JPH10171488A - Method for speech recognition and device therefor and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce reject error and to improve operability of information input by determining a recognition result of an input speech according to weighted likelihood. SOLUTION: Firstry, when a model of each word to be recognized is compared with that of an unknown word in the likelihood, a weight for the model of the unknown word is determined with respect to that of the word (S020). Namely, the weights determined here are as many as the subject words for recognition. Each weight value for the subject words for recognition is stored in a storage device by making them corresponding to words. Thus, by determining weights for each word for recognition and storing them prior to reject processing, it is possible to determine a suitably weighted likelihood (S410) for each word for recognition by reading the weight for calculating a likelihood of an unknown word. After weight determination and storing process have been executed, recognition processing (S100-S601) are repeated every time a speech is inputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method and apparatus, and a storage medium.

[0002]

2. Description of the Related Art In speech recognition, acoustic analysis of an input speech is performed, and the likelihood indicating the similarity between the analysis result and a model of a word to be recognized prepared in advance is calculated. Is output as the recognition result.
Therefore, even when an utterance that is not a recognition target is input as a recognition target voice or the user coughs, the likelihood of those input voices is calculated and the recognition target word having the highest likelihood is recognized. Will be output as As a technique to deal with these, when an unexpected utterance (unknown word) such as a cough or an unrecognized word is input, a model of the unknown word is created by rejecting the recognition result of the utterance. A method of rejecting a recognition result when a value obtained by weighting the likelihood of the word model is larger than the likelihood of the recognition result is generally known. A conventional method of performing a recognition process from a speech input and outputting a recognition result will be described with reference to FIG. First, when a voice is input, the voice is captured (S100), the acoustic analysis of the voice captured in S100 (S200), and the likelihood of the model of the recognition target word prepared in advance with respect to the analysis result in S200. Is calculated (S300). Next, the likelihood of the unknown word model with respect to the analysis result of S200 is calculated (when the weight for the unknown word is set in advance, considering the weight) (S410). (S30
The order of the steps 0 and S410 may be interchanged or may be performed simultaneously.) The likelihood of the obtained recognition result is compared with the likelihood of the unknown word model.If the likelihood of the unknown word model is high, the recognition result is rejected (S601) .If the likelihood of the recognition result is high, the recognition result is high. Is output (S600).

[0003]

When a computer or the like is operated by voice, erroneous recognition may adversely affect subsequent processing. For example, when the word "delete" corresponds to the operation of deleting data on the computer's storage device, when an unknown word (for example, "prize" or "evidence") is entered, it is incorrectly recognized as "delete" In this case, the data will be erased regardless of the intention of the operator,
The subsequent processing will be adversely affected. The likelihood of such a situation occurring must be reduced. Also, conversely,
If the erroneous recognition does not have a significant effect, it is better that the number of times correct speech is rejected by the rejection process is small.

However, if the accuracy of the unknown word model is not good, for example, utterances that should be originally recognized as unknown words are rejected (false rejection), utterances to be rejected are not rejected (rejection omission), and rejection processing is performed. Error occurs. When the weight applied to the likelihood of the unknown word model calculated in S400 is set large, rejection is reduced, but erroneous rejection is increased. Conversely, when the weight is reduced, false rejection is reduced but rejection is increased.

Conventionally, a certain weight is applied to the likelihood of an unknown word model regardless of the recognition target word and the recognition result, and it is determined whether to reject only the likelihood.

Therefore, when the weight applied to the unknown word model is large, there is a problem of rejection processing error such that a word to be rejected cannot be rejected and is erroneously recognized as a word which should not be erroneously recognized. In addition, there is a problem of rejection processing error that words that do not need to be rejected are rejected when the weight is set small.

[0007]

SUMMARY OF THE INVENTION In order to solve the above-mentioned conventional problems, the present invention recognizes an input speech by using a predetermined recognition target word and determines the likelihood of an unknown word model for the input speech. Deriving a likelihood obtained by multiplying the derived likelihood by a weight stored in association with the recognized recognition target word, and obtaining the input speech according to the derived weighted likelihood. Provided are a speech recognition method and apparatus for determining a recognition result of a speech recognition device, and a storage medium.

[0008] In order to solve the above-mentioned conventional problems, the present invention preferably calculates the likelihood using an HMM model.

