JPH075890A - Voice interactive device - Google Patents

Abstract

PURPOSE:To provide a voice interactive device which reduces user's average number of uttering. CONSTITUTION:This device which decides a correct candidate work by the voice input by a user who responds to a candidate word by a candidate word presenting part 8 is provided with a word-oriented likelihood-correct rate corresponding data storage part 6 which stores the data indicating the relationship between the likelihood and correctness rate for every candidate word and a correctness rate threshold value storage part 7 which stores the threshold value of the correctness rate that is used to judge whether to confirm the correctness of candidate words or to instruct of the pronunciation of a word again to obtain the correctness rate of an objective candidate word from the likelihood of the candidate words obtained from a likelihood computing part 2 and the data stored in the part 6. Also, the device is provided with a confirmation control part 5 which compares the correctness rate and the threshold values stored in the section 7, controls the part 8 when the correctness rate is larger than the threshold value, and controls a repronunciation instruction part 9 if the correctness rate is smaller than the threshold value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice interactive apparatus, and more particularly to a voice interactive apparatus for rationally and accurately confirming the correctness of a recognition result of a word voice uttered by a user of the voice interactive apparatus.

[0002]

2. Description of the Related Art A conventional example will be described with reference to FIG. In the voice interaction device, the likelihood between the input word voice and all the words stored and registered in the memory in advance is calculated, and as shown in FIG. In many cases, the candidate words are presented to the user for confirmation in the order of “Ito”, which has the highest frequency, from “Sato” to “Kato”. If the likelihood of the word to be confirmed is extremely small, it is determined that the probability of being the correct answer is low, and the confirmation process is not performed, as shown in FIG. The method of instructing recurrent voice is generally adopted.

Conventionally, in carrying out such confirmation dialogue processing, as to a word to be confirmed, as shown in FIG. 3, a certain threshold value is set for the likelihood so as to determine the word candidate rank. It is determined whether the confirmation processing should be selected or the re-voice instruction should be selected depending on whether or not the likelihood of is larger than this threshold value. Setting a certain threshold value for the likelihood means that when the confirmation process or the re-voice instruction is selected,
This means that a fixed threshold is always used regardless of the uttered word and the candidate word. However, while some words have a high recognition likelihood and a large likelihood difference from other words, it is relatively difficult to make accurate utterances. There are some words that have a low denominator and have little likelihood difference with other words.

Here, for a word with a statistically high likelihood, the likelihood that the likelihood is small means that the likelihood is wrong, so that the re-voice instruction is selected rather than the confirmation process. It is considered more efficient to set the threshold value relatively high in order to make the value large. On the other hand, for a word with a statistically low likelihood, it is relatively difficult to make a correct utterance, so although the likelihood is small, it is more likely that the utterance is wrong. Since it is more likely that this is a characteristic peculiar to the word that the likelihood is inherently small, it is better to select a confirmation process rather than set a relatively low threshold and select a re-voice instruction. Considered to be efficient.

[0005]

However, in the above-mentioned judgment processing, no consideration is given to the likelihood peculiar to a word, and the judgment processing is carried out based on a fixed threshold value. Therefore, for words with a statistically high likelihood, it is more efficient to set a high threshold value, although it is considered to be more efficient. The confirmation process will continue regardless of the target. On the other hand, for words with a statistically low likelihood, a low threshold is considered to be more efficient, but the result is set to a high value, and when the confirmation process should be continued Will be instructed.

The present invention provides a voice interactive apparatus which solves the above problems.

