JPS59201100A - Voice standard pattern registration system - 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 of the Invention The present invention relates to a method for registering standard speech sounds in sound recognition in which words or monosyllables are recognized.

(bl Technological Background With the development and spread of data processing technology in recent years, data processing j18
As part of the data input/output means in the system, voice recognition/synthesis technology, which was originally limited to voice-controlled sorting and guidance services on telephone lines, has also been supported by advances in semiconductors, especially integration technology and circuit design technology. As a result, implementation means that require high-speed arithmetic processing or large-capacity storage can be made smaller and lower in cost, and voice input/output means in Japanese are suitable for distributed processing and interactive formats and require special skill for operators. It has become popular as a data processing device that takes advantage of the conversion function between input voice and Tenkle Day 2, which is easy to operate and does not require any input.

(C1 Prior Art and Problems Conventionally, speech recognition devices usually set a plurality of words and/or monosyllables in the input speech to be recognized for a specific speaker, and pre-input each word or/and monosyllable. The characteristic parameters obtained by sampling each spectrum output obtained by applying it to a group of bandpass filters are stored as data, and this is used as a speech standard button corresponding to each word and/or monosyllable. The device has a function of converting an unknown voice into corresponding digital data each time it is input by comparing the voice bang of the input voice with the standard bang.Therefore, in order to recognize the voice pattern of the input voice, the voice recognition device For example, 4 times each time you register a phonetic standard button corresponding to a single syllable.
For example, 68 l or 101 monosyllables of P'jM each time ~8 times each require a total of 68/10 IX (4 to 8), which is a hassle not found in other data input devices. Repetition of utterances when registering voice/speech buttons not only burdens the speaker, but also increases the storage capacity of RAM and the amount of data processing in the device, so it is desirable to reduce the number of utterances from a cost standpoint. This has the disadvantage of reducing the reliability of the ability (': L m a = m).

tdl OBJECTS OF THE INVENTION The object of the present invention is to eliminate the above-mentioned drawbacks, in order to eliminate the above-mentioned drawbacks by repeating the voice 81 as many times as possible, for example once per word or/and single clause, and for safe subjects 2-3 times. A voice standard that achieves both a reduction in the number of utterances and a guarantee of self-sequence by registering a quasi-pattern to ensure a voice standard pattern that is comparable to the level achieved by the conventional multiple u, rotation) input. This is intended to provide a means of registration.

fel A'1 Achievement of the Invention The purpose of this invention is to recognize a large number of speech standard slam candidates in a speech recognition system that recognizes an unknown input speech by comparing input speech stamps with a plurality of speech standard stamps registered in advance in a dictionary. It has a voice standard slam candidate dictionary consisting of , and at the time of registration, the control unit compares the voice slam obtained by inputting the voice of the specific speaker to be registered into the voice processing unit with the phonetic standard one company slam candidates in the candidate dictionary. The similarity was determined and set for each voice to be registered! According to a threshold of Jii similarity and the number of registrations, candidates for voice stamps that are equal to or greater than the threshold and less than or equal to the number of registrations are selected and stored and registered in a storage unit to form a dictionary consisting of standard speech stamps for a specific speaker. This can be achieved by providing an audio standard bang registration method (J).

+f+ Embodiment of the Invention An embodiment 1 of the present invention will be described below with reference to the drawings. FIG. 1 is a professional diagram of the audio standard button registration method in an embodiment of the present invention, FIG. The figure is a flowchart showing the processing procedure in the audio standard button registration method according to an embodiment of the present invention. In the figure, 1 is a control unit, 2 is a storage unit, 21 is a control program, 22 is control data, 23 is a speech standard slam candidate dictionary, and 23a.

b...p is a voice standard slam candidate group, 23aa,
ab, ac,...8g... are voice standard bang candidates, 24 is sound p5 registered standard bang dictionary, 24a, b
...p is the audio standard button gl, 24aa, ab...
・ah... is a standard voice button. The control unit 1 controls each component according to the control program 21 and control data 22 stored in the storage area of the storage unit 2, selects the standard audio button in response to the audio input signal, and selects the standard audio button corresponding to the specific speaker. A dictionary 24 is created.The storage unit 2 stores in its storage area the control program 21 and the similarity level of each standard bang candidate with the voice bang to be compared, or the standard bang corresponding to each word and/or monosyllable. The recognition (hereinafter referred to as the number of registrations) value of the standard button to be selected from each candidate group 23a, b...p and registered for each standard button group 24a, b...p is set.

It should be noted that multiple speakers (a+b+C
...G)'s audio bang is the standard bang group +nj 23 aa -ag in the standard bang candidate group 23a.
] ξ (standard slam 24aa for "a" of an unknown speaker.

