JP2005257917A - Phonetic interpretion method, phonetic interpreting device, and phonetic interpretation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve the reliability and the flexibility of phonetic interpretation in a voice interaction system which uses computers, etc. <P>SOLUTION: In a phonetic interpretation device, reliability information is successively added to the recognized results of voice recognition by a reliability information adding means 14, moreover, finite-state transducers are generated, respectively by a rule type interpretation means 18a, a specific word extracting means 18b, an example type interpretation means 18c whose forms are different and a combination whose weight is the smallest; in short, whose reliability is the highest is selected from among the combined forms of these finite-state transducers to be outputted as a phonetic interpretation result in an interpretation result selecting means 15. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech interpretation method, speech interpretation apparatus, and speech interpretation program for a speech interpretation system that accepts a human request using speech.

Speech input means for inputting speech to the system; speech recognition means for recognizing speech as a sequence of words; speech recognition reliability evaluation means for evaluating the recognition reliability of each recognized word; and recognized words And a speech interpretation method of a speech interpretation system comprising speech interpretation means for interpreting speech input based on the reliability of recognition of each word, and speech input from the speech input means, A conventional model that recognizes words as a sequence of words by speech recognition means and identifies interpretation results for speech input from a sequence of words. Converts a sequence of words into an intermediate representation, and calculates the co-occurrence probability of the intermediate representation as a statistical model. There is a technique for avoiding an error in the interpretation result due to the influence of a voice recognition error (Non-patent Document 1).
"Spoken language understanding using conversion from input sentence to intermediate representation by statistical processing" Masakazu Masami, Koichi Tanigaki, Yoshinori Sakasaka, IEICE Transactions D-II Vol.J82-D-II No.2 pp. 169-177, February 1999

  In the conventional technique, since speech interpretation is performed using only the word sequence obtained by the speech recognition means, it is not possible to consider which word in the recognized word sequence is a recognition error. As a result, even if a word sequence is obtained with sufficient confidence by the speech recognition means, or even if a word sequence with little confidence is obtained, the word sequence itself is consistent. Thus, the result of the same interpretation is derived, and the effect of reducing the adverse effects due to the error in speech recognition is not sufficient. In addition, a sufficient amount of human utterance database is necessary for learning to use the statistical model used in the conventional technology, and another speech interpretation system is used to collect the data. Or, it was necessary to record a dialogue between a human and a computer system in a simulated manner by a dialogue between humans.

  According to the first aspect of the present invention, a voice input process for inputting voice into the system, a voice recognition process for recognizing the voice as a sequence of words, and a reliability of recognition for each recognized word. A speech interpretation method for a speech interpretation system, comprising speech recognition reliability evaluation processing to be evaluated, and speech interpretation processing for interpreting speech input based on a recognized word sequence and the reliability of recognition of each word. The speech (A) input from the speech input processing is recognized as a word sequence (B) by the speech recognition processing, and the words (B0), (B1),..., (Bn) included in this word sequence , (Cn) is calculated by the speech recognition reliability evaluation process for each of the above, and the word (B0),..., (Bn) and the recognition reliability are calculated. Sex evaluation values (C0), (C1), ..., From cn), we propose a sound interpretation to identify the voice interpretation result (D) for the audio (A) entered.

According to claim 2 of the present invention, the voice (A) input from the voice input process is recognized as a word series (B) by the voice recognition process, and the words (B0),. Bn) and the speech interpretation reliability (D) for the input speech (A) from the speech recognition reliability evaluation values (C0), (C1),..., (Cn) calculated by the speech recognition reliability evaluation process. When a word sequence including a registered specific word sequence (E0), (E1),..., (Ei) is input, the interpretation result corresponding to it is input. A rule-type interpretation process that outputs a sequence (F0), (F1),..., (Fj), and a specification that outputs (G) when a sequence of words including a specific type of word (G) is input. Seed word extraction processing and accumulated as examples in the database , (Hk) and the corresponding interpretation result series (I0), (I1),..., (Il), and (H0),. ), A rule type interpretation process, a specific type word extraction process, and a case type interpretation process are combined based on the case type interpretation process that outputs (I0),..., (Il). A speech interpretation method for specifying the best speech interpretation result (D) by an interpretation result selection process using a finite state transducer is proposed.
According to a third aspect of the present invention, in the speech interpretation method according to the second aspect, as a case of accumulating in the database, a set of word sequences (Ji), (Li) larger than a certain value as a speech recognition reliability evaluation value (Mi) is stored. A speech interpretation method is proposed in which the registered word series (Ji) and (Li) are used as examples in the case type interpretation process.

  According to a fourth aspect of the present invention, in the voice interpretation method according to the third aspect, the speech recognition reliability evaluation value (Mi) used in the case type interpretation process is actually spoken from the dialogue information recorded in the dialogue information recording means. An estimated word sequence (P) is extracted, the content of the dialogue information interpreted from the extracted word sequence (P) is specified as (P) ′, and the words included in the word sequence (P) Are extracted as (Q0), (Q1), (Q2),..., (Qn), the correct answer of the actually spoken word sequence corresponding to the word sequence (P) is identified as (R), and this word The correct answer of the content of the dialog information interpreted from the correct answer (R) of the sequence is specified as (R) ′, and the actually spoken word of the dialog information corresponding to the word string (Qi) is specified as (Si). In order to estimate the extracted word sequence (P) as a result candidate In addition to the index calculated in the above, the index obtained by the dialog information recorded in the dialog information recording means is calculated as the utterance unit index value and the word unit index value for each utterance unit and word unit, and the calculated utterance unit An index value and a word unit index value, a series of words in which the system recognizes the dialogue information read from the dialogue information recording means, and a series of words actually spoken recorded by the dialogue information recording means (P ) In the utterance unit and the word unit by determining the relationship between the correctness in the utterance unit and the word unit obtained by comparing the correct answer (R) of) with an index indicating binary values of 0 and 1 The reliability measure for evaluating the reliability, the index value of the speech recognition result for which the reliability should be evaluated is calculated based on the created reliability measure, and the reliability obtained by applying it to the evaluation formula of the reliability measure Evaluation To propose a sound interpretation methods with.

  According to the present invention, by using rules defined for speech interpretation and a specific type of word to be noticed, speech interpretation can be performed with a simple definition, and more cases can be collected by collecting examples. It is possible to change the voice interpretation so that it can be flexibly interpreted, and it incorporates a framework that rejects words with low reliability for speech recognition errors and incorporates speech interpretation errors. And robust speech interpretation. In addition, cases for case-type interpretation are not only created in a format in which data is collected in advance, the utterances are transcribed, and the corresponding interpretation results are labeled. It is also possible to collect high cases and create them automatically.

FIG. 1 shows the best mode for carrying out the present invention. 1 denotes a speech interpretation apparatus according to the present invention, and 20 denotes a database creation unit that creates a database used in the speech interpretation apparatus 10.
The speech interpretation apparatus 10 according to the present invention relies on the speech input means 12 for capturing the input speech 11, the speech recognition means 13 for speech recognition of the content of the input speech, and the recognition result recognized by the speech recognition means 13. A finite state transducer is generated by outputting the interpretation result according to the mode assigned in accordance with the data registered in the database 17 from the reliability information adding means 14 for adding sex information and the word sequence to which the reliability information is added. The finite state transducer generating unit 18 and the interpretation result selecting unit 15 for selecting the best one using the finite state transducer output from the finite state transducer generating unit 18. The result 16 is output and displayed as speech or text.