[0009] In order to solve the above-mentioned conventional problems, it is preferable that the recognition of an input voice using the recognition target word calculates a likelihood of the input voice with the recognition target word, and the likelihood is calculated. It is assumed that the largest recognition target word is determined as a recognition result.

In order to solve the above-mentioned conventional problems, the present invention preferably compares the likelihood of a word to be recognized obtained in the recognition with the likelihood of the weighted unknown word model,
According to the comparison, when the likelihood of the recognition target word is large,
The recognition target word is output as a recognition result of the input voice.

In order to solve the above-mentioned conventional problems, the present invention preferably compares the likelihood of the word to be recognized obtained in the recognition with the likelihood of the weighted unknown word model,
As a result of the comparison, if the likelihood of the weighted unknown word model is large, the recognition target word is rejected.

[0012] In order to solve the above-mentioned conventional problems, the present invention preferably provides the recognition target word with an attribute indicating whether it is a rejection target word and stores the word, according to the stored attribute. , Is controlled so as to switch whether or not to perform processing for weighting the unknown word model.

[0013] In order to solve the above-mentioned conventional problems, the present invention preferably provides an attribute indicating whether or not rejection processing is required for each item for inputting the input voice, and stores the attribute. Control is performed so as to switch whether or not to perform processing for weighting the unknown word model according to the attribute.

In order to solve the above-mentioned conventional problems, the present invention preferably provides that the weight information corresponding to the recognition target word is:
It is stored as a table associated with the recognition target word.

In order to solve the above-mentioned conventional problems, the present invention preferably provides that the weight information corresponding to the recognition target word is:
The recognition target word is included in the speech recognition grammar and stored.

[0016] In order to solve the above-mentioned conventional problems, the present invention preferably provides weight information corresponding to the recognition target word,
This is performed prior to the likelihood calculation processing of the unknown word model.

In order to solve the above-mentioned conventional problem, the present invention preferably calculates the weight information corresponding to the recognition target word every time the recognition target word is updated.

In order to solve the above-mentioned conventional problems, the present invention preferably displays the recognition result on a display means.

In order to solve the above-mentioned conventional problems, the present invention preferably inputs the voice from a microphone.

In order to solve the above-mentioned conventional problems, the present invention determines a weight to be applied to the likelihood of an unknown word model for each recognition target word, calculates the likelihood of the input speech with the recognition target word, Calculate the likelihood of the input speech with the unknown word model, and determine the weight determined for the recognition target word for which the likelihood has been calculated,
Multiplying the likelihood of the unknown word model obtained by the above calculation,
Find the weighted likelihood of the unknown word model, compare the likelihood of the word to be recognized with the weighted likelihood of the unknown word model, and decide whether to reject the recognition result for the input speech according to the comparison result. The present invention provides a speech recognition method and apparatus for determining the number and a storage medium.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram illustrating an example of a schematic configuration of a speech recognition device according to the present invention. Figure
FIG. 3 is a flowchart of a voice recognition process according to the present invention.

Processing such as voice recognition for a user's utterance described below is performed by the CPU (200) according to a program stored in the storage device (400). The program of the flowchart described later is also stored in the storage device 400. Further, the program stored in the storage device 400 may be downloaded from a removable external storage medium such as FD or CDROM, or from another device via a public line or a LAN.

First, in step S020, for each word to be recognized, when comparing the likelihood between the model of the word and the unknown word model, the weight applied to the model of the word by the unknown word model is determined. That is, the weight determined here is the same as the number of words to be recognized. Alternatively, the weight may be determined for each group that is grouped according to an attribute (for example, the degree of risk of the word) that is predetermined for each recognition target word.

The value of the weight for each recognition target word determined in S020 is stored in the storage device 400 as a table in association with the word.
May be stored in the grammar information for speech recognition used in the likelihood calculation processing in S300. Thus, prior to the rejection processing, the weight of each recognition target word is determined and stored in advance, so that the weight is read out in the likelihood calculation of the unknown word model in S410, and an appropriate weight is set for each recognition target word. The likelihood multiplied is obtained.

The weight determination process in S020 may be performed prior to the rejection process, and may be performed before S100 as shown in the flowchart of FIG. 3, or may be performed after S300, for example.

