JPS5823097A - Voice recognition apparatus - 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] This invention is a voice iia! that introduces words, syllables, etc. input by voice! It is related to the identification device.

Speech recognition devices that recognize words or syllables such as numbers, place names, etc. input by voice are in practical use in the fields of sorting luggage and inputting data from terminals. The configuration of a conventional speech recognition device is shown in FIG.

In Fig. 1, prayers that are input more than input terminal 1 are band-limited in the voice analysis section 2, and then converted into digital voices by A/D conversion, and voice sections are extracted using power information, etc. , a feature parameter is calculated in the extracted speech section, and the input speech is converted into a percentage feature parameter. The MU section 3 compares the standard patterns stored in the standard pattern storage section 4 with the input speech converted into feature parameters by the analysis section 2, and calculates the distance (or similarity or similarity) between the input speech and each standard pattern. The corresponding t) is calculated. The distances to all standard patterns are calculated, and the word corresponding to the standard pattern with the minimum distance (in the case of similarity, the strict similarity) is output from the output terminal 5 as a recognition result. Although the following discussion will be based on distance, it can also be applied to similarity by replacing the minimum value with the maximum value.

Conventionally, in this type of speech recognition apparatus, it has been usual to store only standard patterns of words to be recognized in the standard pattern storage section 4. However, the user of the speech recognition device does not necessarily correctly input only the single chain of agt & target. Sometimes I have a cough, dark circles, etc.
It is unavoidable that voices such as "and" and "um-" are input.Also, if you hear voices, shoes, doors opening and closing, etc., these may also be picked up.Furthermore, When inputting audio through a telephone line, busy tones, ringback tones, etc. may be input depending on the situation.When these voices or sounds are input,
If these audio trigger sounds are determined to be close to any of the internally placed standard patterns and are output from the output terminal 5, an undesirable result will be output. Therefore, for example, when trying to enter data into a computer by voice, or when trying to receive services such as banking or seat reservations by voice, this can cause serious problems.

In order to avoid such a situation, a conventional method sometimes allows an output to be rejected. This is to prevent the above-mentioned recognition error by setting a threshold (mD) and determining that the distance value d associated with the result of one fiber is larger than D is rejected. However, Even with this method, if a voice similar to the voice to be recognized or a Fi prayer is input, recognition errors will inevitably occur.

FIG. 2 is a diagram showing this situation, and schematically shows the voice characteristic parameters as points in space. 4
@Area 6 is the distribution of single chains to be recognized as "yes", and point 7 is the point of its standard pattern.

In order not to reject correct utterances, the threshold value needs to be large enough to cover region 6. At this time, points inside the sphere 8 with a radius around the point 7 are recognized as "yes". Region 9 shows the distribution of M1 non-targeted speech, which is "Tani 1", which overlaps with Muzuka 8, and point 10 is that 1.
The position of the two utterance data is shown in the overlapping multi-portion.

In this case, point 10 will be erroneously recognized as 1, ``Yes''.This situation is an unavoidable problem in conventional voice 111g Makoto devices, regardless of the recognition method.

In order to eliminate this drawback, the present invention creates and stores in advance a standard pattern corresponding to the voice or sound to be rejected, and outputs a rejection when the recognition result becomes the standard pattern for rejection. It is a voice recognition device that is characterized by
The purpose of the present invention is to correctly reject voices or sounds that are easily recognized incorrectly and cause serious trouble in conventional devices.

FIG. 3 shows an embodiment of the present invention, and parts corresponding to those in FIG. In addition to the storage section 41 for storing the data, a storage section 42 for expansion rate patterns corresponding to voices or sounds to be rejected is provided.

The operation of this device is almost the same as that of the speech recognition device shown in FIG. In this inventive device, the identification section 3 compares the input voice with the standard single-chain pattern of the wt recognition object stored in the storage section 41, and also rejects the input voice stored in the storage section N42. It also performs matching with standard patterns of appropriate speech or sounds, and finds the standard pattern with the minimum distance among them. Further, if the obtained standard pattern is a standard pattern stored in the storage section 41, the same processing as in the apparatus shown in FIG. 1 is performed and the result is output. If the standard pattern stored in the memory unit 42 as the [4 pattern with the minimum distance is selected, the result is output to the output terminal 5 as a rejection.

FIG. 4 illustrates the principle of effective operation of the device of the present invention, and parts corresponding to those in FIG. 2 are given the same reference numerals. Point 11 on the space of the standard pattern prepared for “Voices that should be rejected” and point lO of the utterance data
Since this point 10 is the closest to the standard pattern 11, a normal output of rejection is obtained for the voice at this point 10.

As explained above, #cTijt of the present invention prepares a standard pattern of speech or sound to be rejected, and when it is determined that the input speech is closest to the standard pattern of speech or sound to be rejected, In order to produce an output of rejection, voices or sounds that are unavoidable to be input to the voice introduction device such as coughing, coughing, busy tones, and ringback tones are correctly rejected and an undesirable durable weave is obtained. It has the advantage of avoiding

4. Simplified drawings, E94 Figure 1 is a block diagram showing the configuration of a conventional audio cable device.
Fig. 2 shows an example in which a word to be rejected is incorrectly recognized in a conventional speech recognition device, and Fig. 3 is a block diagram showing the configuration of an embodiment of the inventive device. FIG. 4 is a diagram showing that when this invention is used, a case that should be rejected is correctly rejected.

l two-person power terminal, 2: analysis section, 3: bean paste pupil section, 4: standard pattern storage section, 41: standard pattern storage section for certified acupuncture target candy,
42: Input O4J quasi-pan LnH self-memory part to be rejected, 5: Output terminal, 6: Distribution of candy “yes” for Mk target, 7: 4s quasi-pattern of “yes”, 8: “yes” and V , range of input to be displayed, 9: distribution of ``tani'' to be rejected, 10: position of specific utterance of ``tani'', 11: standard pattern of ``dani''.

% Permit 1 person Nippon Telegraph's NU Service Corporation Agent Kusano Kuruma Figure 71 E 2 Diagram Figure 3 42

Claims (1)

[Claims] (1) a speech analysis section that analyzes input speech and extracts characteristic parameters; a standard pattern storage section that stores standard patterns; a feature parameter extracted from the input speech by the speech analysis section; It is composed of a voice ktk section that performs matching with the yajunno (turn) stored in the pattern memory section, and in the standard pattern memory section, the voice to be recognized or the voice to be recognized is A speech recognition device whose main purpose is to memorize standard patterns.