in particular, it relates to a device for inputting a human voice signal and recognizing data expressed by the voice. ” and other meaningless words mixed in to prevent misrecognition.

[Conventional technology]

As one way to improve the man-machine interface,
2. Description of the Related Art Speech recognition devices that recognize speech that is divided into words and pronounced have reached the stage of practical use in recent years. -For example, in a bank's cash card service,
2. Description of the Related Art Devices have been put into practical use that recognize a card's personal identification number based on voice input from a depositor when inquiring about the personal identification number of the card.

Such voice recognition devices are used in different ways depending on the field of use.
Standard patterns of speech spectra for various specific words (for example, numbers, etc.) are held in advance. When a voice signal is input, the voice recognition device spectrally analyzes the voice signal to obtain spectrum data, and sequentially performs a comparison operation between the spectrum data and all of the standard patterns. Then, the standard pattern with the highest degree of similarity to the spectrum of the input audio signal is selected from among all the standard patterns, and data indicating a predetermined word corresponding to the standard pattern with the highest degree of similarity is converted into voice recognition data (
output as data indicating the words recognized from the input speech).

[Problems to be Solved by the Invention] However, in such conventional speech recognition devices, not only speech expressing necessary data but also other meaningless speech is input. As a result of calculation and selection, the standard pattern with the highest relative similarity is selected,
Words prepared in advance corresponding to the pattern are erroneously recognized as input data expressed by the voice.

Taking the above-mentioned cash card service as an example, when a depositor inputs his/her PIN number by voice, he/she says, "Um, 2, 6, 1.
, 5J. Then, in this speech recognition device, the above-mentioned comparison operation is also performed for the first voice "uh-to", and the number having the standard pattern with the highest relative similarity to the spectrum of this voice "uh-to" is used. (Let's assume it is "3") is selected. As a result, the depositor's voice was "3°2.
It is recognized as 6,1,5J, and the first four digits [3°2.6.
The IJ is determined to be the depositor's PIN number. In addition to meaningless words like "um",
This phenomenon can also occur when coughing or certain noises are detected.

As described above, conventional speech recognition devices have the problem of erroneously recognizing data due to input of meaningless speech other than speech expressing necessary data.

The present invention has been made in view of the above-mentioned points, and it is an object of the present invention to provide a speech recognition device capable of accurately recognizing data.

[Means and effects for solving the problems] The speech recognition device according to the present invention uses a standard pattern of the speech spectrum for predetermined data according to the applicable field of use, as well as a meaningless pattern that is easy for humans to utter. Standard patterns of sound spectra for involuntary sound absorption such as coughing, noise, etc. are held in the storage means. In the comparison calculation device, the acquired spectrum is successively compared not only with the standard pattern for the data but also with all the standard patterns for the meaningless paths and the like.

Similarity is required.

When the determining means selects one of the standard patterns for the predetermined data as having the greatest degree of similarity to the spectralized speech, the determining means outputs the data expressed by the standard pattern to a host computer, etc. do. However, when one of the standard patterns for the meaningless path is selected as having the highest degree of similarity, the meaningless path expressed by the standard pattern is ignored and not output.

As a result, when inputting predetermined data by voice, even if a nonsensical path whose standard pattern is stored in the storage means is uttered, the meaningless path is ignored and the correct data input by voice is uttered. only that is accurately output from the determining means.

〔Example〕

Hereinafter, one embodiment of the present invention will be described in detail with reference to the accompanying drawings.

In FIG. 1, a voice analysis section 1 is for sequentially analyzing the waveform of a detected voice to obtain spectrum data of the voice. For example, in the analysis section 1, 800
Waveform analysis is performed at a rate of ~1600 Hyde. Although the analysis section 1 normally converts the waveform into a digital signal before analyzing the waveform, the analysis section 1 may analyze the analog waveform as it is.

A data compression section 2 is provided to sequentially take in the speech spectrum data obtained by the speech analysis section 1. The data compression unit 2 performs data compression operations such as thinning or averaging of the spectrum data in units of predetermined time intervals. As mentioned above, 800~
The enormous amount of spectral data provided at the rate of 1600 bytes is thereby reduced to an amount suitable for sequential arithmetic processing. This thinning or averaging should be done within a range that does not impair the characteristics of the voice.
fl+ If is performed at time intervals that compress the spectrum data approximately 10 times.

The comparison calculation unit 3 is provided to sequentially take in the spectral data thinned out or averaged by the data compression unit 2 and perform a comparison calculation between the spectral data and a standard pattern stored in advance. It consists of a computer or a dedicated hardware device.

According to this invention, a memory 4 and a memory 5 are provided for storing standard patterns. The memory 4 stores standard patterns of the audio spectrum representing the required data depending on the field of use. For example, in the case of train reservations, standard patterns for the number 9 train name, distinction between ordinary cars and green cars, station names, etc. are stored in the memory 4. On the other hand, the memory 5 is for storing standard patterns of spectra for sounds other than the sounds representing the data that are likely to occur during data input. The memory 5 stores, for example, "um" and "um".

Standard spectral patterns are stored for nonsensical words that Iruma usually tends to pronounce, such as ``Ano'', as well as for coughing, noises specific to the place of use, and the like.

When the spectral data starts to be supplied from the data compression section 2, the comparison calculation section 3 sequentially reads out all these standard patterns from the memory 4 and the memory 5, and sequentially performs a comparison calculation between the spectral data and all the standard patterns. To go. As a result of this comparison calculation, the comparison calculation unit 3 sequentially outputs signals indicating the degree of similarity between the spectrum data and each standard pattern.

