JPS616694A - Voice 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) The present invention relates to a voice registration method in a voice recognition device.

.

(Prior art) The general method of voice recognition m is to perform pattern matching between a standard pattern registered in advance and an input unknown voice pattern, and use the most similar standard pattern as the recognition result. be. In this case, the widely used method for registering the voice that should be registered in advance is to analyze and detect the voice by uttering - times to create a registration pattern, and then register the created registration pattern as the voice. .

In this case, if the utterance for boarding fails, the voice recognition will not go well. Furthermore, even if the same speaker utters the same word, it cannot be said that each utterance is exactly the same. Furthermore, if the registered utterance is completely different from usual, for example due to a cold, recognition becomes difficult, as is the case when the utterance fails.

However, it is difficult to conclude that the cause of this difficulty in speech recognition is a failure in utterance, and there are also cases in which a voice different from the voice that should be recognized is erroneously recognized in the recognition process. This may be caused by poor vocalization during recognition, but it can also occur if the quality of the registered standard pattern is poor.

In order to eliminate such registration failures due to vocalizations, it may be possible to create a standard pattern of recordings if you practice in advance and then perform voice registration. Furthermore, since practice during registration can be expected to have an effect on utterances during speech recognition, misrecognition is less likely to occur and the recognition rate is expected to improve. However, in order to easily perform the voice recognition process, it is not desirable in terms of device operation to require vocal practice at the time of voice registration.

(Object of the Invention) An object of the present invention is to solve the above-mentioned problems, eliminate the need for vocal practice to create a high-quality standard pattern during voice registration, and prevent errors caused by poor vocalization during voice registration. The object of the present invention is to provide a voice registration method that can eliminate recognition.

(Configuration of the Invention) The configuration of the present invention is such that in a voice registration method of a voice recognition device that performs a matching process between a registered voice standard pattern and an input voice pattern, 'C A plurality of first speech patterns obtained by uttering the speech to be recognized a plurality of times (As t AI-
..., An), and calculate the average matching distance between the same rings for these first voice patterns (Al to An).
A second voice pattern (Bt, Bt,
..., Bn) and the first audio pattern (At -A
A matching distance is calculated between different words with I*...t A n ) to obtain an average of the different words matching distance, and the first speech pattern (A8. bird) is calculated.

. . . , An), the one with the largest difference between the average matching distance between the same words and the average matching distance between different words is registered as the standard pattern of the first speech (5). shall be.

(Example) Next, the present invention will be explained in detail with reference to the drawings.

FIG. 1 is a block diagram of one embodiment of the present invention, and FIG. 2 is a block diagram of an embodiment of the present invention.
It is a figure which shows the matching distance of an example of the audio|voice pattern in a figure. In the figure, 1 is a speech analysis detection section, 2 is a speech registration section, 3 and 5 are storage sections, 4 is a matching processing section, 5 is an averaging processing section, 7 is a subtracter, and 8 is a comparison section.

Now, there are three types of voices (A, B, and C) to be registered, and voice A is uttered three times, and each of these utterance patterns is A1 * A2 = AI. In addition, sound B is played four times (B1*B2- Bs-B4) and sound C is played once (C1
) and select only one of these voices to register as a standard pattern. In FIG. 1, first, the speech analysis detection unit 1 analyzes the uttered speech pattern A1tA(e)'stBt*Bt
*Bs t B41 C is stored and held in the memory s3.

At this time, phonetic word information such as which pattern and which pattern are the same ring is also stored in the storage unit 3. Next, the matching processing unit 4 reads out the audio patterns stored in the storage unit 3, calculates matching distances for each, and stores the results in the storage unit 5.

The matching distance column in FIG. 2 was obtained in this way. Regarding this matching process,
For example, it is described on page 107 of "Speech Recognition" by Yasunaga Niimi, Information Science Course E19.3, published by Kyoritsu Shuppan Co., Ltd. If the vocalizations are exactly the same and the vocal patterns are exactly the same,
The matching distance becomes zero, indicating that the smaller the matching distance between two voice patterns, the more similar the two voices are, and the larger the matching distance, the more different they are.

Next, the voice A that should be recognized as the same edict = A (, A
For each pattern of -s, the average processing unit 6 calculates the average X of the matching distance between same words and the average Y of the matching distance between different words, and the subtractor 7 calculates the difference (Y-X) between them. Y for each pattern At thus obtained, bird-A
The comparison unit 8 compares the values of -X, and the registration unit 2 registers the pattern giving the largest value as a standard pattern. In this example, since AI gives the largest difference Y-X, this utterance A1 is registered as the standard pattern for the voice person. Next, for Bs-Bt-'%lB4, calculate the average matching distance between same words and the average matching distance between different words,
Pattern B4 that gives the largest difference by finding the difference between them
Register as a standard pattern. Since C is uttered only once, C1 is registered as is.

This matching distance between similar words becomes small, but if it becomes smaller than the matching distance with the speech pattern that should be recognized, it will result in incorrect recognition.

For things that should be recognized as different voices, it is advantageous to use patterns with a large matching distance as standard patterns. Also, among patterns that should be recognized as the same word, patterns that occur when pronunciation fails should be excluded,
The matching distance between the same rings for this pattern when vocalization fails becomes relatively large. In the example, the standard pattern is the pattern in which the value obtained by subtracting the average matching distance between different words from the average matching distance between same words is the largest, so that the pattern that is similar to the speech that should be recognized as another word, Moreover, this means that a pattern is selected that excludes patterns of failure in vocalization. The voice pattern selected in this way is suitable as a standard pattern and can reduce misrecognition.

(Effects of the Invention) As explained above, according to the present invention, it is possible to eliminate the complexity of re-registration due to failure of vocalization, eliminate the need for vocal practice for good registration, and avoid misrecognition of different voices. This has the effect of creating registered voice data that further improves the recognition rate.

[Brief explanation of drawings]

FIG. 1 is a schematic diagram showing an embodiment of the present invention, and FIG. 2 is a diagram showing the relationship between patterns and matching distances used to explain this embodiment. In the figure, 1... Voice analysis detection unit, 2... Voice registration unit, 3° 5...
・Storage unit, 4...Matching processing unit, 6...
... Average processing section, 7 ... Subtractor, 8 ...
. . . Comparison unit, 10 . . . Input terminal. Figure 2

Claims (1)

[Claims] In a voice registration method of a voice recognition device that performs a matching process between a registered voice standard pattern and an input voice pattern, a plurality of voices to be recognized as the same ring as a first voice to be registered are recognized. A predetermined matching distance is calculated between the plurality of first speech patterns obtained by uttering the same words twice, and the average matching distance between the same words is calculated for these speech patterns. Calculate the matching distance between the second speech pattern obtained by uttering the speech to be recognized as the first speech pattern and the first speech pattern to obtain the average of the matching distance between the first speech pattern and the first speech pattern. A voice registration method characterized in that a voice pattern having a large difference between the average matching distance between the same words and the average matching distance between different languages is registered as a standard pattern of the first voice.