The finite state transducer generation unit 18 extracts a specific word sequence registered in the database 17 from the word sequence to which the reliability information is added by the reliability information adding means 14, and outputs a corresponding interpretation result. The specific type word extracting unit 18B that extracts and outputs a specific type of word registered in the database 17 from the interpretation unit 18A, and a series of words to which the same reliability information is added, and the reliability information is also added. The word type stored in the database 17 as a case is extracted from the word series, and the case type interpreting means 18C is configured to refer to and output the value of the corresponding interpretation result in the database 17.
The database creation unit 20 registers the case in the database 17 from the case file 21 through the case registration unit 22, and adds the speech recognition reliability information to the dialogue information from the dialogue information recording unit 23 through the case reliability information addition unit 24. A configuration may be considered in which highly reliable information is selected by the case selecting means 25 from the dialogue information to which the voice recognition reliability information is added, and the selected information is written in the database 17 through the case registering means 22.

The operation of each part will be described below according to each step display shown in FIG.
Step 1 The database 17 describes the speech interpretation rules used by the rule type interpretation means 18A, the specific type words used by the specific type word extraction means 18B, and the cases in the database 17 used by the case type interpretation means 18C. Prepare a file (system description file).
Step 2 Based on the speech interpretation rules defined in the system description file, the rule type interpretation means 18A for interpreting the word sequence of the speech recognition result speech-recognized by the speech recognition means 13 is activated. Specifically, it is expressed as a weighted finite state transducer that takes as input a series of words defined by rules and outputs the corresponding interpretation result. Here, a certain weight is added to the application of one rule.
Step 3 If the word recognized by the speech recognition means 13 includes a word of a type representing the specific contents defined in the system description file, the specific seed word having a function of interpreting it in the form of outputting the word The extraction means 18B is activated. Specifically, when a specific type of word is input, it is clearly indicated that it is a specific type of word, and the word is output with a constant weight, and other words are output with a higher weight. It is expressed as a weighted finite state transducer that outputs a word as it is.
Step 4 Case type interpretation means for performing speech interpretation using interpretations based on past cases, based on interpretation results corresponding to word sequences stored in the database 17 described in the system description file Start 18C. The operation is almost the same as in the case of the regular interpretation means 18A, and the weighting method of the weighted finite state transducer is different. As the weight when applying the interpretation by the case, a value obtained by inverting the logarithm sign of the occurrence probability of the case in the database 17 is used.
Step 5 The three weighted finite state transducers generated by the rule type interpretation unit 18A, the specific-type word extraction unit 18B, and the case type interpretation unit 18C are combined and input to the interpretation result selection unit 15.
Step 6 The input voice 11 is received by the voice input means 12.
Step 7 Voice recognition processing is performed by the voice recognition means 13.
Step 8 The reliability information adding means 14 converts the word sequence obtained as a result of the speech recognition into a finite state transducer in which the speech recognition reliability information is added. The higher the reliability evaluation value, the smaller the weight is given so that the word with high reliability is given priority.
Step 9 A finite state transducer to which reliability information of speech recognition is added is combined with a weighted finite state transducer that is a combination of a regular interpretation unit, a specific-type word extraction unit, and a case type interpretation unit. The path with the smallest weight is selected from the transducer (interpretation result selection means), and the output is output as the speech interpretation result 16.

In the database creation unit 20, as one method,
Step A1 A file (example file 21) is prepared that records a series of words actually spoken by a human and information describing the corresponding interpretation result.
Step A2 The case registration means 22 registers the word sequence and interpretation result set in the case file 21 in the case of the database 17.
As an alternative,
Step B1 The case reliability information adding means 24 calculates an utterance unit reliability evaluation value for the word sequence of the speech recognition result of the dialogue record, and the utterance unit reliability information for the word sequence of the speech recognition result of the dialogue record. Is added. As a method for calculating the utterance unit reliability evaluation value, the “voice recognition reliability evaluation method” proposed in the prior application “Japanese Patent Application No. 2003-27926” can be used. The calculation method of the utterance unit reliability evaluation value of the prior application will be described in detail later.
Step B2 The case selection means 25 selects a speech recognition result whose reliability evaluation value is larger than a separately designated threshold value, and saves it as a file (selected case file).
Step B3 The case registration means 22 registers the word series of the speech recognition result with high reliability written in the selected case file and the corresponding interpretation result in the case of the database 17.

Examples of each part will be described below. As an example, a speech interpretation system for querying a computer system for a bus timetable is used. In this system, a human can convey five types of information to the system: “bus stop to get on”, “bus stop to get off”, “bus stop”, “departure time or arrival time”, and “day of the week”.
Reliability information adding means 14
The sequence of words recognized by the speech recognition unit 13 is converted into a weighted finite state transducer to which information on the reliability of speech recognition of each word is added. As a reliability evaluation method used here, for example, the technique described in Reference 1 can be used.
(Reference 1)
“Recognition confidence scoring and its use in speech understanding systems”, Timothy J. Hazen, Stephanie Seneff and Joseph Polifroni, Computer Speech and Language 2002 vol.16 pp.46-67
The higher the reliability evaluation value, the smaller the weight is given so that the word with high reliability is given priority. In this embodiment, the number obtained by inverting the sign of the reliability evaluation value is used as the weight. Then, a framework for rejecting words with a constant weight at the time of state transition of the finite state transducer is introduced so that an erroneously recognized word can be rejected when there is a speech recognition error.

As an example, the weight of state transition when rejecting a word is +3.0, and the first candidate of the speech recognition result is “from bus center (reliability evaluation value +2.0) (reliability evaluation value +1.2)”, The expression of the weighted finite state transducer when the second candidate is “bus center (reliability evaluation value +2.0) to (reliability evaluation value +1.0) is (reliability evaluation value−0.8)” is It becomes as follows. In the following example, each row represents a transition rule, the first column is the current state number, the second column is the destination state number, the third column is the input symbol for the state transition, and the fourth column is the state transition. The output symbol at the time, the fifth column is the weight of state transition (0 in case of blank). If state number 0 is the start state and an element is included only in the first column, the state number is the end state. epsilon indicates a state transition in which nothing is input or nothing is output. The rejection of the word is realized by a state transition in which the output is epsilon.
========
1st row 2nd row 3rd row 4th row 5th row 0 1 epsilon epsilon
1 2 Bus Center Bus Center -2.0
1 2 Bus Center epsilon 3.0
2 3 to Bus Center -1.2
2 3 to epsilon 3.0
3
0 4 Bus Center Bus Center -2.0
0 4 Bus Center epsilon 3.0
4 5 to Bus Center -1.0
4 to 5 epsilon 3.0
5 6 is 0.8
5 6 is epsilon 3.0
========
Regular interpretation means 18A
A sequence of words of the speech recognition result is interpreted based on a predefined speech interpretation rule. Specifically, it is expressed as a weighted finite state transducer that takes as input a series of words defined by rules and outputs the corresponding interpretation result. A certain weight is added to the application of one rule.