After the weight is determined and stored in S020, S100 to S601 are repeated each time a voice is input. For example, when the number of words to be recognized increases or when the user instructs, S020 may be executed.

When a voice is input from a voice input / output device (500) such as a microphone or a telephone, the voice is captured in step S100 and A / D conversion is performed.

Next, the A / D-converted voice is stored in step S200.
Acoustic analysis with.

Next, in step S300, the CPU (200) calculates the likelihood representing the similarity between the result of the acoustic analysis and the model in which the recognition target word stored in the storage device (400) is represented by an HMM or the like. The model with the highest likelihood is set as the recognition result.

Next, in step S410, the CPU (200) calculates the likelihood representing the similarity between the acoustic analysis result and a model in which the unknown word stored in the storage device (400) is represented by an HMM or the like. At this time, by reading the weight corresponding to the word of the recognition result in step S300 from the table or the grammar stored in the storage device 400, the weight determined to be appropriate for each recognition target word determined in step S020 is determined. Multiply the likelihood of the unknown word model.

Next, in step S500, the likelihood of the recognition result calculated in step S300 is compared with the likelihood of the unknown word model calculated in step S410.

If the likelihood of the recognition result is equal to or greater than the likelihood of the unknown word model in step S500,
Proceeding to S600, the recognition result is displayed on the display device (300).

If the likelihood of the recognition result is smaller than the likelihood of the unknown word model in step S500, the flow advances to step S601 to discard the recognition result.

In step S601, the rejection may be displayed on the display device (300). In step S600, the recognition result may be stored in the storage device (400).

An example in which the processing in steps S020 and S300 in the above description is changed will be described with reference to the flowchart of FIG. The other steps are almost the same as those in the flowchart of FIG. 3, and thus the same step numbers are given and the description is omitted here. Hereinafter, the changed steps will be described.

When determining the weight to be applied to the likelihood of the unknown word model for each word to be recognized in step S020, the words to be recognized are classified into words to be rejected and words not to be rejected. The weight to be applied to the likelihood of the unknown word processing model for words not to be set is set to 0 (S040). This S040 is S020
Is performed at the same timing as the storage device 400 is performed, and the storage method in the storage device 400 is the same.

In step S300 in the first embodiment,
After the recognition result is output, the storage device 400 is checked whether the word of the recognition result output in S300 is to be rejected (whether the weight applied to the likelihood of the unknown word model is 0) (S350).

If it is determined in S350 that the word is not to be rejected (when the weight is 0), the likelihood calculation of the unknown word model (S410) and the likelihood comparison between the recognition result and the unknown word model are performed. Instead, the process proceeds to step S600, and the word output in S300 is displayed as a recognition result.

If it is determined in S350 that the word is a word to be rejected (when the weight is other than 0), the flow advances to step S410 to calculate the likelihood of the unknown word model. Do the same.

Next, it is determined before the likelihood calculation whether or not the content of the utterance is to be rejected by using the flowchart of FIG.
Processing for switching whether to perform rejection processing based on the result will be described.

Step S0 in the flowchart of FIG.
After step 20, a step S050 is provided for determining whether or not the utterance content is to be rejected.
Speech recognition using rejection processing (S
100 to S601). If it is determined in step S050 that the utterance is not the object of rejection, the rejection processing of steps S400 and S500 is not performed, only S100 to S300 are performed, and the recognition result is displayed in the same manner as normal speech recognition.

For example, in a system for inputting data of a plurality of items by voice, whether or not rejection processing is required for each item is stored in the storage device 400 in advance, and the data is input as a voice input target. It can be used to refer to the storage device 400 to determine whether an item is an item that requires rejection processing or an item that is not necessary by referring to the storage device 400.

[0043]

As described above, according to the present invention,
The input speech is recognized using a predetermined recognition target word, a likelihood of the unknown word model for the input speech is derived, and the derived likelihood is stored in association with the recognized recognition target word. Deriving a weighted likelihood, and determining the recognition result of the input speech according to the derived weighted likelihood, speech recognition with less rejection errors can be realized, and information by speech recognition can be realized. The operability in the case of performing the input of can be improved.

As described above, according to the present invention, the calculation of the likelihood is performed using an HMM model, thereby enabling speech recognition with few rejection errors in accurate speech recognition.