It is assumed that ab...ab is accommodated in the conflict Bakun group 24a. However, the standard batan group sleeve 23aa is the standard batan 24
AA does not correspond to direct judgment. Therefore, 23 a -1) of the standard batan candidate group, 24 of the material 2 rod Bakun Ei I
The numbers a to p are in blue and each single ml or/and single row 11
It corresponds to the total four prongs of 1j. Example (+: (68-6:
Taba 101 is a talented person.
The standard slam candidates a-g that are common to ~p correspond to the pre-accumulated multi-V and speaker g's words, and are standard slam candidates -24a-
The standard batons a - h common to p are Kim Ue ((corresponding to
N [j is 6. Although not shown in this embodiment of the present invention, input No. 16 in the form of an analog electrical signal obtained by inputting a particular speaker's utterance "A" into a microphone is input to the audio processing m3. and extracts its characteristic parameters to create a voice button.

That is, the spectral characteristics obtained from the input signal by means of sampling the audio frequency of 200 to 5400 Hz into m channels, e.g., 16 bandpass filters, and the temporal changes into n, e.g., 16 or 32, are expressed as 256 or 512 pieces of data. Comparison unit 4-digit comparison unit 1 to which the audio bang Xa is applied, which is converted into an audio slam Xa and sent out, selects the standard slam candidates 23aa-ag corresponding to the circle marks in the standard baton candidate group 23a shown in FIG. RrlJ is compared with the control data 22 in advance, and the standard slam candidates with the highest degree of similarity within the threshold range set in the control data 22, i.e., those with the closest distance to the data, are selected, for example, by the number of meters of the control data 22, for example, Select the standard button group 24a to select the standard button aa~a.
In hko\, six standard buttons aa-af are registered. Similarly, the other voice button Xp is 0 in the standard button candidate group 23p.
The data of the standard slam candidates 23pa to 23pg corresponding to the marks are compared and the standard batons pa to pf are registered in the standard slam group 24p. The deformed ellipse showing the standard button candidate group 23a in FIG. 2 is not an external line surrounding the area, but a display line connecting the mark A% / standard] (The circle indicating the tongue group 24a is also indicated by an X mark)
It can also be erased by a single voice click from the fi person.
A standard batan dictionary 24 is created by selecting it from the d弗 batan candidate dictionary 23, which is composed of the data of the voice batans of many speakers who have KA.
We have replaced the trouble of having to pronounce each word multiple times to register a, b...p with a standard slam dictionary that can be used by a specific speaker with just one utterance for each word or monosyllable, p times in total. It is useful because 24 can be recorded.

In addition, in my own explanation, the sound data accompanying the vocalization, for example, the X mark in Figure 2 of Xa, that is, the sound (for Kun Xa, the standard of the standard button 24a) is not adopted for aa - a f, but the sound The analogy for Batan Xa self-fang is bv<
The structure of the standard notch group 24a as a quasi-button ag can be expected to provide higher density data.

When the voice sound due to the input signal on the lever (as shown in the ■ mark in Figure 12, which is significantly different from the conventional voice button candidate 23aa-ag) is obtained with a deviation from the distribution as 1 degree, If the control unit 1 determines this as an erroneous input signal or a malfunction of the audio processor 11 section 3, suppresses the intermediary data processing, and controls so as not to set the standard buttons aa to ah, an incorrect standard button dictionary 24 will be registered. to be done (1
I'm not talented. At this time, it is necessary to omit the illustrations, and the system should display a warning to that effect, for example, 1! For example, the speaker is asked to utter 17''' for the same sound.Also, in this embodiment, the voice bangs of many speakers are used as standard bang candidates, but the average number of voice bangs from a large number of voice bangs is It goes without saying that the same effect can be achieved by using a synthesized button or both of the methods described above1, To register a standard slam dictionary for p nest words and/or monosyllables, we now have to deal with the hassle of having to utter each word multiple times and process the data once or fewer times. This is useful because it can greatly reduce the troublesomeness of the speaker's utterances and the amount of processing time required.

[Brief explanation of the drawing]

Figure 1 is a professional diagram of the voice standard registration method in one embodiment of the present invention, and Figure 2 is a diagram of the voice standard registration method (Japanese).

Claims (3)

[Claims] (1) In a voice recognition system that recognizes an unknown input voice by comparing the number of voice standard bangs registered in advance during a speech with the input sound 1-bang, a voice standard consisting of a large number of voice standard bang candidates is used. It has a slam candidate dictionary, and the control unit at the time of registration compares the voice button obtained by inputting the voice of a specific product to the voice processing unit and the voice el button candidates in the candidate dictionary and calculates the degree of similarity between them. , according to the similarity threshold 2 set for each voice to be registered and the number of registrations, select voice single-bang candidates that are above the threshold and below the number of registrations, and store and register them in the storage unit. The dictionary is made up of quasi-bakuns, and the speech standard baptan registration method is fully characterized. (2) The control unit stores and registers the selected audio standard button candidate and the audio button obtained from the input voice at the time of registration in the storage unit to form a dictionary consisting of the audio standard button of the specific product name. A voice standard button registration method according to claim 1, which is characterized by: (3) When the control unit detects that the voice beat obtained by inputting it to the voice processing unit at the time of registration deviates from the distribution according to the voice standard slam candidates in the voice standard beat candidate dictionary, the control unit performs the voice standard based on the voice beat. 2. The audio standard button registration method according to claim 1, wherein storage registration of a button is suppressed.