As an example, if the weight for applying the rule is +1.0 and the word sequence “From the laboratory to the bus center” is input, “stopfrom = (busstop = (before the laboratory)), stopto = ( A weighted finite state transducer applying a speech interpretation rule that outputs a speech interpretation result of “busstop = (bus center))” is as follows.
========
0 1 epsilon stopfrom = (in front of the laboratory)), stopto = (busstop = (bus center)) 1.0
1 2 In front of the laboratory epsilon
2 3 to epsilon
3 4 Bus Center epsilon
Up to 4 5 epsilon
5
========
Specific word extraction means 18B
A function of interpreting a word of a type representing a specific content defined in advance by outputting the word is realized. Specifically, when a specific type of word is input, it is clearly indicated that it is a specific type of word, and the word is output with a constant weight, and other words are output with a higher weight. It is expressed as a weighted finite state transducer that outputs a word as it is.

For example, the weight for outputting a specific seed word is +2.0, and the weight for outputting a word included in a vocabulary that is not explicitly defined as a specific seed word but can be handled by the system is +3. , The word “bus center” of the type “busstop” representing the content “bus stop name is bus center”, and the word “8 o'clock” of the type “hour” representing the content “hour is 8 o'clock”. A weighted finite state transducer that handles the undefined word "is" is
========
0 1 Bus Center busstop = (Bus Center) 2.0
0 1 8:00 hour = (8) 3.0
0 1 is 3.0
1
========
Case type interpretation means 18C
Based on the case of the interpretation result corresponding to the word sequence stored in the database 17, speech interpretation is performed using the interpretation in the past case. The construction method is almost the same as in the case of the regular interpretation means 18A, and the weighting method of the weighted finite state transducer is different. As the weight when the interpretation by the case is applied, a value obtained by inverting the sign of the logarithm of the occurrence probability of the case in the database is used.

As an example, a series of words “10 o'clock in front of the laboratory” and an interpretation result of “stopfrom = (busstop = (in front of the laboratory)), departtime = (hour = (8))” are stored in the database. If the occurrence probability in the database is 0.001, the weighted finite state transducer to which this case is applied is as follows.
========
0 1 epsilon stopfrom = (in front of laboratory)), departtime = (hour = (8)) -log (0.001)
1 2 In front of the laboratory epsilon
2 3 10 o'clock epsilon
3 4 epsilon
4
========
Interpretation result selection means 15
The rule type interpretation means 18A, specific-type word extraction means 18B, and case type interpretation means 18C can be used to give interpretation results for word sequences. Since it is impossible to determine whether the obtained interpretation result is valid, a series of words accepted by the combined weighted finite state transducer by combining the weighted finite state transducers configured in parallel with each other. The one with the smallest weight is selected and adopted as the interpretation result. In that case, to interpret the word sequence of arbitrary length and the sequence of words in the speech comprehension rule, the specific species word, or the sequence of words in the case of the example in the database, rule-type interpretation The weighted finite state transducers of the means 18A, specific-species word extraction means 18B, and case type interpretation means 18C are combined in parallel, and the closure (the path from each end state of the transducer to the start state exists, any number of times This is used as a means for selecting the interpretation result.
Case registration means 22

This is a means for registering a series of words and a corresponding interpretation result as examples stored in the database used by the case type interpretation means 18C. A file in which a set of interpretation results corresponding to a word sequence is written is input, and the set is registered in the database as an example. There are two types of files that are input. The first type is a sequence of words (transcriptions) actually spoken by humans based on the conversation records between humans and computer systems, and the interpretation results. The labeled file is written (case file), and the second type is a file created by the case selecting means 25 (selected case file).
Case selection means 25
The database 17 used by the case type interpretation means 18C uses a dialogue record that is recorded only by the system but is not maintained, instead of using a maintained dialogue record that is transcribed or labeled with meaning content. This is a method for enabling the registration of cases. From the set of interpretation results corresponding to the word sequence of the speech recognition result to which the reliability information is added by the case reliability information adding means 24, the one having a reliability evaluation value higher than a certain value is selected and stored in the database 17 Write to a prepared file (selected case file).
Case reliability information adding means 24
The case reliability information adding means 24 calculates an utterance unit reliability evaluation value for the word sequence of the speech recognition result of the dialog record, and adds the utterance unit reliability information to the word sequence of the speech recognition result of the dialog record. To the case selection means 25. As the reliability evaluation method used at that time, the “speech recognition reliability evaluation method proposed in Japanese Patent Application No. 2003-27926” can be applied.

[Flow of implementation process]
First, the operation of the speech interpretation apparatus 10 will be described. In the following example, for simplification of description, only the part of the speech interpretation system that handles the bus timetable that interprets information related to “route points” will be described. The same processing is performed for the part that interprets other contents.
The speech interpretation rules are recorded in an XML format file (A.xml: system description file) in the following format.
<class name = “specify_via” type = “Action”>
<entry> via Kannonzaka </ entry>
<entry> The waypoint is a ship </ entry>
<class>
In the above example, the type of utterance “specify_via” is defined.

The definition of the specific type word is recorded in the same XML file (A.xml) in the following format.
<class name = “via” type = “Key”>
<entry> Kannonzaka </ entry>
<entry> Ferry </ entry>
<entry> Hiromachi Bridge </ entry>
</ class>
In the above example, a word of the type “via” is defined.

On the other hand, the database referred to by the case type interpretation means 18C is recorded in the system description file provided in the database 17 as an example of the interpretation result corresponding to the word series in the following format.
Via Kannonzaka: specify_via via ＝ (Kannonzaka)
Go through Funako: specify_via via ＝ (Hiromachi Bridge)
A file containing the above examples is a case file in a format in which a set of interpretation results corresponding to a word sequence actually spoken by a human is collected in advance, the utterances are transcribed, and the corresponding interpretation results are labeled. Or by referring to the dialogue record between the human and the system, and using the speech recognition reliability evaluation method of Japanese Patent Application No. 2003-027926 described later, the speech unit reliability evaluation value is calculated, and the speech recognition result speech A file having the following format to which the unit reliability evaluation value is added is created, and a selection example file in which a reliability evaluation value larger than a certain threshold (for example, greater than 0) is selected is created.
20021114-02-02_08,0.861485: [CONCEPT] specify_via via = (Ship)
20021114-02-02_21, -1.65177: Until 1 o'clock [CONCEPT] specify_time hour = (1)
The format of the above file is an utterance ID, an utterance unit reliability evaluation value, a word sequence of a speech recognition result, and a speech interpretation result corresponding to the word sequence of the speech recognition result below [CONCEPT].

The cases stored in the database are taken out and recorded in the system description file (A.xml) in the following format.
<example>
<entry acttype = “specify_via” concept = “via = (Kannonzaka)” prob = “0.01”> Via Kannonzaka </ entry>
<entry acttype = “specify_via” concept = “via = (Hiromachibashi)” prob = “0.001”> Passing the boat </ entry>
</ example>
In the above example, the word sequence “Via Kannonzaka” is specified in the database with a probability of 0.01, which specifies the transit location called Kannonzaka, and the transit location called Hiromachi Bridge is also specified. This means that the word sequence “passing through Hiromachi Bridge” is included in the cases in the database with a probability of 0.001.