[0007]

A registered word parameter storage unit that includes a voice analysis unit 1 that analyzes an input voice signal and calculates a characteristic parameter thereof, and stores the characteristic parameters of a plurality of preregistered words. 3, the input voice feature parameters obtained from the voice analysis unit 1 and the registered word feature parameters stored in the registered word parameter storage unit 3 or each word is assumed to be a hidden Markov model, and the feature parameters are statistically processed. A likelihood calculation unit 2 that calculates the likelihood for each registered word of the input speech from the state transition probability and the symbol output probability obtained by the above is provided, and the candidate word number and the likelihood of the candidate word are arranged in descending order of likelihood. A candidate word output unit 4 for outputting and a candidate word presentation unit 8 for presenting the candidate words output from the candidate word output unit 4 to the user in the order of likelihood are provided. In the voice dialogue device, which is provided with a re-voiced voice instruction unit 9 for instructing the user to perform again, a correct dialogue candidate word is determined by the user's voice input in response to the candidate word presented by the candidate word presentation unit 8. A word-by-word likelihood-correctness rate correspondence data storage unit 6 that stores data indicating a correspondence relationship between the likelihood and the correctness rate for each word is provided, and the correctness or incorrectness of the candidate word is confirmed or the word is uttered. The correctness rate threshold value storage unit 7 stores a threshold value of the correctness rate for determining whether to instruct again, and the likelihood of the candidate word obtained from the likelihood calculation unit 2 and the word-specific likelihood-correctness rate correspondence data storage unit. The correct answer rate of the target candidate word is obtained from the data of No. 6, the correct answer rate is compared with the threshold value stored in the correct answer rate threshold storage unit 7, and if the correct answer rate is larger than the threshold value, the candidate word presentation / presentation unit 8 Drive and the accuracy rate is smaller than the threshold Re-utterance instruction unit 9
A voice interaction device including a confirmation control unit 5 for driving the device is constructed.

[0008]

Embodiments of the present invention will be described. This invention
Instead of the likelihood that has been conventionally adopted in a voice dialog device, the correct answer rate of the candidate word to be processed is recalculated from the likelihood, and the confirmation processing or the selection judgment of the re-voice instruction is made based on the calculated correct answer rate. Is to do. As will be described in detail later, this correct answer rate is obtained by substituting the likelihood into a calculation formula representing the characteristics of the likelihood distribution obtained in advance for each candidate word.

The correct answer rate thus obtained takes into consideration not only the likelihood of the corresponding candidate word alone, but also the identification rate with other words. As a result, even when the likelihood distribution is greatly different depending on the candidate word, the confirmation and the re-emergence voice can be determined based on the common correct answer rate. By executing the confirmation and re-evaluation judgment processing based on the correct answer rate, the above-mentioned conventional speech dialogue apparatus that executes judgment processing based on a certain threshold value without giving any consideration to word-specific likelihood. The drawbacks of can be eliminated.

Here, the accuracy rate will be described. First,
When a candidate word O _k having a word number k (k = 1, 2, ... N, where N is the number of recognition target words) is observed at the i-th place with the likelihood L, the probability P _ikL that the observation result is the correct answer. Is expressed using the frequency distribution of the likelihood of the recognition result. In addition, the uttered true word is represented by w _j (j = 1, 2, ... N, where j is a word number),
The occurrence frequency (prior probability) of the spoken word is P (w _j ).

It is assumed that the occurrence likelihood distribution for each generated word, that is, the probability distribution D _{j, ikL} in which the candidate word O _k appears at the i-th place with the likelihood L when the word w _j is uttered is obtained in advance. To do. D _{j, ikL} = P (i, O _{k, L} | w _j ), where i, j, k = 1, 2, ... N (1) D _{j, ikL} should be obtained at the evaluation stage before system operation. You can On the other hand, the data that can be obtained by the system during the actual recognition processing is a list of candidate word strings O _k received from the recognition processing unit and the likelihood L corresponding to each k.

Therefore, the probability P _{ikL, j} that the true word is wj can be calculated from the information available to the system, that is, the list of the word numbers k and the likelihoods L as shown in FIG. If possible, P _{ikL, j = k gives the} probability that the observation result is correct (correct answer rate) P _ikL , and determines whether to control the system, that is, whether confirmation processing or re-voice processing is to be selected. be able to. P _{ikL, j} has a _meaning of P _{ikL, j} = P (w _j | i, O _k , L) (2), but can be expressed as follows from the Bethe's theorem.

