JPH11282485A - Voice input device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice input device for solving the problem that a shake by the accidental factor of the place is present at the time of voice input and a recognition rate is lowered because of that. SOLUTION: This device is provided with a voice analysis part 101 for calculating the feature amount of each fixed time of the plural times of input voices, a dictionary part 102 for holding the voice patterns of all words assumed as input, a matching part 103 for performing the pattern matching of the feature amount and word models in the dictionary part and calculating and outputting coarse likelihood to the respective word models of the input voice, a coarse likelihood storage part 104 for holding the coarse likelihood with all the input voices of the respective word models and an output means 105 for outputting the words for a predetermined fixed piece number from the one of the large weighted average of the coarse likelihood with the input voices of the corresponding word models. By utilizing the plural pieces of the input, adverse influence of the accidental factor is eliminated. Also, by performing recognition from the utterance of different forms, influence from an utterance form is reduced as well.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice input device, and more particularly to a voice registration / recognition device that performs word recognition by a plurality of voice inputs.

[0002]

2. Description of the Related Art As a speech registration / recognition apparatus for performing word recognition by a plurality of speech inputs, for example, Reference 1 (Kiyoshi Kano, "Basics of Speech Recognition", May 17, 1996, pp. 32-34, 39- Four
3) etc. are referred to.

[0003] Word input by voice input device is still 100
In other words, the input device requires a means for improving the reliability of speech recognition.

[0004]

By the way, it is conceivable that a human voice, even if it is the same person's utterance of the same word, greatly varies depending on the environment and the state of the speaker each time it is uttered. In particular, when trying to recognize only one utterance content, the disturbance of the single utterance or the like greatly affects the speech waveform, and erroneous recognition is likely to occur.

[0005] For this reason, in various aspects of using the voice input device, there is a phase in which the number of voice inputs can be set to be large, assuming high recognition performance.

Accordingly, it is an object of the present invention to provide a speech input device that realizes speech recognition with a high recognition rate using a plurality of input speeches in such a situation.

[0007]

In order to achieve the above-mentioned object, the present invention integrates information of two or more utterances, thereby making it possible to produce a unique "sway" which is likely to occur only once. The effect is reduced, and highly accurate voice input is realized. In addition, the present invention eliminates the “fluctuation” of the voice waveform generated depending on each utterance style by adopting inputs having different utterance styles.

More specifically, the present invention provides a voice analysis means for calculating a feature amount of a plurality of times of input voices at predetermined time intervals, and a voice pattern of all words assumed as input (referred to as a "word model"). ), Pattern matching is performed between the feature amount and the word model in the dictionary unit, and the likelihood (“coarse likelihood”) of the input speech with respect to each of the word models is calculated and output. Matching means, a coarse likelihood storage unit for holding the coarse likelihood between all of the input voices of the respective word models, and the word between the input voices of the corresponding word models. Output means for outputting a predetermined fixed number from the largest weighted average of the coarse likelihoods.

[0009]

Embodiments of the present invention will be described. Normally, in speech recognition, the likelihood between an input speech and an existing word model is calculated by pattern matching, and a word model having the highest likelihood among a set of word models assumed in advance is used as a recognition result. However, it is often the case that the likelihood of the word model corresponding to the word actually uttered is calculated to be low due to the disorder of the utterance.

In the present invention, the influence of the likelihood decrease of individual utterances due to such accidental factors is reduced by using the average value or the maximum value of the likelihoods of a plurality of utterances as the likelihood.

It is also conceivable that the likelihood decreases due to the style of utterance. This is an acoustic change such as “namake” that appears when a specific syllable set is connected. In the present invention, in order to reduce the influence of the likelihood decrease caused by such a utterance style, from the likelihood of utterance by a plurality of utterance styles, the average value or the maximum value as described above is obtained, and the likelihood is determined. Is used.

The speech input device according to the present invention, in a preferred embodiment thereof, includes a speech analysis means (101 in FIG. 1) which receives a plurality of speeches as input, calculates acoustic features for each frame of each input speech, and outputs them. A dictionary unit (102 in FIG. 1) for holding a word model, and inputting a feature amount of each input voice output from the voice analysis unit, obtaining a likelihood of the input voice for each word model, and outputting as a coarse likelihood. Matching means (103 in FIG. 1), a coarse likelihood storage unit (104 in FIG. 1) holding the coarse likelihood of any combination of the input speech and the word model, Output means (105 in FIG. 11) for calculating an average value of the coarse likelihood for each input voice, ordering the word models in descending order of the value, and outputting a fixed number of word models according to the order; It is provided.

The operation of the embodiment of the present invention will be described.

[0014] The voice analysis means performs a frequency analysis of the voice of the section for each fixed time (frame) of the input voice,
Calculate and output feature values. As the feature amount, the power of the voice, the power change amount, the cepstrum, the cepstrum change amount, and the like are used.