  The system description file (A.xml) in which the contents as described above are recorded is processed, and each weighted finite state transducer is created by the rule type interpretation means 18A, the specific seed word extraction means 18B, and the case type interpretation means 18C. To do. In addition, when a specific type word is included in the word series defined by the rule type interpretation unit 18A or the case word series used by the case type interpretation unit 18C, the same rule is applied even if the word type is replaced. Make changes as you can. It should be noted that the weight when applying the rule by the rule type interpretation means 18A is +1.0, and the specific kind word extraction means 18B extracts the specific kind words (in this example, “Kannonzaka” “Funako” “Hiromachi Bridge”). The weight when accepting +2.0, the weight when accepting words that are not defined as a specific species (all words that are not defined as specific species: in this example, “Ut”, “Bus”, “Is”, etc.) 0.0. FIG. 2 shows an example of a weighted finite state transducer in which they are combined in parallel and closed.

In the figure, the initial state is 0, each state is indicated by a circle, and the number in the circle is the state number. A state indicated by a double circle indicates an end state. In addition, the state transition arrows display information on input symbols, output symbols, and weights at the time of state transition in the format of “input symbol, output symbol / weight”. The epsilon of the input symbol and the output symbol indicates a state transition in which there is no input symbol or output symbol.
At this time, by using this weighted finite state transducer and obtaining a state transition sequence that minimizes the weight, a speech interpretation result can be obtained.

Based on the above speech interpretation rules and the definition of a specific species word, the word sequence “via (word) included in (route) species” and “passage (word included in“ route ”species)” Interpretation is possible. As an example, the interpretation result for the word sequence “passage is Hiromachi Bridge” is as follows: specify_via via = (Hiromachi Bridge). The state transition sequence at that time is 0->1-> 6- (Enter "route")-> 7- (Enter "ha")-> 8- (Enter "Hiromachibashi")->9-> 10 (end state), weight is 1. At this time, the output from the speech interpretation device 10 is
1.0 specify_via via = (Hiromachibashi)
And output as a set of interpretation weights and interpretation results.

Also, like “Et Kannonzaka”, it is not included in the speech interpretation rules, but via = (Kannonzaka) (Kannonzaka) Interpretation results are obtained. The state transition sequence at that time is 0->1-> 11- (input "Eto")->12->13->24->0->1-> 11- (input "Kannonzaka") ->12-> 13 (end state), the weight is 5. At this time, the output from the speech interpretation device 10 is
5.0 via = (Kannonzaka)
become that way.

Furthermore, when the word sequence “passing through the boat” is input, the word sequence “passing through the boat” is included in the database examples, and is interpreted by the case type interpretation means 18C. Since the word that is not defined as the specific type word “is a bus” is treated as a non-specific type word by the specific type word extracting means 18B, an interpretation result of “specify_via via = (ship)” is obtained. In this case, the state transition sequence is 0->1-> 19- (enter "Ship")-> 20- (enter "O")-> 21- (enter "Pass")-> 22- >23->24->0->1-> 11- (enter "bus")->12->13->24->0->1-> 11- (enter "is")-> At 12-> 13 (end state), the weight is 8 (-log (0.001) = 3). At this time, the output from the speech interpretation device 10 is
8.0 specify_via via = (Funako)
And output as a set of interpretation weights and interpretation results.

And when the following speech recognition results are obtained,
Kannonzaka (reliability evaluation value +1.5) via (reliability evaluation value -0.5) (reliability evaluation value +1.0)
Kannonzaka (reliability evaluation value +1.5) via (reliability evaluation value -1.0) (reliability evaluation value +1.0)
When the weight when rejecting the word of the recognition result is +3.0, the reliability information is added by the reliability information adding means 14 (the value obtained by inverting the sign of the reliability evaluation value is used as the weight). The finite state transducer is as shown in FIG. Two weighted finite state transducers are treated in such a manner that the output symbol of the weighted finite state transducer representing the word sequence of the speech recognition result of FIG. 3 is treated as the input symbol of the weighted finite state transducer representing the speech interpreting means of FIG. By synthesizing and obtaining a state transition sequence that minimizes the weight on the synthesized weighted finite state transducer, an optimal speech interpretation result can be obtained. In the example of Fig. 2 and Fig. 3, the state transition sequence of Fig. 3 is 0->1-> 2- ("Kannonzaka" is input, "Kannonzaka" is output)->3-("Ha" is input , "Ha" is rejected)-> 4- (input "via", output "via")-> 5 (end state), output "via Kannonzaka", state transition on the side of Figure 2 The series has the smallest weight when 0->1-> 2- (input "Kannonzaka")-> 3- (input "via")->4-> 5 (end state) The weight on the side of FIG. 3 is 0.5, and the weight on the side of FIG. Finally,
1.5 specify_via via = (Kannonzaka)
The interpretation result is obtained. Suppose that there was no framework for rejecting the words in Fig. 3, and applying the sequence of words "Kannonzaka via" to the state transition of the weighted finite state transducer in Fig. 3, Because it cannot be interpreted,
8.0 via = (Kannonzaka)
Therefore, it cannot be interpreted that the utterance type is “designating a waypoint”.

The speech recognition reliability evaluation method proposed in Japanese Patent Application No. 2003-27926, which is a prior application, will be described below. In the previous application, a speech recognition reliability evaluation method and apparatus provided in the speech dialogue system has been proposed. The following description will focus on the speech recognition reliability evaluation method and apparatus as much as possible.
FIG. 4 shows an example of a procedure for executing the speech recognition reliability evaluation method used in the previously filed spoken dialogue system. Here, the description will be made assuming that the speech recognition reliability evaluation process is started after the end of the dialogue.

In step SP1 shown in FIG. 4, dialogue information is read from the dialogue information storage means.
In step SP2, a sequence (P) of words estimated to be spoken by the user for requesting information is extracted.
In step SP3, the user's information request content interpreted from the estimated word sequence (P) is specified as (P) ′.
In step SP4, the words included in the word sequence (P) extracted in step SP2 are identified as (Q0) (Q1) (Q2).
In step SP5, the correct answer of the word sequence actually spoken by the user corresponding to the word sequence (P) for information request is specified as (R).
In step SP6, the correct answer of the information request content of the user interpreted from the correct answer (R) of the word series is specified as (R) ′.
In step SP7, the word actually spoken by the user corresponding to the word string (Qi) for the information request is specified as (Si).

In step SP8 shown in FIG. 5, the voice recognition means recognizes the user's voice at each point of time between the user and the system, and estimates a word sequence (P) that is estimated to be spoken by the user as a result candidate. In addition to the calculated index, the conversation unit index value and the word unit index value are calculated for the utterance unit and the word unit for the index obtained when the exchange between the user and the system ends.
In step SP9, the calculated utterance unit index value and the word unit index value, the word sequence in which the system recognizes the user's utterance, the correct answer of the actually uttered word sequence recorded by the dialog information recording means, A measure for evaluating the reliability of the utterance unit and the word unit is obtained by obtaining the relationship between the correctness in the utterance unit and the word unit obtained by comparing the two values and the index indicating the binary value of 0 and 1. create.
In step SP10, the word series (P) is extracted as in step SP2. In step SP11, as in step SP3, the user's information request content interpreted from the word sequence (P) is specified as (P) ′.

