JPH06161488A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To greatly reduce a period of time and a throughput required for spotting and efficiently and reliably perform speech recognition. CONSTITUTION:This speech recognizing device has a speech input part 1 which inputs a vocalized speech and a spotting part 6 which extracts >=1 key words from the speech inputted from the speech input part 1. Further, key words to be extracted are classified in plural groups and held in a key word holding part 4 and recognition order on the respective groups are determined as recognition order rules and held in a recognition order rule holding part 5. The spotting part 6 spots the input speech in the order based upon the recognition order rules by regarding the groups as vocabularies to be recognized to extract the key words from the input speech in order. Consequently, the key words can be extracted from the input speech efficiently and reliably at the minimum spotting frequency.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for recognizing spoken voice by word spotting.

[0002]

2. Description of the Related Art A method of recognizing a voice by word spotting, instead of recognizing the entire voice section of a uttered voice without a gap (one phrase per phrase), avoids problems such as addition of non-words and pauses. It is known that it is suitable for spoken dialogue systems and speech understanding systems.

A technique for recognizing speech by this type of word spotting is shown in a document "Acoustic Society of Japan Lecture Collection (March 1992, p139-140)". The technique of this document is extracted. Perform word spotting (which words have a score from where (starting) to where (ending)) in the speech section by spotting all the target keywords, and then combine the keywords. It parses syntactically to recognize the set of keywords in the spoken voice.

[0004]

However, in the above-mentioned conventional speech recognition method, word spotting is executed for all keywords to be extracted without any linguistic restrictions such as syntax and meaning. Therefore, there is a drawback that a lot of calculation time is required for spotting, a large amount of processing is required, and a reliable recognition result cannot be obtained.

It is an object of the present invention to provide a voice recognition device which can significantly reduce the time and processing amount required for spotting and can perform voice recognition efficiently and reliably.

[0006]

In order to achieve the above-mentioned object, the invention according to claims 1 to 3 is a voice input means for inputting a uttered voice, and an input from the voice input means. In a voice recognition device having a spotting means for extracting at least one or more keywords from the speech, the keywords to be extracted are classified into a plurality of groups, and the recognition order of each group is defined as a recognition order rule. It is characterized in that the spotting means performs spotting on the input voice as a group to be recognized as a recognition target vocabulary in an order according to the recognition order rule, and sequentially extracts keywords from the input voice. As a result, the keyword can be efficiently and reliably extracted from the input voice with the minimum number of spotting times.

According to the fourth aspect of the invention, when extracting a keyword from the input speech, the spotting means should spot a speech section excluding a speech section in which the keyword already spotted is present. The feature is that the range is a voice section.
As a result, the spotting area can be narrowed, the keywords can be extracted more efficiently and reliably, and the recognition efficiency can be further improved.

Further, in the invention described in claims 5 to 7, when spotting is performed according to the recognition order rule with the group as a recognition target vocabulary, the score of the extracted keyword is equal to or less than a predetermined threshold value. In this case, a message output means for outputting a predetermined message to the user is further provided.
As a result, a dialogue can be promoted with the user and correct recognition can be performed.

The invention according to claim 8 is characterized in that at least one of the plurality of groups includes a keyword indicating that the previous recognition result is erroneous. Thereby, the recognition efficiency can be further improved.

Further, the invention according to claim 9 is characterized in that at least one group of the plurality of groups includes a keyword indicating that the dialogue is interrupted and the processing is started again. As a result, when the user utters a keyword for interrupting the dialogue, for example, "escape", spotting of a group of functional words is performed first, and the recognition device is immediately detected when "escape" is extracted. Since the state of 1 is returned to the initial state, it is possible to perform efficient processing without performing unnecessary spotting.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an embodiment of a voice recognition device according to the present invention. Referring to FIG. 1, the voice recognition device includes a voice input unit 1 such as a microphone or a handset for inputting a voice, a feature extraction unit 2 for converting the input voice into a time series of feature vectors, and a feature extraction result. A basic unit recognizing unit 3 for recognizing a basic unit of a voice such as a phoneme, a syllable, a VCV (vowel-consonant-vowel), and a keyword holding for storing keywords for performing word spotting classified into a plurality of groups. The group 4, the recognition order rule holding unit 5 holding the recognition order rules for a plurality of groups, and the recognition order rule held in the recognition order rule holding unit 5 are used to select a group, and the keywords belonging to this group are selected. The spotting unit 6 that extracts from the keyword holding unit 4, the message holding unit 7 that holds a message, and the spotting 6 in accordance with the extraction result of taking out a predetermined message from the message storage unit 7, and a message output unit 8 for outputting.

