JPS6331793B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Technical Field> The present invention relates to a speech recognition method that performs speech recognition based on a matching result between a standard pattern registered in advance and an input speech pattern.

<Prior art> Speech recognition, especially speaker-specific speech recognition methods,
During registration, one word is uttered several times to have one or more standard patterns. That is, as shown in FIG. 1, M standard patterns are stored in the storage unit 1 for one word, and each of N words has M standard patterns.

As the number of words increases as shown in FIG. 1, the standard pattern group in memory 1 includes standard patterns similar to standard patterns in other groups. Therefore, when recognizing input speech, when matching is performed using a similar standard pattern, erroneous recognition often occurs, such as when the word is recognized as a word corresponding to this standard pattern. Conventionally, there was no means to prevent such misrecognition.

<Objective of the Invention> An object of the present invention is to prevent erroneous recognition, which is a drawback of the conventional method, and to suppress a decrease in the recognition rate.

<Embodiment> In the present invention, as shown in FIG. 1, a memory 1 stores M standard patterns for N words. In FIG. 1, P, i, j means the J-th standard pattern of the i-th word group. In contrast to the standard pattern memory 1 shown in FIG. 1, a flag memory 2 is prepared as shown in FIG. This flag memory 2 includes standard pattern memories P, i, j, memories F, i,
j is the corresponding one, and as will be explained later, if the standard pattern stored in the memory P, i, j is not suitable for speech recognition, set "1", and if appropriate, set "0". . In the present invention, by preparing the memories 1 and 2 shown in FIGS. 1 and 2, it is possible to reliably recognize the speaker's voice without erroneously recognizing it.

FIG. 3 is a flowchart showing the flow of the speech recognition method of the present invention. The present invention will be explained with reference to this figure.

First, in S1, the speaker selects M for one word.
Perform voice input to register standard patterns. When creating and registering standard patterns, well-known techniques can be used, such as dividing input audio through a number of bandpass filters, extracting audio characteristic data, and using this information. It is registered (recorded) as a standard pattern at a predetermined position in the memory 1, for example, P, i, and j. In this way, when the registration of M standard patterns for N words by the speaker is completed, the process moves to the next step S2. In this S1, the conventional technology is used as is.

Next, in S2, matching between the registered standard patterns of the memory 1 is performed. This extracts an arbitrary pattern from the standard pattern memory 1, uses it as the speaker's voice input, and executes voice recognition in the same way as usual. As a result of this matching,
If the extracted standard pattern belongs to the same group, "0" is set in the flag memory 2 corresponding to this standard pattern. To explain a little, the standard pattern P, 1, 1 of the first word of the first group is read out from memory 1 in FIG.
This pattern P, 1, 1 is used as the speaker's voice input and is matched with all other standard patterns.

In this case, matching with the standard pattern P, 1, 1 is excluded. As a result of this matching, if the standard pattern P,1,1 is not the first group but matches with other groups, this standard pattern P,1,1 is the standard pattern for recognition. is considered inappropriate, and P,1,1
F,1,1 of flag memory 2 corresponding to “1”
Make it. In other words, if a word from a matched group is input by voice, it should normally be recognized as a word from this group, but assuming that the matching with the standard pattern of P, 1, 1 is maximum, 1st
It will be mistakenly recognized as a word in the second group. In order to prevent misrecognition in this way, if the extracted standard patterns P, 1, 1 are not in the same group (first word) by matching,
Set flag memory F,1,1 to "1". Conversely, if they are in the same group, F,1,1 is set to "0".

The above processing is performed as P,1,2,P,1,3,P,
1, 4..., PN, M in order.

That is, any P,i, of the standard pattern
j _i=1 ... _N ; _j=1,2 ... _M is extracted and matched with all standard patterns other than P, i, and j. As a result, if it is not the i-th word group, the flag memory F, i, j is set to "1", and if it is the i-th word group, F, i, j is set to "0".
In this way, when the matching process between all the standard patterns from P, 1, 1 to P, N, M is completed, the position of the flag memory corresponding to the inappropriate standard pattern is "1" as shown in Figure 4. ” becomes. As shown in Fig. 4, since F,2,1, F,2,j is "1", the corresponding P,2,1, P,2,j of standard pattern memory 1 is standard. It can be seen that this is inappropriate as a pattern.

When the matching between the standard patterns is completed as described above, the process moves to S3, and the flag memory 2 is sequentially checked. If the flag is "1", it is determined that there is an inappropriate pattern among the standard patterns. Let the speaker know what you are doing. If "1" exists in the flags F, i, and j, the speaker decides whether or not to re-register the standard pattern at that location. If re-registration is to be performed here, the process moves to re-registration processing in S5. This re-registration instructs the speaker to re-register only the portion of the flag memory that is set to "1". That is, the speaker inputs the word corresponding to the "1" part of the flag memory 2 by voice, and the information from which the features are extracted is registered as an adjustment pattern in the same manner as the initial registration. Like this,
When the re-registration of all the portions of the standard pattern memory 11 corresponding to the portions of the flag memory 2 that are set to "1" is completed, the process moves to S2 and the matching process between the standard patterns is repeated. In this way, when the inappropriate standard pattern in the standard pattern memory is eliminated and all of F, i, and j in the flag memory 2 become "0", the process moves to the normal speech recognition process in S6. That is, when F, i, and j all become "0", there are no inappropriate standard patterns, and by performing recognition processing using this standard pattern memory 1, it is possible to suppress a decrease in the recognition rate.

Note that if the speaker does not instruct re-registration in S4, the process moves to speech recognition processing in S6. However, in speech recognition, the standard pattern in the corresponding memory 1, which is set to "1" in the flag memory 2, is not used for recognition processing. In other words, voice recognition is performed by performing a matching process between the standard pattern in the memory 1 corresponding to "0" in the flag memory and the input voice, thereby preventing erroneous recognition and improving the recognition rate.

<Effects of the Invention> According to the speech recognition method of the present invention, an arbitrary pattern is extracted from a group of standard patterns that have already been registered, the extracted pattern is matched with other standard patterns, and based on this matching result, By determining whether the pattern is inappropriate or not, and storing the fact that it is an inappropriate standard pattern as a result of this determination, the standard pattern stored as inappropriate can be excluded during speech recognition. Since speech recognition is performed based on other optimal standard patterns, the recognition rate can be significantly improved.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the storage state of the standard pattern memory according to the present invention, FIG. 2 is a diagram showing a flag memory provided corresponding to the standard pattern memory of FIG. 1, and FIG. 3 is a diagram showing the storage state of the standard pattern memory according to the present invention. FIG. 4 is a flowchart showing the procedure according to the recognition method, and is a diagram showing a flag memory based on the matching result between standard patterns according to the present invention. 1: Standard pattern memory, 2: Flag memory.

Claims (1)

[Claims] 1. Recognize the input voice by analyzing the input voice for voice recognition and matching the extracted data with a pre-registered standard pattern. In the speech recognition method, a step of registering M standard patterns for each word for N words, a step of selecting an arbitrary pattern from the group of registered standard patterns, and a step of selecting the selected standard pattern as one. a step of matching one standard pattern with all other standard patterns except for the standard pattern, using one as voice input data, and as a result of the matching, the standard pattern selected above matches a group of words that are the same as the standard pattern; If not, the selected standard pattern is stored in an unusable state, and the input voice is matched with other standard patterns excluding the stored standard pattern to perform voice recognition. Speech recognition method.