In step SP12, the word string (Q0) (Q1) (Q2)... (Qn) is specified as in step SP4.
In step SP13, the index value is calculated for each utterance and each word as in step SP8.
In step SP14, the calculated utterance unit index value and word unit index value are applied to the utterance unit reliability scale and word unit reliability scale created in step SP9, and the utterance unit reliability evaluation value and word unit reliability evaluation value are calculated. Calculate

FIG. 6 shows an embodiment of a voice dialogue system realized by a computer and a voice recognition reliability evaluation apparatus for evaluating the voice recognition reliability of the voice dialogue system.
As is well known, the computer includes a CPU 31 for decoding and executing a program, a read-only memory ROM 32, a RAM 33 for storing and executing the program, an input port 34, an output port 35, and the like. The Note that a loudspeaker 41 is connected to the output port 35, a speaker is driven by the loudspeaker 41, and a response from the system is output by voice.

The RAM 33 includes a program for configuring the information request input unit 33A, a program for configuring the dialog information recording unit 33B, a program for configuring the format conversion unit 33C, a program for configuring the speech recognition unit 33D, and a dialog end determination unit 33E. The programs constituting the information providing means 33F and the programs constituting the information providing means 33F are stored, and the voice dialogue system 100 is constituted by these programs and the CPU 31.
In addition to the configuration of the speech dialogue system 100, the speech recognition reliability evaluation apparatus 200 is configured to configure a program that constitutes the word sequence (P) extraction means 33G in the RAM 33 and an information request content (P) 'identification means 33H. Program, word string Q0, Q1, Q2... Program constituting extraction means 33I, program constituting correct word specifying means 33J, program constituting correct information content specifying means 33K, program constituting speech word specifying means 33L, speech Program that constitutes unit index value calculation means and word unit index value calculation means 33M, program that constitutes utterance unit reliability scale creation means and word unit reliability scale creation means 33N, utterance unit and word unit reliability evaluation value calculation means 33P is stored, and these programs are spoken dialogue systems. Has been the voice recognition reliability evaluation method consists voice recognition reliability evaluation device 200 executed by the CPU31 with programs constituting 00 is executed.

Examples of each part will be described below. As an example, a bus timetable guidance system using voice dialogue is used. In this system, the user has six types of information: “bus stop to get on”, “bus stop to get off”, “bus stop”, “time”, “departure time / arrival time”, and “day of the week”. The system informs the user of the departure time of the corresponding bus.
Details of the dialogue information recording means, index value calculating means, and reliability measure creating means in the system are as follows.

Dialog information recording means The voice dialog system understands user information requests as attribute-value pairs. This is called a slot. The system uses a data structure consisting of a plurality of slots to hold user information requests and to proceed with the dialogue. This data structure is called a dialog state. The dialogue state in the bus timetable guidance system is composed of the following six slots.
(STOP_FROM.value): "Bus stop to get on"
(STOP_TO.value): "Get off the bus stop"
(VIA.value): "Bus stop"
(TIME.value): "Time"
(TIME_TYPE.value): “Specify either departure / arrival time” (DAY.value): “Day of the week”
A semantic expression included in the user's utterance, such as “fill value (s) in slot (S)”, obtained as a result of the system interpreting the user's utterance, is called a dialogue action. For example, the dialogue action for the user's utterance “From the bus center to ** Gakuin University” is expressed as follows.
(SET-STOP_FROM (XX Bus Center)) (SET-STOP_TO (** Gakuin University))
The system records the user's utterance voice at each point of dialogue with the user and the information request contents of the user who have been confirmed when the dialogue is completed as a dialogue record in the form of a dialogue state.

Utterance unit index value calculation means For each user's utterance voice recorded, the same voice recognition means as that used during conversation recording is used to recognize the user's utterance voice and estimate a maximum of five sequences of words spoken by the user. Then, for each recognition result candidate, index values U1 to U19 in units of utterances shown in FIG. 7 are calculated based on the word sequence of each recognition result candidate and the score used in the recognition process.
Utterance unit reliability scale creation means Based on the index value calculated by the utterance unit index value calculation means, a sequence of words obtained as a result of recognizing the user's utterance (P), and a corresponding utterance is obtained by human transcription. A scale for calculating the reliability with which the dialogue action when (P) is interpreted coincides with the dialogue action when (Q) is interpreted, with the correct answer of the sequence of words as (Q). Literature [“Recognition confidence scoring and its speech understanding systems”, Timothy J. Hazen, Stephanie Seneff and Joseph Polefroni, Computer Speech and Language 2002 vol.16 pp.46-67]

式２の、ｔはしきい値、p（r|correct）、p（r|incorrect）は、それぞれrが正しい発話音声認識結果に基づく指標値、誤った発話認識結果に基づく指標値であったときの、ガウス密度関数（式３、式４）であり、P（correct）、P（incorrect）は、それぞれ正しい発話音声認識結果、誤った発話音声認識結果を観測する事後確率である。発話音声認識結果が正しいかったときのrの平均をμ（correct）、分散をσ²（correct）、発話音声認識結果が誤っていたときのrの平均をμ（incorrect）、分散をσ²（incorrect）とすると、p（r｜correct）、p（r｜incorrect）は、それぞれ式３、式４で計算できる。

Dialogue action when (P) is interpreted as dialogue action when (P) is interpreted using the column vector f ^ and the column vector p ^ of the same dimension calculated by the utterance unit index value calculation means Is evaluated using a one-dimensional value R according to Equations 1 and 2.

In Equation 2, t is a threshold value, and p (r | correct) and p (r | incorrect) are index values based on the correct speech recognition result for r and index values based on the incorrect speech recognition result, respectively. Are Gaussian density functions (Equation 3 and Equation 4), where P (correct) and P (incorrect) are posterior probabilities of observing correct utterance speech recognition results and incorrect utterance speech recognition results, respectively. When the speech recognition result is correct, the average of r is μ (correct), the variance is σ ² (correct), when the speech recognition result is incorrect, the average of r is μ (incorrect), and the variance is σ ² If (incorrect), p (r | correct) and p (r | incorrect) can be calculated by Equation 3 and Equation 4, respectively.

The utterance unit reliability scale creation means interprets the utterance speech recognition result and the utterance recognition result by using a small amount of records in which the transcript of the corresponding utterance speech exists as a correct answer to the utterance speech recognition result. The optimal value of each element of the vector p ^ necessary for applying the equations (1)-(4) is obtained from the relationship with the binary expression as to whether the dialogue action to be performed is correct, and the corresponding t, Find μ (correct) σ ² (correct), μ (incorrect), and σ ² (incorrect). P ^ is determined by setting an initial value based on Fisher's linear discriminant analysis method and repeatedly updating the value of each element based on the hill-climbing method so that the correct / incorrect classification error is minimized.