FIG. 2 is a diagram showing a concrete example of the configuration of the keyword holding unit 4. In this example, the keyword holding unit 4
Is configured to group keywords having the same meaning into one group and hold the keyword for each group. More specifically, the keywords representing the subject, the person, the place, the date, and the time are each grouped. In addition, the keyword holding unit 4 includes keywords "escape", "mistake", and "mistake" that represent a function word that is a set of words that indicate that the dialogue should be redone and that the previous recognition result was incorrect. I am aware of "is also held as a group. In addition,
“Escape” is a keyword indicating that the dialogue is interrupted and redone, and “mismatch” and “misrecognized” are keywords indicating that the previous recognition result is an error.

FIG. 3 shows an example of holding keywords that belong to the "time" group. In this example, for times such as "1:30", "2:30", "3:30", etc. Syllabic arc as a basic unit of speech
Described and maintained by one automaton.

FIG. 4 is a diagram showing a specific example of the configuration of the recognition order rule holding unit 5. In this example, the recognition order rule holding unit 5 describes the recognition order rule for each keyword. That is, if the voice to be recognized is the first (that is, the first) utterance, the "requirement" is spotted first, and if the utterance other than the first utterance (for example, the second utterance), " The recognition order rule for spotting "function words" is held, and the recognition order rule for which group should be spotted from the next time onward is held for each keyword (meeting, annual leave) indicating "message". Has been done. For example, the keyword "meeting" holds order rules such as "place" followed by "date" and "date" followed by "time". The spotting unit 6 identifies a group according to such an order rule, performs spotting with the group as a recognition target vocabulary (for example, performs spotting with an automaton in the group as a recognition target), and extracts a predetermined keyword. Has become. In addition, at this time, the spotting unit 6 sets the voice section excluding the section in which the already spotted keyword is present as the range of the voice section to be spotted. Here, as can be seen from the above description, the “order” does not mean the temporal position of the keyword appearing in the input voice, but the order in which the voice recognition device attempts spotting on the input voice. That is.

FIG. 5 is a diagram showing a concrete example of the configuration of the message holding unit 7. In this example, a message is held for each group. This message prompts the user to utter a keyword in the above-mentioned recognition target group. For example, in the "message" group,
There is a message saying "Please have a request." When the message output unit 8 performs spotting according to the ordering rule in the spotting unit 6 with the group as a recognition target vocabulary, the score of the extracted keyword is less than or equal to a predetermined threshold value, or the keyword cannot be extracted. In this case, the message corresponding to that group is retrieved from the message holding unit 7,
It is designed to be output to the user.

Next, the processing operation of the speech recognition apparatus having such a configuration will be described. When the voice of the speaker is input from the voice input unit 1, the feature extraction unit 2 converts this voice into a time series of feature vectors. More specifically, the feature extraction unit 2
Is converted into a feature vector for speech recognition such as a spectrum by a bandpass filter group or an LPC cepstrum.

Next, the basic unit recognition section 3 recognizes the basic unit of voice. For example, a syllable is recognized as a basic unit of voice. In this case, the recognized result (score) is called syllable lattice, and cv_rst (cv, i
s, dur). Where cv
Is a syllable number, is is a syllable start frame number, and dur is a syllable continuous frame number. This syllable lattice may be calculated before the spotting unit 6 described later, or may be calculated sequentially when the spotting unit 6 needs it. Further, as a concrete calculation method of the syllable lattice, there is a method of calculating a standard pattern for each syllable by DP matching, or an HMM for each syllable.
(Hidden Markov Model) can be prepared and recognized.

The spotting unit 6 connects the syllable lattices by automaton control to find an optimum one path on the automaton, and uses this as the extracted keyword.
A method of controlling using an automaton is conventionally known. For example, the cumulative score S (i, j) is calculated using the recurrence formula shown below, and the largest score is calculated in the terminal node of the automaton. It is possible to set a matching path having “” as a recognition result (spotting result).

[0019]

[Equation 1]

Where i is the frame number of the input voice, and j
Is the node number of the automaton, and S (i, j) is the cumulative score of the matching path at which the i-th frame of the input speech reaches the j-th node of the automaton. Also, j'is the j-th
The node number of the parent of the node, cv ′ is an arc (syllable) connecting the j′th node and the jth node, and rst (c
v ', i-dur, dur) is the recognition result of the above syllable.

By the way, in the present embodiment, the keywords to be extracted are classified into a plurality of groups, and the keywords are spotted according to the recognition order rule that defines the order in which the groups should be recognized. The keyword can be extracted from the voice by the number of times, and the meaning of the uttered voice can be efficiently understood. Therefore, when this is applied to a schedule management device, for example, efficient schedule management can be performed.