As described above, according to the present invention, input speech recognition using the recognition target word calculates the likelihood of the input speech with the recognition target word and calculates the recognition target having the maximum likelihood. By performing the process of determining a word as a recognition result, it is possible to perform a rejection process in consideration of a weight suitable for a recognition target word having the highest likelihood.

As described above, according to the present invention, the likelihood of the word to be recognized obtained in the above recognition is compared with the likelihood of the weighted unknown word model.
If the likelihood of the recognition target word is large, the recognition target word is output as the recognition result of the input voice, and it is determined that the recognition target word is appropriate after performing rejection processing in consideration of a weight suitable for each of the recognition target words. The determined recognition target word can be obtained as a recognition result.

As described above, according to the present invention, the likelihood of the word to be recognized obtained in the recognition is compared with the likelihood of the weighted unknown word model.
If the likelihood of the weighted unknown word model is high, rejection of the recognition target word allows the rejection to be determined at an appropriate level for each of the recognition target words.

As described above, according to the present invention, the recognition target word is added with an attribute indicating whether it is a rejection target word and stored, and the unknown word is stored in accordance with the stored attribute. By performing control to switch whether or not to apply weighting to the model, rejection processing is performed as necessary, so that the processing efficiency of speech recognition can be improved.

As described above, according to the present invention, an attribute indicating whether or not rejection processing is required is added and stored for each item to which the input voice is input, and according to the stored attribute, By controlling whether or not to perform the process of weighting the unknown word model, the rejection process is performed as needed, so that the processing efficiency of speech recognition can be improved.

As described above, according to the present invention, the weight information corresponding to the recognition target word is separately provided for each recognition target word by storing the weight information as a table corresponding to the recognition target word. Weight information can be obtained quickly and appropriately.

As described above, according to the present invention, the weight information corresponding to the recognition target word is separately provided for each recognition target word by storing it in the speech recognition grammar of the recognition target word. Weight information can be obtained quickly and reliably.

As described above, according to the present invention, the weight information corresponding to the recognition target word is performed prior to the likelihood calculation processing of the unknown word model, so that the weight information different for each unknown word target word is obtained. Can be determined at an appropriate timing.

As described above, according to the present invention, the calculation of the weight information corresponding to the recognition target word is performed every time the recognition target word is updated, so that even when the recognition target word changes. In addition, different weight information can be reliably prepared for each recognition target word.

As described above, according to the present invention, by displaying the recognition result on the display means, the recognition result can be confirmed accurately.

As described above, according to the present invention, voice input can be easily performed by inputting the voice from the microphone.

As described above, according to the present invention, the weight applied to the likelihood of the unknown word model is determined for each recognition target word, the likelihood of the input speech with the recognition target word is calculated, and the input speech is calculated. The likelihood of the unknown word model is calculated by multiplying the weight determined for the recognition target word for which the likelihood has been calculated by the likelihood of the calculated unknown word model. Finding weighted likelihood, comparing the likelihood of the recognition target word with the weighted likelihood of the unknown word model, and determining whether to reject the recognition result for the input speech according to the comparison result. Accordingly, rejection errors can be reduced, and operability when inputting information by voice can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of a configuration of a speech recognition device according to the present invention.

FIG. 2 is a flowchart of a conventional voice recognition method.

FIG. 3 is a flowchart of a voice recognition method according to an embodiment of the present invention.

FIG. 4 is a flowchart of a speech recognition method 2 according to an embodiment of the present invention.

FIG. 5 is a flowchart of a voice recognition method 3 according to an embodiment of the present invention.

Claims (18)