Utterance unit reliability evaluation value calculation means Apply the utterance unit index value calculated for the speech recognition result in the conversation recording different from the time when creating the reliability scale to the utterance unit reliability scale (Equation 2), The utterance unit reliability evaluation value is calculated.
Word unit index value calculation means Using the same user utterance speech recognition results as those used in the utterance unit index value calculation means, each recognition result based on the word sequence of each recognition result candidate and the score used in the recognition process Index values W1 to W13 are calculated in units of words shown in FIG. The “utterance score” in FIG. 8 is calculated using the scale created by the utterance unit reliability scale creating means.

Word unit reliability scale creation means , where (R) is a word obtained as a result of recognizing a user's utterance, and (S) is a word answer obtained by human transcription of the corresponding utterance. By performing the same processing, a reliability measure in units of words can be created.
Word unit reliability evaluation value calculation means Apply the word unit index value calculated for the speech recognition result in the conversation recording different from the time when creating the reliability scale to the word unit reliability scale (Formula 2), The word unit reliability evaluation value is calculated.

Data collection is performed using a spoken dialogue system. An example of dialogue is shown below. In the dialogue example, S indicates an utterance of the system, and U indicates a sequence of words obtained as a result of the system recognizing the user's utterance. The parentheses are the transcripts of actual user utterances.
U: XX bus center at 8:00 1:00 to Gakuin University
(Depart from XX Bus Center, bound for Gakuin University)
S: Yes U: 3pm at 7pm bus leaving 11am
(Please tell me the departure bus for 11 cars)
S: Yes S: Where is your departure?
U: XX bus center
(This is the ○○ Bus Center)
S: Yes S: Where is the bus stop?
U: ** At Gakuin University
(※※ Gakuin University)
S: Yes S: Do you want to specify a stopover?
U: Not (Not)
S: Yes S: Which timetable is weekday, Saturday, Sunday?
U: Holiday (holiday)
S: Yes From ○○ Bus Center ** No way to go to Gakuin University. Is there a timetable for arrival by 19:00 on holidays?
U: Timetable for departure at 12:00
(Departure timetable at 11 o'clock)
S: Desune leaving at 12:00?
U: Depart at 11:00
(It's 11 o'clock)
S: Desnay leaving at 11 o'clock?
U: Yes
(Yes)
S: From XX bus center to ** Gakuin University.
U: Yes
(Yes)
S: From the bus center ○○ Bus Center ○○ 36 Routes △△ ※※ To Gakuin University ※※ The departure time for the 11:00 bus to Gakuin University is 5 minutes 55 minutes.

The dialogue information recording means 23 records the user's utterance voice at each point in the dialogue and the user's information request content (explained below) determined at the end of the dialogue.
(TIME_TYPE.DEPARTURE) (TIME.11) (DAY. Holiday)
(STOP_TO. ** Gakuin University) (VIA.ARBITRARY) (STOP_FROM.XX Bus Center)
Next, the user's utterance voice recorded at each time point of the dialogue is recognized using the same voice recognition means (corresponding to the voice recognition means 13 shown in FIG. 1) when data is collected, and a maximum of five recognitions are recognized. The result candidate and the speech recognition score file 1 are output. The contents of file 1 are as follows (only the most likely candidates are extracted below).

sentencel: at 11 o'clock
wseq1: </ s> at <s> 11:00 departure
phseq1: silB | ju: i ch iji | haq sh a | de | silE
score1: −6708.832031
=== word alignment begin ===
id: from to n_score CM−meam CM−var CM−min CM−max
applied HMMs (logical [physical] or [pseudo])
――――――――――――――――――――――――――――――――――
(0: 0 52 −22.2466−1.0495 0.7000−2.8045 0.0000 silB)
(1:53 166−26.9416−2.204 1.8629−9.2010 0.0000 j + u: [j + u] j−u: + i [y−u: + i] u: −i + ch [u−i + ch] i−ch + i ch−i + j i−j + i j− i + h [v-i + h])
(2: 167 232-27.1568-2.9200 1.8325-7.3065 0.0000 i−h + a h−a + q a−q + sh q−sh + a sh−a + d [y−a + d])
(3: 233 250 −27.1151−2.9540 2.4125−8.7062 −0.4162 a−d + e d−e)
(4: 251 254 -28.4927-6.0016 1.6088-8.5824 -4.3282 silE)
re-conputed AM score: −6645.894531
(log−likelihood−ratio: −576.595320 (normalized to−2.261158))
Acoustic-score: −6645.894531 (normalized to−26.062331)
=== Word alignment end ===
In the contents of file 1,
A value obtained by dividing the re-computed AM score by the number of words corresponds to the index U6 in FIG.

The difference between to and from of each word corresponds to the index W7 in FIG.
The n-score of each word corresponds to the index W12 in FIG.
CM-mean of each word corresponds to the index W5 in FIG.
The parallel root of CM-var of each word corresponds to the index W4 in FIG.
CM-min of each word corresponds to the index W3 in FIG.
Further, the number of recognition result candidates corresponds to the index W9 in FIG.
Then, from the recognition result candidates output to the file 1, the word appearance frequencies in the recognition result candidates necessary for obtaining the indices U10 and U12 in FIG. 7, the indices U10 and U12 in FIG. 8, and the index W11 in FIG. calculate. As shown below, the positions of matching words between the recognition result candidates are aligned, and the frequency at which a certain word in the recognition result candidate appears in the same place in other recognition results is calculated. In the following example, the appearance frequency of the word “departure” is 100%, and the appearance frequency of the word “1 o'clock” is 60%.

Maximum likelihood candidate: 11:00 departure, 2nd candidate: 10: 1 departure, 3rd candidate: 15: 1 departure, 4th candidate: 14: 1 departure, 5th candidate: 12:00 departure until departure Using vocabulary definitions and phrase definitions (excerpts below) used to parse word sequences,
(1 hour 1 hour nil 1)
(11:00 Juichiji time nil 11)
(12 o'clock Juniji time nil 12)
(14:00 Jyuyoji time nil 14)
(15 o'clock jugoji time nil 15)
(De in particle de * *)
(Until particle made: made *)
(Departure Departure)
(Common noun_starting phrase
(Common noun_departure (opt (or particle c)
(Time clause
(Time (opt common noun _ table))
(opt (or
Particle no particle particle particle particle (particle particle (opt particle particle))
Presented word final particle interjective usage))))
A series of words (referred to as unknown phrases) that cannot be correctly parsed is searched from the result of parsing each recognition result candidate (excerpted below), and index U19 in FIG. 7 and index W13 in FIG. 8 are calculated.

(11 o'clock) time clause | (by departure) common noun _ departure clause
(10 o'clock) time clause | (1 o'clock) time clause | (by departure) common noun _ departure clause
(15 o'clock) time clause | (1 o'clock) time clause | (by departure) common noun _ departure clause
(14:00) Time clause | (1) Time clause | (On departure) Common noun _ Departure clause
(At 12 o'clock) time clause | (departure) common noun_starting clause | (until) unknown clause And the dialogue action (excerpt below) obtained as a result of interpretation by each recognition result candidate system for each user utterance during dialogue Then, the index U18 in FIG. 7 is compared with the information request contents of the user determined at the end of the dialogue, and U16 and U17 in FIG. 7 are calculated.