As a first example, when the input voice is as shown in FIG. 6 (a) and this is the first utterance (first utterance), the spotting unit 6 determines the recognition order. According to the recognition order rule held in the rule holding unit 5 as shown in FIG.
Using the correspondence table between groups and keywords held in the keyword holding unit 4 as shown in FIG. 2, a predetermined keyword is extracted from the “message” group, that is, spotting is performed. That is, it is assumed that, in the voice section (area A) of FIG. 6A, the group “work matter” is recognized as a recognition target, and as a result, the keyword “meeting” is extracted in the section of area B. In this case,
Subsequent recognition order is based on the keyword “meeting” in the recognition order rule of FIG. 4, and is “place”, “date”, “time”.
It is determined in the order of groups. That is, first,
[A- of FIG.
Spotting is performed in the area [B]. As a result, when the keyword “first conference room” is extracted in the section of the area C,
Second, spotting is performed in the area [A- (B + C)] in FIG. 6 with the "date" group as the recognition target. As a result, when the keyword "after tomorrow" is extracted in the section of the area D, thirdly, spotting is performed in the area [A- (B + C + D)] of FIG. 6 with the "time" group as the recognition target. In such a manner, spotting can be sequentially performed according to the recognition order rule.

As a second example, the input voice is shown in FIG.
In the case of something like (b), if it is possible to recognize that the input speech is "annual leave", the next recognition should be "according to the recognition order rule of FIG. Only the "Date" group is used, and spotting is not attempted with the "Place" or "Time" groups as recognition targets. Since the spotting area can be narrowed in this way, it is possible to extract keywords from the voice with the minimum number of spotting times without unnecessary spotting. Also, by narrowing the spotting area and limiting the recognition target, highly reliable results can be obtained efficiently.

If the recognition result obtained in this way is not what the user intended and is different, or is erroneously recognized, the user "does not ...",
Next, say a functional word such as "I am erroneously recognizing ...". For example, in the above-mentioned first example, although the user uttered “first conference room” in the first utterance,
If this is mistakenly recognized as the "second meeting room," the user then says, "No, it is the first meeting room." ”.

Since this input voice is not the first utterance, but the next uttered voice, the spotting section 6 attempts to spot the "function word" according to the recognition order rule of FIG. In other words, we try to spot the words to be re-dialogue and the words showing that the previous recognition result is wrong. Here, when the keyword of “not correct” or “erroneously recognized” can be extracted, the previous recognition target is recognized according to the recognition order rule of FIG. That is, in the first example, the keyword groups “message”, “location”, and “date” of the previous recognition result are displayed.
Try spotting. As a result, the keyword "first conference room" can be extracted in the "place" group next to the "message" group. Even in this case, the target keyword can be extracted only in the "place" group without targeting the "date" and "time" groups, thus reducing the number of spotting and reliable recognition. The result can be obtained.

That is, in the past, considering that the recognition result is incorrect and the user corrects it, the object to be recognized this time (ie, “place”) and the previously recognized object (ie, “message”, “message”). All of "Date", "Time") must be recognized, that is, extracted, but according to the present invention, a keyword (mismatch, misrecognize, ...) indicating a recognition error is first detected. Only if it can be extracted, only the target that was not recognized last time is recognized.
Therefore, it is possible to efficiently recognize an input voice such as “No, it is the first conference room” without targeting the “time” group.

Further, in the present embodiment, as a result of performing spotting according to the recognition order rule of FIG. 4, when spotting of a certain group fails (when it cannot be recognized and when it is not actually recognized in the input voice, Including),
A message prompting the user to utter this group is output with a simple structure (without considering the history of the dialogue). Specifically, in the above-mentioned first example, "meeting" was extracted as the "requirement", "first meeting room" was extracted as the "place", and "the day after tomorrow" was extracted as the "date". However, it is assumed that the score of “time” = “2:00” is less than or equal to a predetermined threshold. At this time, since the message output unit 8 could not recognize the "time" group, the message holding unit 7 outputs the message "What time is it?" In the "Time" column, and the message is displayed to the user. In the next utterance, encourage them to say the word "time."

As a result, when the user utters "time", the input voice is not the first utterance, but the next uttered voice. Attempt to spot function words according to the ordering rules. For example, suppose the user utters "Well, it's 2 o'clock." However, this input voice does not include the “function word”, and therefore the spotting unit 6 cannot extract the “function word”. When the "function word" cannot be extracted in this way, the spotting unit 6 attempts spotting with a group to which a keyword having a score of a predetermined threshold value or less in the previous recognition belongs as a recognition target. That is, in the above example, spotting is attempted with the "time" group as the recognition target. As a result, “2 o'clock” can be extracted as a keyword. Also in this case, the “requirement”, which is the previous recognition target,
Since the group of "place" and "date" is not the target of recognition, efficient and reliable recognition can be performed.