The dictionary section holds word models of all words to be recognized, that is, words assumed in advance.

[0016] The matching means comprises:
Pattern matching of all word models in the dictionary is performed, and a coarse likelihood for each word model of the input speech is calculated and output.

The coarse likelihood storage unit stores the likelihood of each word model output by the matching unit for each input speech.

The output means sums the coarse likelihoods stored in the coarse likelihood storage unit for each word model to obtain an average value. After that, the words represented by the word model are output by a certain number in order from the word model having the larger average value. The matching means and the output means may be realized by program control executed by a processor or the like constituting the voice input device.

[0019]

Next, in order to explain the above-mentioned embodiment of the present invention in more detail, an embodiment of the present invention will be described below with reference to the drawings.

A first embodiment of the present invention will be described.
FIG. 1 is a diagram showing the configuration of the first exemplary embodiment of the present invention.
Referring to FIG. 1, a first embodiment of the present invention is a speech analyzing unit 101 which calculates and outputs an acoustic feature amount for each frame of each input speech, and a dictionary unit 102 which holds a word model.
A matching unit 103 that receives a feature amount of each input voice output from the voice analysis unit as input, obtains a likelihood of each of the input voices with respect to each word model, and outputs the likelihood as a coarse likelihood; A coarse likelihood storage unit 104 that holds the coarse likelihood of any combination, and for each word model, calculates the average value of the coarse likelihood for each input speech, and orders the word models in descending order of their values; Output unit 105 that outputs a certain number of word models according to the order
And is provided.

FIG. 2 is a flowchart showing the operation of the first embodiment of the present invention. The operation of the first embodiment of the present invention will be described with reference to FIGS.

It is assumed that the dictionary unit 102 holds three word models, "Sato", "Kato", and "Goto". The number of words output by the output unit 105 is two. Further, it is assumed that the utterance “Sato” is made three times as the input voice.

First, in step 1, the voice analysis unit 1
01 calculates and outputs the feature amount of each frame from the first utterance of “Sato”.

Next, in step 2, the matching unit 1
Reference numeral 03 denotes a pattern matching of the feature amount of each frame for each word model stored in the dictionary unit 102, as described in the above-mentioned document 1, and Viterbi (Viterbi).
The coarse likelihood of each word model for the first "sato" input speech is obtained by using the method.

Next, in step 3, the matching unit 1
03 stores the coarse likelihood of each word model for the first “sato” input speech in the coarse likelihood storage unit 104.

Thereafter, steps 1 to 3 are repeated for the second and third input speeches of "Sato", as shown in FIG.
4 to be stored.

After the coarse likelihood is stored for all combinations of the input speech and the word model, in step 4, the output unit 105 outputs the coarse likelihood of all the utterances of each word model as shown in FIG. Find the average value.

Thereafter, in step 5, the output unit 1
05 is two words “Sato” having a large average value of the coarse likelihood,
Outputs "Kato".

According to one embodiment of the present invention, a weak point that stable likelihood cannot be obtained by pattern matching using only one utterance is compensated for.

Next, another embodiment of the present invention will be described. In order to prevent the influence of the likelihood decrease caused by the accidental factor as described above, only the utterance having the highest likelihood may be selected and recognized from a plurality of utterances. The second embodiment of the present invention is different from the first embodiment shown in FIG.
5 is configured by replacing the average value with the maximum value. Even in this case, it is possible to prevent the influence of the decrease in the likelihood due to an accidental factor.

Next, a third embodiment of the present invention will be described. The decrease in likelihood can be caused not only by accidental factors but also by differences in utterance style. In order to reduce the influence of the utterance style, a method using a plurality of utterance styles can be considered. The third embodiment of the present invention is a modification of the operation of the matching unit 103 in the above-described embodiment shown in FIG.

In the third embodiment of the present invention, the matching unit 105 changes the matching process according to the format of the input voice. When the input voice is a word utterance such as “Sato”, the operation described in the first embodiment is performed. In the case of a syllable utterance such as “Sa,...,. The syllable recognition is performed for each input syllable, and the likelihood of each syllable is averaged to obtain a coarse likelihood.

The fourth embodiment of the present invention is obtained by adding a syllable model for all syllables included in each word to the contents of the dictionary unit 102 in the above-described embodiment shown in FIG. For example, the word model "Sato"
Is added to the syllable models “sa”, “to” and “u”.

The operation of the fourth embodiment of the present invention will be described. Regarding the operation of the fourth embodiment of the present invention, the first embodiment
The description will be given in accordance with the flowchart shown in FIG.

The dictionary unit 102 includes a word model “Sato”,
In addition to "Kato" and "Goto", the syllable models "Sa", "To", "U", "K",
Hold five “go”.