(SET−TIME (11)) (SET_TYPE (DEPARTURE))
(SET−TIME (10)) (SET_TYPE (DEPARTURE))
(SET−TIME (15)) (SET_TYPE (DEPARTURE))
(SET−TIME (14)) (SET_TYPE (DEPARTURE))
(SET−TIME (12)) (SET_TYPE (DEPARTURE))
In addition, a dialogue sequence obtained from a sequence of transcribed words and a dialogue obtained from a recognition result candidate indicate whether a sequence of words transcribed by humans and each word of each recognition result candidate match. Compare whether all actions match, and express the result as binary values of 0 and 1.

A series of transcribed words and dialogue actions obtained from them
11 o'clock departure (* "departure" is not included in the recognition vocabulary)
(SET−TIME (11)) (SET−TIME_TYPE (DEPARTURE))
The index values obtained from the above are summarized, and the utterance unit index is written in the file 2 and the word unit index is written in the file 3, respectively. The index W11 in FIG. 8 is represented by a symbol because it cannot be calculated yet.
Contents of file 2 (excerpt)
ID correct U6 U10 U12 U16 U17 U18 U19
20021125−02−04_10 1 −26.062331 84 76.5517241379311 0.5 0.5 0.2 0
Contents of file 3 (excerpt)
ID correct W3 W4 W5 W7 W9 W10 W11 W12 W13
11 o'clock 1 −9.2010 1.8629 −2.2024 113 0.04 5 [UTTERANCE_SCORE]
−2.69416 0
Departure 0 −7.3065 1.8325 −2.9200 65 1 5 [UTTERANCE_SCORE] −27.1568
0
0 −8.7062 2.4125 −2.9540 17 1 5 [UTTERANCE_SCORE] −27.1151
0
The above processing is performed on the user utterances of a plurality of dialogue records, and the utterance unit reliability scale is created using the file 2 in which the utterance unit indexes are collected. Each parameter of the created reliability measure is shown below. The value corresponding to each index name is each element of the vector p in Equation 1, and the values corresponding to CORRECT_GAUSSIAN-INCPRRECT_GAUSSIAN are the mean μ, variance σ ² , and posterior probability P from the left, respectively.

top_choice_average_acoustic_score 0.4071058
top_choice_average_nbest_purity 0.0590583
average_nbest_purity 0.0049878
top_choice_average_consistent_concept_rate 2.2632635
top_choice_inconsistent_concept_rate −8.8873672
top_choice_average_concept_frepuency 6.3129411
top_choice_unparsed_bunsetsu_rate 0.0453356
CORRECT_GAUSSIAN: −25.7632580 1.4265214 0.47988507
INCORRECT_GAUSSIAN: −30.7077694 4.5412450 0.52011490
The reliability evaluation value of the utterance unit is obtained by calculating Expression 2 by applying the utterance unit index value to the scale having the above parameters. The symbol [UTTERANCE_SCORE] in the file 3 is calculated as the utterance unit reliability evaluation value of the corresponding utterance, and is replaced with that value.

Updated file 3 contents (excerpt)
11 o'clock 1 -9.2010 1.8629 -2.2024 113 0.04 5 -2.07607612356411 -26.9416 0
Departure 0−7.3065 1.8325−2.9540 17 1 5 −2.07607613256411−27.1151 0
0−8.7062 2.4125−2.9540 17 1 5 −2.07607613256411−27.1151 0
The parameters of the reliability measure for which the word unit reliability measure is created using the updated file 3 are shown below.
minimum_acoustic_score 0.1656796
acoustic_score_standard_deviation −0.0581996
mean_difference_from_maximum_score 2.0266259
number_of_acoustic_observations 0.0187305
square_nbest_purity 1.6731033
number_of_nbest −0.1339422
utterance_score 0.1814701
noumalized_score −0.1636039
unparsed_bunsetsu_violation −0.5617342
CORRECT_GAUSSIAN: 1.3205144 2.4545729 0.58425194
INCORRECT_GAUSSIAN: −3.6866648 4.1660976 0.41574803
The word unit reliability evaluation value is obtained by calculating the formula 2 by applying the word unit index value to the scale having the above parameters.

Here, the utterance unit index value is written in the file 4 and the word unit index value is written in the file 5 for the utterance in another dialogue recording whose reliability is to be evaluated.
Contents of file 4 (excerpt)
20021127-01-03_06-24.188667 100 90 1 0 0.157894736842105 0
Contents of file 5 (excerpt)
○○○○ Station -6.5979 1.2797-1.4625 123 1 4 [UTTERANCE_SCORE] -24.1581 0
Departure 1-6.3325 1.4448-1.9726 67 1 4 [UTTERANCE_SCORE] -25.0312 0
The utterance unit reliability evaluation value is calculated using the utterance unit index value of the file 4 and the utterance unit reliability scale.
20021127-01-03_06 1.84255589061485
In addition, the symbol [UTTERANCE_SCORE] of the file 5 is replaced with the utterance unit reliability evaluation value, and the updated content of the file 5 (excerpt)
○○○○ Station -6.5979 1.2797-1.4625 123 1 4 1.84255589061485-24.1581 0
Departure 1-6.3325 1.4448-1.9726 67 1 4 1.8425558906 1485-25.0312 0
The word unit reliability evaluation value is calculated using the updated utterance unit reliability evaluation value of the file 5 and the word unit reliability scale.
○○○○ Station 2.24738419391751
Departure 1.19691584999685
According to the speech recognition evaluation method described above, the reliability of the evaluation value for speech is high. Therefore, for example, by applying it to a speech decoding device or a dialogue system, the degree of understanding of utterances can be improved, and dialogue with people can be executed smoothly.

  The speech interpretation method and apparatus according to the present invention described above is realized by causing a CPU provided in a computer to decode and execute a speech interpretation program described by a computer-readable code string. The speech interpretation program may be recorded on a computer-readable recording medium and installed in the computer for execution, or may be installed in the computer through a communication line.

  The speech interpretation method and apparatus according to the present invention can be used in, for example, a speech dialogue system, an automatic guidance system, and the like.