In this way, normal interaction can be continued with the device to efficiently obtain the correct results.

When the user wants to interrupt the dialogue, the user may say "escape". When "escape" is uttered, a group of functional words is first spotted, and the state of the recognizer is immediately returned to the initial state when "escape" is extracted. This can prevent useless spotting.

[0031]

As described above, according to the inventions of claims 1 to 3, the keywords are classified into a plurality of groups, and the keywords are input in accordance with a recognition order rule that defines the order in which each group should be recognized. Since the keyword is extracted by performing spotting on the voice, the keyword can be efficiently and reliably extracted from the input voice with the minimum number of spotting times.

According to the fourth aspect of the invention, the spotting of the group to be recognized next is performed excluding the keyword section already extracted within the section of the input voice. It is possible to narrow the region in which the keyword is performed, so that the keyword can be extracted more efficiently and reliably, and the recognition efficiency can be further improved.

According to the fifth to seventh aspects of the invention, spotting is performed in order according to the recognition order rule, and a message is output to the user about the spotting that was not recognized but could be recognized. Therefore, it is possible to carry out a dialogue with the user and make a correct recognition.

Further, according to the invention of claim 8, only the group to which the keyword having an incorrect recognition result belongs is targeted for recognition, so that the recognition efficiency can be further improved.

According to the ninth aspect of the invention, when the user utters a keyword for interrupting the dialogue, for example, "escape", first, spotting of a group of function words is performed, and " Since the state of the recognition device is immediately returned to the beginning when "escape" is extracted, it is possible to perform efficient processing without performing unnecessary spotting.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of a voice recognition device according to the present invention.

FIG. 2 is a diagram showing a specific configuration example of a keyword holding unit.

FIG. 3 is a diagram showing an example of holding keywords belonging to a “time” group.

FIG. 4 is a diagram illustrating a specific configuration example of a recognition order rule holding unit.

FIG. 5 is a diagram showing a specific configuration example of a message holding unit.

6A and 6B are diagrams showing an example of an input voice.

[Explanation of symbols]

 1 voice input unit 2 feature extraction unit 3 basic unit recognition unit 4 keyword storage unit 5 recognition order rule storage unit 6 spotting unit 7 message storage unit 8 message output unit

Claims (9)

[Claims] 1. A voice recognition device having voice input means for inputting spoken voice and spotting means for extracting at least one or more keywords from the voice input from the voice input means, as an extraction target. Keywords are classified into a plurality of groups, and the recognition order of each group is determined as a recognition order rule, and the spotting means sets the groups as recognition target vocabulary in the input voice in an order according to the recognition order rule. A voice recognition device characterized by performing spotting and sequentially extracting keywords from an input voice. 2. The voice recognition device according to claim 1, wherein
It has a voice basic unit recognition means for recognizing a basic unit of voice, and in this case, the keywords in each group are described for each group by an automaton whose basic unit of voice is an arc, and the spotting means. When spotting a group of input speeches in the order according to the recognition order rule as a recognition target vocabulary and extracting a keyword from the input speech, an optimum path on the automaton in one group is obtained, and the group is extracted. A voice recognition device characterized by being adapted to perform extraction of a keyword inside. 3. The voice recognition device according to claim 1, wherein
The voice recognition device, wherein the recognition order rule is described for each keyword. 4. The voice recognition device according to claim 1,
The speech recognition, wherein the spotting means, when extracting a keyword from the input speech, sets a speech section excluding a speech section in which a keyword already spotted is present as a range of a speech section to be spotted. apparatus. 5. The voice recognition device according to claim 1,
When performing the spotting with the group as the recognition target vocabulary according to the recognition order rule, when the score of the extracted keyword is equal to or less than a predetermined threshold value,
A voice recognition device, further comprising message output means for outputting a predetermined message to a user. 6. The voice recognition device according to claim 5,
The voice recognition device, wherein the message output to the user is predetermined and held for each group. 7. The voice recognition device according to claim 5, wherein
The voice recognition device, wherein the message output to the user is a message prompting the user to re-speak a keyword in a group to which the keyword whose score is equal to or lower than a predetermined threshold value. 8. The voice recognition device according to claim 1,
At least one group of the plurality of groups includes
A voice recognition device characterized in that a keyword indicating that the previous recognition result is incorrect is included. 9. The voice recognition device according to claim 1, wherein
At least one group of the plurality of groups includes
A voice recognition device characterized in that it includes a keyword indicating that the dialogue should be interrupted and started over.