[0013]

[Equation 1]

If the denominator of the above equation is replaced with the sum of products of the condition and the conditional probability and the equation (1) is substituted,

[0015]

[Equation 2]

_Therefore , the correct answer rate P _ikL is

[0017]

[Equation 3]

The probability distribution D represented in advance and obtained in advance
The correct answer rate of the candidate can be _obtained from _{j, ikL} . The method of the present invention is _{premised on} that the distribution D _{j, ikL} is obtained in advance at the system design evaluation stage. Although it is possible to collect distributions of correct answer candidates in a relatively short period of time, it is necessary to perform a large number of evaluation experiments in order to obtain likelihood distributions of incorrect answer candidates, especially candidates with a low rank. Not necessarily realistic. Therefore, at the initial stage of system operation, the initial distribution is obtained by the method described below, and when sufficient data has been accumulated since the start of operation,
It is considered that the method of exchanging the distribution is realistic and has few problems. As a condition for determining the initial distribution, the condition "the correct answer rate of the recognition candidate is determined only by the likelihood and does not depend on the rank." That is, the _assumption P _{ikL, j} = P _{kL, j} = P (w _j | O _k , L) (6) is introduced for setting the initial distribution.

As a result, the distribution of appearances by observed word is simplified as follows. D _{j, kL} = P (O _k , L | w _j ) j = 1,2, ... N (7) In this case, the likelihood of all the candidate words can be obtained with one utterance, and therefore several tens per word can be obtained. Sufficient vocalization can acquire sufficient data as initial data. When the correct answer rate P _kL is obtained again based on the above assumption, the result is as _follows .

[0020]

[Equation 4]

The numerator of this correct answer rate P _kL is D _{k, kL} P (w
_k), the word k is its own when the word w _k is utterance means the probability of occurrence in the likelihood L. The denominator is the sum of all the probabilities that the word k appears with the likelihood L when the word w _l including the word w _k itself is uttered. Therefore, when a word w _l other than the word w _k itself is uttered, the word k
That is, the probability of appearing at the likelihood L is taken into consideration. By carrying out the confirmation and the judgment of the re-voiced voice based on the common correct answer rate P _kL thus considered, the judgment corresponding to the likelihood peculiar to the word is finally carried out, and the likelihood peculiar to the word is peculiar. It is possible to eliminate the drawbacks of the conventional voice dialog device in which the judgment processing is performed based on a certain threshold value without giving any consideration to the degree.

Embodiments of the present invention are shown in FIGS. 4, 5 and 6.
It will be specifically described with reference to. First, a prompt voice output is made from the voice interaction device, for example, "Please have ++". The user responds to this by, for example, “OOO
Please say. " The voice signal corresponding to this utterance is subjected to voice analog-digital conversion processing and feature parameter extraction processing in the voice analysis unit 1. As the characteristic parameter, for example, a cepstrum sequence of a registered word, a delta cepstrum sequence, a power sequence, or a vector quantization sequence obtained by vector quantization (VQ) of these parameters is generally adopted. In the voice interaction device of the present invention, any of these characteristic parameters can be applied. The extracted feature parameter is sent to the likelihood calculation unit 2. Likelihood calculation section 2 includes speech analysis section 1
A characteristic parameter of the input voice extracted in
From the state transition probabilities and the symbol output probabilities obtained by statistically processing the characteristic parameters of a plurality of registered words or each word stored in the registered word parameter storage unit 3 or assuming each word as a hidden Markov model (HMM). , Calculate the likelihood between the input speech and each registered word. The likelihoods of the calculation results are sent to the candidate word output unit 4, where the candidate word numbers and the likelihoods are rearranged in descending order of likelihood as shown in FIG.