The block diagram for demonstrating one Example of the speech interpretation apparatus by this invention. The flowchart for demonstrating operation | movement of the speech interpretation method by this invention. The same flowchart as FIG. The flowchart for demonstrating the procedure of the speech recognition evaluation method proposed previously. 5 is a flowchart for explaining the continuation of FIG. The block diagram for demonstrating the structure of the speech recognition evaluation apparatus proposed previously. The figure for demonstrating operation | movement of the speech recognition evaluation apparatus shown in FIG. The same figure as FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speech interpretation apparatus 11 Input speech 12 Speech input means 13 Speech recognition means 14 Reliability information addition means 15 Interpretation result selection means 16 Speech interpretation result 17 Database 18 Finite state transducer generation part 18A Regular type interpretation means 18B Characteristic type word extraction means 18C Case type interpretation means 21 Case file 22 Case registration means 23 Dialog information recording means 24 Case reliability information addition means 25 Case selection means

Claims (9)

On the computer,
Speech input processing for inputting speech information to the system, speech recognition processing for recognizing speech as a sequence of words, speech recognition reliability evaluation processing for sequentially evaluating the recognition reliability for each recognized word, and recognition A speech interpretation method of a speech interpretation system that executes speech interpretation processing for interpreting speech information input based on a series of words and reliability of recognition of each word,
The voice information (A) obtained by the voice input process is recognized as a word series (B) by the voice recognition process, and the words (B0), (B1),..., (Bn) included in this word series. , (Cn) is calculated by the speech recognition reliability evaluation process for each of the words, and the words (B0),. A speech interpretation method characterized by identifying speech interpretation results (D) for speech (A) input from reliability evaluation values (C0), (C1),. The speech interpretation method according to claim 1,
When a sequence of words including a registered sequence of specific words (E0), (E1),..., (Ei) is input, a sequence of interpretation results (F0), (F1),. Fj) to output a regular interpretation process;
A specific word extraction process for outputting the specific type of word (G) when a series of words including a specific type of word (G) is input;
There are pairs of word sequences (H0), (H1),..., (Hk) stored as examples in the database and corresponding interpretation result sequences (I0), (I1),. A case type interpretation process for outputting a series (I0), (I1),..., (Ii) of the interpretation result when a series of words including the word series (H0),. The speech that identifies the best speech interpretation result (D) by executing interpretation result selection processing using a finite state transducer that combines the above-mentioned regular type interpretation processing, specific type word extraction processing, and case type interpretation processing based on Interpretation method. The speech interpretation method according to claim 2,
As a case of accumulating in the database, a set of word sequences (Ji), (Li) having a speech recognition reliability evaluation value (Mi) larger than a certain value is registered, and the registered word sequences (Ji), (Li) Is used as a case in the case type interpretation process described above. The speech interpretation method according to claim 3,
The speech recognition reliability evaluation value (Mi) used in the case type interpretation process is
Extracting a sequence (P) of words estimated to be actually spoken from the dialogue information recorded in the dialogue information recording means,
The content of the dialogue information interpreted from the extracted word sequence (P) is specified as (P) ′,
Extract the words included in the word series (P) as (Q0), (Q1), (Q2), ..., (Qn),
The correct answer of the word sequence actually spoken corresponding to the word sequence (P) is identified as (R),
The correct answer of the content of the dialogue information interpreted from the correct answer (R) of the word sequence is specified as (R) ′,
The word actually spoken in the dialogue information corresponding to the word string (Qi) is identified as (Si),
In addition to the index calculated to estimate the extracted word sequence (P) as a candidate for the result, the index obtained from the dialog information recorded in the dialog information recording means is expressed in utterance units and word units, respectively. Utterance unit index value and word unit index value for the calculated utterance unit index value and word unit index value, a series of words that the system has recognized the dialogue information read from the dialogue information recording means, The utterance unit obtained by comparing the correct answer (R) of the actually spoken word sequence (P) recorded by the dialog information recording means and the correctness in word units are shown as binary values of 0 and 1. The reliability measure for evaluating the reliability of the utterance unit and the word unit is created by obtaining the relationship with the index, and the reliability of the speech recognition result whose reliability is to be evaluated by the created reliability measure. Audio interpretation, characterized in that the target value is calculated, a reliability evaluation value obtained by applying the evaluation formula confidence measures. Speech input means for inputting speech information to the system; speech recognition means for recognizing the speech information as a sequence of words; speech recognition reliability evaluation means for evaluating the recognition reliability of each recognized word; A speech interpretation system comprising a sequence of recognized words and speech interpretation means for interpreting speech input based on the reliability of recognition of each word,
The voice (A) input from the voice input means is recognized as a word series (B) by the voice recognition means, and the words (B0), (B1),..., (Bn) included in the word series are recognized. Voice recognition reliability evaluation values (C0), (C1),..., (Cn) are calculated for the respective voice recognition reliability evaluation means, and the word (B0),. A speech interpretation apparatus that identifies a speech interpretation result (D) for an input speech (A) from evaluation values (C0), (C1),..., (Cn). The speech interpretation apparatus according to claim 5,
When a series of words including registered specific word series (E0), (E1),..., (Ei) is input, a series of interpretation results (F0), (F1),. ), A specific type word extracting unit for outputting the specific type word (G) when a series of words including a specific type of word (G) is input, and a database There are pairs of word sequences (H0), (H1),..., (Hk) accumulated as examples and corresponding interpretation result sequences (I0), (I1),. Based on the case type interpretation means for outputting the interpretation result series (I0),..., (Il) when a series of words including the series (H0),. , A finite state transaction that combines specific word extraction means and case type interpretation means The interpretation result sorting means utilizing inducer, speech interpreter to identify the best speech interpretation results (D). The speech interpretation apparatus according to claim 6,
As examples stored in the database, the speech recognition reliability evaluation value (Mi) includes case registration means for registering a set of word sequences (Ji), (Li) larger than a certain value, and this registered word sequence ( A speech interpretation apparatus in which Ji) and (Li) are used as examples by the case type interpretation means. 8. The speech interpretation apparatus according to claim 7, wherein the speech recognition reliability evaluation value (Mi) used in the case type interpretation means is a sequence of words estimated to be actually uttered from dialogue information recorded in the dialogue information recording means (P ) To extract word series,
Information content specifying means for specifying (P) ′ as the content of the information interpreted from the word sequence (P) extracted by the word sequence extracting means;
Word extraction means for extracting the words included in the word series (P) as (Q0), (Q1), (Q2),..., (Qn);
Correct word specifying means for specifying, as (R), the correct answer of the series of words actually spoken corresponding to the word series (P);
Correct information specifying means for specifying (R) ′ as the correct answer of the information content interpreted from the correct answer (R) of the series of words specified by the correct word specifying means;
An utterance word specifying means for specifying the actually spoken word corresponding to the word string (Qi) as (Si);
In addition to the index calculated to estimate the extracted word sequence (P) as a candidate for the result, the index obtained from the dialog information recorded in the dialog information recording means is expressed in utterance units and word units. An utterance unit index value calculating means and a word unit index value calculating means for calculating an utterance unit index value and a word unit index value for each;
The calculated utterance unit index value and word unit index value, a series of words in which the system recognizes the dialog information read from the dialog information recording means, and the actual utterance recorded by the dialog information recording means The reliability of the utterance unit and the word unit is obtained by obtaining the relationship between the utterance unit obtained by comparing the correct word sequence and the index indicating the correctness in the word unit by binary values of 0 and 1. An utterance unit reliability scale creation means and a word unit reliability scale creation means for creating a scale for evaluation;
An utterance unit reliability evaluation value calculation means and a word for calculating an index value of a speech recognition result whose reliability should be evaluated by the created reliability measure, and obtaining the reliability evaluation value by applying the evaluation value of the reliability scale A speech interpretation apparatus characterized by being calculated by unit reliability evaluation value calculation means.   5. A speech interpretation program, characterized in that the computer is written in a readable program language and causes the computer to execute the speech interpretation method according to any one of claims 1 to 4.

Images