The output signal of the comparison calculation section 3 is sequentially input to the determination section 6. The determination unit 6 is comprised of a microcomputer or a dedicated hardware device, and is configured to select, based on the signal, the standard pattern that has the greatest similarity to the spectrum data of each input voice. If the standard pattern selected as a result is one stored in the memory 4 (that is, if it expresses meaningful data), the determination unit 6 determines the word (- word or (including words, phrases, etc.) is output to a host computer (not shown) as speech recognition data of the input speech. Furthermore, if the selected standard pattern is stored in the memory 5 (that is, if it corresponds to a meaningless word), the determination unit 6 ignores this and does not output speech recognition data. .

Incidentally, assuming that the comparison calculation section 3 and the determination section 6 are composed of a microcomputer, a program to be executed by the microcomputer is schematically shown in FIG. 2.

Steps 7 and 8 correspond to the functions of the comparison calculation unit 3, and in step 7, the spectrum data given from the data compression unit 2 and the true standard pattern read from the memory 4 are sequentially compared, and each Find the similarity of.

When step 7 is completed, the process proceeds to step 8.

In step 8, the spectral data is successively compared with the standard patterns for meaningless words etc. read out from the memory 5, and the degree of similarity of each is determined. When step 8 is completed, the process proceeds to step 9.

Steps 9 to 11 correspond to the functions of the determining section 6, and step 9 selects the standard pattern having the highest degree of similarity to the spectrum theta based on the comparison results of steps 7 and 8.

Subsequently, in step 10, it is determined whether the standard pattern selected in step 9 belongs to the standard patterns for idioms stored in the memory 5. If the pattern belongs to the standard pattern in the memory 5, YES is determined in step 10, and the process returns (therefore, the speech recognition data is not output to the host computer). On the other hand, if it does not belong to the standard pattern VC in the memory 5 (that is, if it belongs to the true standard pattern stored in the memory 4), it is determined NO in step 10, and the process proceeds to step 11. The speech recognition data corresponding to the standard pattern selected in step 9 is output to the host computer. When step 11 is completed, the process returns.

Next, an example of the operation of the speech recognition device according to the present invention will be described with reference to a specific example. When making a train reservation, when the person making the reservation inputs the train name by voice, [Hikari,
Let's say that when inputting the station name by voice, you pronounce it as ``Ano, Shin-Osaka''. At this time, the voice is spectralized by the voice analysis section 1 and the data compression section 2 and thinned out or averaged by 1, and then in the comparison calculation section 3, the numbers 2, the train name, and the numbers stored in the memory 4 and the memory 5,
Comparisons are made with all standard patterns for station names, meaningless words, etc., and a signal indicating the degree of similarity is sent to the calculation unit 3.
is given to the determination unit 6 from.

At this time, the determination unit 6 ``゛゛゛ is the station name whose standard pattern is stored in the memory 4 for the voice spectra of ``Hikari J, R217'' and ``Shin-Osaka.''

"Hikari", one of the data regarding train names, 12
No. 17'' and ``Shin-Osaka'' are selected as having the highest degree of similarity, respectively, and output to the outside. On the other hand, “Eh-”, “Um-”
Regarding the spectrum of the voice "E-" and "Ano-", which are one of the Japanese words for which the standard pattern is stored in Memory 5VC, we select them as having the greatest similarity, respectively. ignore. Therefore, from this voice recognition, the train name is "Hikari 217".

The correct data for the station name, ``Shin-Osaka,'' will be output.

In this way, when inputting the data whose mark pattern is stored in the memory 4 by voice, the voice recognition device ignores the uninformed American words, etc., and accurately inputs the data inputted by voice. It is recognized and output to the outside.

It should be noted that the memories 4 and 5 do not need to be separate memories in terms of hardware, but can store both the standard pattern corresponding to the true data and the standard pattern for non-American words, etc. in a single memory hardware circuit. There may be.

〔Effect of the invention〕

As described above, according to the speech recognition device according to the present invention, even when Japanese words such as Japanese are input during voice input of data,
It is possible to accurately recognize true input data without being affected by such unspoken words. Therefore, even if an unwritten American word is input during voice input of data, there is no need to input the voice of the data again, so the overall data input time can be shortened. As a result, in various service operations using voice recognition devices,
It is possible to improve the service.

[Brief explanation of the drawing]

FIG. 1 is a schematic block diagram showing an embodiment of the speech recognition device according to the present invention, and FIG. 2 shows the execution of the microcomputer when the microcomputer is used as the comparison calculation section and determination section in the embodiment. 1 is a flowchart schematically illustrating a program to be executed. 1...Speech analysis section, 2...Data compression section, 3...
Comparison calculation unit, 4, 5... memory, 6... determination unit.

Claims (1)

[Claims] Speech analysis means for sequentially analyzing an input speech signal to obtain spectrum data of the speech, and storing standard patterns of the speech spectrum for a plurality of specific words depending on the field of use. a first storage means that inputs the spectral data analyzed by the voice analysis means, compares the spectral data with all standard patterns stored in the storage means, and calculates the degree of similarity thereof. and determining means for selecting the standard pattern with the greatest degree of similarity based on the output of the comparison calculation means and outputting speech recognition data corresponding to the standard pattern, the speech recognition device comprising: further provided is a second storage means storing standard patterns of each voice spectrum for a predetermined plurality of meaningless words and other meaningless sounds that are easy to utter, and in the comparison calculation means, the voice analysis means The spectral data analyzed in the second storage means is further compared with the standard patterns of all meaningless words, etc. stored in the second storage means, and the determination means compares the standard patterns of the meaningless words, etc. with the degree of similarity. 2. A speech recognition device, wherein the speech recognition data is not output when the speech recognition data is selected as the largest one.