First, with respect to the candidate word having the first highest likelihood, the correct answer rate of the target candidate word is obtained from the likelihood and the data in the word-wise likelihood-correct answer rate correspondence data storage unit 6. Next, the obtained correct answer rate is compared with the threshold value stored in the correct answer rate threshold storage unit 7, and if the correct answer rate is higher, the word number of the candidate word is passed to the candidate word presenting unit 8. The candidate word presenting section 8 asks whether the corresponding candidate word is correct or not.
Or ask the user to answer "No". Example:
"Is it XX?"

"Yes" as a result of recognizing the response voice of the user
If so, the candidate word with the first highest likelihood is the correct word, and if “no”, the same process is repeated for the candidate with the second highest likelihood. If the correct answer rate is smaller, the re-voice instruction unit 9 is controlled to instruct the user to utter another word "Please try again".

[0025]

As described above, the dialogue device of the present invention can perform confirmation and judgment of recurrent voice based on the common correct answer rate even when the likelihood distributions differ greatly depending on the candidate words. Therefore, it is possible to proceed with the dialogue without performing unnecessary re-voice and unnecessary confirmation processing, and as a result, it is possible to reduce the average number of times the user speaks.

In the interactive apparatus of the present invention, the calculation result of the correct answer rate for each word is tabulated in advance and stored in the word-likelihood-correct answer rate correspondence data storage unit, and this is referred to during the actual interactive processing. The processing speed is improved by executing the processing in this way.

[Brief description of drawings]

FIG. 1 is a diagram illustrating an example of a dialogue for confirming the correctness of a candidate word.

FIG. 2 is a diagram illustrating an example of a dialogue instructing re-voice.

FIG. 3 is a diagram illustrating a conventional example of confirmation dialogue processing using a certain threshold of likelihood.

FIG. 4 is a diagram showing an example of information passed from a recognition processing unit.

FIG. 5 is a diagram showing an embodiment of the present invention.

FIG. 6 is a diagram illustrating an example of confirmation dialogue processing according to the present invention.

[Explanation of symbols]

 1 Speech analysis unit 2 Likelihood calculation unit 3 Registered word parameter storage unit 4 Candidate word output unit 5 Confirmation control unit 6 Word-specific likelihood-correctness rate correspondence data storage unit 7 Correctness rate threshold storage unit 8 Candidate word presentation unit 9 Recurrent voice Indicator

Claims (1)

[Claims] 1. A voice analysis unit for analyzing an input voice signal and calculating a characteristic parameter thereof, and a registered word parameter storage unit for storing characteristic parameters of a plurality of words registered in advance, Input speech feature parameters obtained from the analysis unit and registered word parameter feature parameters of registered words stored in the storage unit or state transition probabilities and symbols obtained by statistically processing the feature parameters assuming each word as a hidden Markov model A likelihood calculation unit that calculates the likelihood of each registered word of the input speech from the output probability and a candidate word output unit that outputs the candidate word number and the likelihood of the candidate word in descending order of likelihood are provided. , Providing a candidate word presentation unit for presenting the candidate words output from the candidate word output unit to the user in the order of likelihood, and instructing the user to utter the word again In a voice dialog device that is equipped with a re-vocalization instruction unit that determines the correct candidate word by the user's voice input in response to the candidate word presented by the candidate word presenting unit, the likelihood and the correct answer rate are calculated for each candidate word. It is equipped with a word-likelihood-correctness rate correspondence data storage unit that stores data indicating a correspondence relationship between the words, and the accuracy rate of the correctness rate that determines whether the candidate word is confirmed to be correct or incorrect or the word is uttered again. It has a correctness rate threshold value storage unit that stores a threshold value, and the likelihood of candidate words and the likelihood by word obtained from the likelihood calculation unit −
The correct answer rate of the target candidate word is obtained from the data in the correct answer rate corresponding data storage unit, and this correct answer rate is compared with the threshold value stored in the correct answer rate threshold storage unit, and if the correct answer rate is higher than the threshold value, the candidate word A voice interaction device, comprising: a confirmation control unit that drives a presentation unit and drives a re-voice instruction unit when the correct answer rate is smaller than a threshold value.