[Claims] An input speech is recognized using a predetermined recognition target word, a likelihood of an unknown word model for the input speech is derived, and the derived likelihood is used as the recognized recognition target word. A speech recognition method, comprising: deriving a weighted likelihood stored in association with the input speech; and determining a recognition result of the input speech according to the derived weighted likelihood. 2. The speech recognition method according to claim 1, wherein the calculation of the likelihood is performed using an HMM model. 3. Recognition of an input speech using the recognition target word is a process of calculating a likelihood of the input speech with the recognition target word, and determining a recognition target word having the maximum likelihood as a recognition result. 2. The speech recognition method according to claim 1, wherein: 4. Comparing the likelihood of the word to be recognized obtained in the recognition with the likelihood of the weighted unknown word model, and as a result of the comparison, if the likelihood of the word to be recognized is large,
2. The speech recognition method according to claim 1, wherein the recognition target word is output as a result of the recognition of the input speech. 5. The likelihood of the word to be recognized obtained in the recognition and the likelihood of the weighted unknown word model are compared, and the likelihood of the weighted unknown word model is large by the comparison. 2. The speech recognition method according to claim 1, wherein in the case, the recognition target word is rejected. 6. A method for assigning an attribute indicating whether the word is a rejection word to the recognition target word and storing the word, and determining whether to perform a process of weighting the unknown word model according to the stored attribute. 2. The voice recognition method according to claim 1, wherein control is performed to switch between them. 7. A process for assigning an attribute indicating whether rejection processing is required for each item for inputting the input voice and storing the attribute, and performing a process of weighting the unknown word model according to the stored attribute. 2. The speech recognition method according to claim 1, wherein control is performed to switch whether or not the voice recognition is performed. 8. The weight information corresponding to the recognition target word is:
2. The speech recognition method according to claim 1, wherein the speech recognition method is stored as a table associated with a recognition target word. 9. The weight information corresponding to the recognition target word is:
2. The speech recognition method according to claim 1, wherein the speech is included in a speech recognition grammar of the recognition target word and stored. 10. The speech recognition method according to claim 1, wherein weight information corresponding to the recognition target word is performed prior to a likelihood calculation process of the unknown word model. 11. The speech recognition method according to claim 1, wherein calculation of weight information corresponding to the recognition target word is performed each time the recognition target word is updated. 12. The speech recognition method according to claim 1, wherein the recognition result is displayed on a display. 13. The speech recognition method according to claim 1, wherein the speech is input from a microphone. 14. A weight for the likelihood of the unknown word model is determined for each recognition target word, a likelihood of the input speech with the recognition target word is calculated, and a likelihood of the input speech with the unknown word model is calculated. Multiplying the weight determined for the recognition target word for which the likelihood has been calculated by the likelihood of the calculated unknown word model to obtain a weighted likelihood of the unknown word model; A speech recognition method comprising comparing a likelihood of a word with a weighted likelihood of an unknown word model, and determining whether to reject a recognition result for the input speech according to the comparison result. 15. A recognition target word recognition unit for recognizing an input speech using a predetermined recognition target word, an unknown word model likelihood derivation unit for deriving a likelihood of an unknown word model for the input speech, Unknown word model weight accounting likelihood deriving means for deriving a likelihood obtained by applying a weight stored in association with the recognized recognition target word to the derived likelihood, and the derived weighted A recognition result determining unit that determines a recognition result of the input voice according to the likelihood. 16. A weight determining means for determining a weight to be applied to the likelihood of the unknown word model for each recognition target word, a recognition target word likelihood determining means for calculating the likelihood of the input speech with the recognition target word, Unknown word model likelihood determining means for calculating the likelihood of the input speech with the unknown word model; and an unknown word obtained by calculating the weight determined for the recognition target word for which the likelihood has been calculated. Unknown word model weighting likelihood deriving means for obtaining the weighting likelihood of the unknown word model over the likelihood of the model; and likelihood comparison comparing the likelihood of the recognition target word and the weighting likelihood of the unknown word model. Means, and rejection determination means for determining whether or not to reject the recognition result for the input voice according to the comparison result. 17. A computer-readable storage medium, comprising: a step of recognizing an input speech using a predetermined recognition target word; and a step of deriving a likelihood of an unknown word model for the input speech. Deriving a likelihood obtained by weighting the derived likelihood stored in association with the recognized word to be recognized, and extracting the likelihood of the input voice according to the derived weighted likelihood. A storage medium storing a program for a step of determining a recognition result. 18. A computer-readable storage medium, comprising: determining a weight to be applied to the likelihood of an unknown word model for each recognition target word; and calculating a likelihood of the input speech with the recognition target word. Calculating the likelihood of the input speech with the unknown word model; and determining the weight determined for the recognition target word for which the likelihood has been calculated, the likelihood of the unknown word model obtained by the calculation. Determining the weighted likelihood of the unknown word model, comparing the likelihood of the recognition target word and the weighted likelihood of the unknown word model, and recognizing the input speech according to the comparison result. A storage medium storing a program for a step of determining whether to reject the result.