The number of words output from the output unit 105 is set to 2, as in the first embodiment. In addition, as the input voice, the word “Sato” is once spoken, and “Sa,
…,…, U ”are uttered once.

The operations from step 1 to step 3 for word utterance are the same as in the first embodiment.

Steps 1 and 3 are the same in syllable utterance, but step 2 is different. In step 2 for syllable utterance, the matching unit 103
Is the matching between the “s”, “ka”, and “go” syllable models corresponding to the first syllable of the word model stored in the dictionary unit 102 and the first syllable of the input. And the likelihood of the first syllable of the input to each syllable model is obtained.

For the second syllable, matching between the syllable model of “to”, which is the second syllable of each word, and the input second syllable is performed in the same manner.

Matching is similarly performed for the third syllable. After that, the matching unit 103 calculates and outputs the average value of the likelihood of each input syllable with the corresponding syllable model of each word as the coarse likelihood of each word.

Thereafter, step 4 is performed in the same manner, and the two words are output in descending order of the average value of the coarse likelihood.

Using this embodiment, as described in the example, information from two utterance modes, word utterance and syllable utterance, can be obtained. This can lead to errors due to the utterance style, such as "naming" between syllables in a word,
It is possible to reduce the influence of utterance disorder at the beginning of a syllable, and the like.

In addition, erroneous recognition that cannot be prevented by using these means may be corrected manually. The fifth embodiment of the present invention provides manual correction means to the above embodiment.

Referring to FIG. 5, in a fifth embodiment of the present invention, a speech analysis unit 101 for calculating and outputting an acoustic feature value for each frame of each input speech, a dictionary unit 102 for holding a word model, and A matching unit 103 that receives a feature amount of each input speech output from the speech analysis unit, obtains a likelihood of the input speech for each word model, and outputs the likelihood as a coarse likelihood.
And a coarse likelihood storage unit 104 that holds the coarse likelihood of any combination of the input speech and the word model;
For each word model, an average value of the coarse likelihood for each input speech is calculated, the word models are ordered in descending order of the value, and an output unit 105 that outputs a certain number of word models according to the order; An input unit 106 for inputting a selection from a number of options, and a final selection unit 107 for obtaining and outputting a word model as a final output from the output from the output unit 105 and the output from the input unit 106 Have been.

FIG. 6 is a flowchart showing the operation of the fifth embodiment of the present invention. Next, the operation of the fifth embodiment of the present invention will be described with reference to FIGS.

The operations of the dictionary unit 102, the voice analysis unit 101, the matching unit 103, the coarse likelihood storage unit 104, and the output unit 105 are the same as those in the above-described embodiment. However, the output of the output unit 105 is also performed to the final selection unit 107. It is assumed that the input unit 106 is an input device that can input numbers from 1 to 3. This may be something like a keyboard or switch device, or another voice input device. Also, the person who operates the input means 106 does not necessarily have to be the same person who uttered the input voice. The input unit 106 outputs the input number to the final selection unit 107.

The final selection unit 107 holds the outputs from the output unit 105 and their order, and outputs a word model in which the numbers output by the input unit 106 are in order. Also,
It is assumed that the input given to the input means 106 is the numeral 1.

First, the operations from step 1 to step 5 are performed in the same manner as in the above embodiment.

At step 6, the input means 106 outputs the given input number 1 to the final selection unit 107.

In step 7, the final selection unit 107
Outputs only the first output “Sato” corresponding to the number 1 as the final output.

In this embodiment, the reliability of the input is increased by the final selection. Although the final selection itself is also used for ordinary single-input speech recognition, in this embodiment, since the probability that a correct answer is included in the word models up to the certain number is extremely high, the final selection is not determined. Probability is high, and the probability of giving the user pain such as selecting from a plurality of candidates that do not include a correct answer is remarkably lower than that in the case of using a single input.

[0052]

As described above, according to the present invention, the following effects can be obtained.

The first effect of the present invention is to solve the problem that a stable likelihood cannot be obtained by pattern matching using only one utterance, and to prevent the influence of a decrease in likelihood due to an accidental factor. ,That's what it means.

The reason is that, in the present invention, by using the average value or the maximum value of the likelihoods of a plurality of utterances as the likelihood, the influence of the likelihood reduction of the individual utterances due to such accidental factors is considered. This is due to the reduction.

Further, the second effect of the present invention is to reduce the effects of errors caused by the utterance style, for example, the effects of "slowness" between syllables in a word and disorder of utterance at the beginning of a syllable. That's what it means.

The reason is that, in the present invention, in order to reduce the influence of the likelihood decrease caused by such a vocal style,
This is because the average value or the maximum value as described above is obtained from the likelihoods of utterances in a plurality of utterance styles, and is set as the likelihood.

As described above, according to the present invention, it is possible to realize a voice input device that is robust against errors caused by accidental factors or utterance styles.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of one embodiment of the present invention.

FIG. 3 is a table showing a list of coarse likelihoods between each word model and each input voice in one embodiment of the present invention.

FIG. 4 is a diagram showing, in a table format, a list of coarse likelihoods and averages of the coarse likelihoods between each word model and each input voice in one embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of another embodiment of the present invention.

FIG. 6 is a flowchart showing the operation of another embodiment of the present invention.

[Description of Signs] 101 voice analysis unit 102 dictionary unit 103 matching unit 104 likelihood storage unit 105 output unit 106 input unit 107 final selection unit

Claims (10)

[Claims] 1. A speech analysis means for calculating a feature amount of a plurality of input speeches at predetermined time intervals, and a dictionary unit for holding speech patterns of all words assumed as inputs (referred to as "word models"). A matching unit that performs pattern matching between the feature amount and a word model in the dictionary unit, calculates and outputs a likelihood (referred to as “coarse likelihood”) of the input voice with respect to each of the word models, A coarse likelihood storage unit for holding the coarse likelihood between the word model and all of the input voices; and a weighted average of the coarse likelihood between the word and the corresponding input voice of the word model. Output means for outputting a predetermined fixed number from the largest one. 2. A speech analysis means for calculating a feature amount of a plurality of times of input speech at regular time intervals, and a dictionary unit for holding speech patterns of all words assumed as input (referred to as "word model"). A matching unit that performs pattern matching between the feature amount and the word model in the dictionary unit, calculates a likelihood of the input speech for each of the word models (referred to as “coarse likelihood”), and outputs the likelihood. A coarse likelihood storage unit for holding the coarse likelihood between all the input voices of each word model, and a maximum of the coarse likelihood between the word and the input voice of the corresponding word model. Output means for ordering values and outputting a predetermined number in accordance with the order. 3. The dictionary unit holds, for all words assumed as the input, speech patterns (referred to as “syllable models”) of all syllables appearing in the words. Only when the input speech is a word utterance, pattern matching is performed between the feature amount and the word model in the dictionary unit, and the likelihood of the input speech for each of the word models is output as a coarse likelihood. Only when the input voice is a syllable utterance, pattern matching is performed between the feature amount and the syllable model in the dictionary unit, and the likelihood of each syllable in the input voice with respect to the syllable model is output as a coarse likelihood. Be characterized in that
The voice input device according to claim 1. 4. The likelihood output by the matching means for an input speech that is the syllable utterance is an average value of the likelihood of each syllable in the input speech with respect to a corresponding syllable model. The voice recognition device according to claim 3, which performs the operation. 5. A speech recognition apparatus according to claim 1, wherein said output means has a weight used for calculating said weighted average of one. 6. The output means according to claim 1, wherein the weight used for calculating the weighted average is a geometric progression having a predetermined constant value as a common ratio in chronological order of the input voice. Voice input device. 7. The voice input device according to claim 1, wherein the number of words output by said output means is one. 8. An input means for allowing a user to select and input a correct answer from the predetermined fixed number of word models, and an optimal word based on an input from the input means and an output from the output unit. The voice input device according to any one of claims 1 to 6, further comprising: final selection means for selecting and outputting one model. 9. (a) A plurality of input speeches each having a characteristic amount calculated at predetermined time intervals by a voice analysis means, and the calculated characteristic amount and voice patterns ("") of all words assumed as inputs. Pattern matching with a word model in a dictionary unit that holds the word model, and calculates a likelihood (referred to as “coarse likelihood”) of the input speech with respect to each of the word models. Matching means for outputting to the coarse likelihood storage unit which holds the coarse likelihood between all of the input voices; (b) calculating the word of the coarse likelihood between the input voice of the corresponding word model; From the one with the largest weighted average,
A recording medium for recording a program for causing a computer constituting the voice input device to function each of the above-mentioned means (a) and (b) of an output means for outputting a predetermined fixed number. (A) A feature amount of a plurality of times of input speech is calculated for each fixed time by speech analysis means, and the calculated feature amount and speech patterns ("") of all words assumed as input are calculated. Pattern matching with the word model in the dictionary unit holding the word model is performed, and the likelihood of the input speech for each of the word models (referred to as “coarse likelihood”) is calculated. Matching means for outputting to the coarse likelihood storage unit which holds the coarse likelihood between all of the input voices, and (b) the coarse likelihood between the word and the input voice of the corresponding word model. Order the maximum degree,
A recording medium for recording a program for causing a computer constituting the voice input device to function each of the means (a) and (b) of the output means for outputting a predetermined number in a predetermined order according to the order.