JP2006184421A - Speech recognition device and speech recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition device which completes speech recognition processing within a short time after a speech for recognition is inputted and identifies a speech-inputted vocabulary. <P>SOLUTION: The speech recognition device comprises: a speech input means 102 for converting and outputting an inputted speech into a speech signal; a beginning/ending speech recognition means 103 capable of recognizing each one syllable of the beginning and ending of a word in the speech signal outputted from the speech input means 102, a vocabulary narrowing-down means 111 for narrowing down a vocabulary range for recognition using either one syllable of the beginning or that of the ending or both of them recognized by the beginning/ending speech recognition means 103; and a vocabulary speech recognition means 104 for recognizing the vocabulary in the above speech signal using the vocabularies for recognition in the range narrowed down by the above vocabulary narrowing-down means 111. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a voice recognition device and a voice recognition method.

  Recently, many voice interface methods for recognizing operator's utterance information and inputting information have been proposed. This is attracting attention as a method that can reduce the trouble of manually operating the switch, and is particularly effective for information search operations (name search, telephone number search, etc.).

  In order to realize a voice interface, generally, a recognizable vocabulary is registered in the system as a dictionary, and the spoken information is compared with the contents of the dictionary to identify one word with the highest likelihood. Do. At this time, since collation processing with all prepared vocabularies is performed, if the capacity of the dictionary increases, there arises a problem that the time required for collation increases. This is because the recognition dictionary is held in an external storage device such as a CD-ROM or DVD, and it takes a long time to read the dictionary. Of course, as the number of vocabularies increases, the time required for collation itself increases. In general, in the case of a dictionary having a level of several hundred words, the collation process is completed within one second. However, when the dictionary size exceeds 10,000 words, a processing time on the order of several seconds may be required in some cases. In such a case, since it takes time until the recognition result is output after the utterance is completed, there is a possibility that the interface tempo is hindered and the interface becomes unnatural.

JP 2001-039881 A

  In the prior art, the above-described collation processing for speech recognition may require a processing time on the order of a few seconds, and as an invention corresponding to the problem of an unnatural interface, for example, in Patent Document 1 below, As described, a speech recognition system has been proposed in which frequently used vocabulary is stored in a place where high-speed access is possible, and vocabulary with low usage frequency is stored in a place where access speed is low.

  However, according to such a method, it is necessary to analyze in advance the vocabulary usage frequency. In general, when performing an information search, it is difficult to predict the frequency of information to be searched, and in general, some constraints must be imposed. For example, when you are in Kanagawa Prefecture, there is a high possibility that the information of Kanagawa Prefecture will be checked, so that information is stored in a place that can be accessed at a high distance. If such a restriction is applied, there is a problem that if the operator's intention is not satisfied, the usability is deteriorated and the usability as an interface is deteriorated. It is desirable that information retrieval of a large amount of vocabulary can be performed as flat as possible.

  The present invention has been made paying attention to such problems, and the problem to be solved by the present invention is to finish the speech recognition processing within a short time after the speech for recognition is input. Another object of the present invention is to provide a speech recognition apparatus and speech recognition method for specifying a vocabulary inputted by speech.

  When a vocabulary is input as speech, it converts the input speech into a speech signal, recognizes only one or both of the first syllable and the last syllable in the speech signal, and recognizes using the recognition result A speech recognition device that narrows down the range of the target vocabulary and recognizes the vocabulary in the speech signal, that is, the input vocabulary, using the recognition target vocabulary within the narrowed down range is configured. Here, the “recognition target vocabulary” means a vocabulary that may match the vocabulary input by voice.

  By implementing the present invention, it is possible to provide a speech recognition device and a speech recognition method for identifying a vocabulary input by speech after finishing speech recognition processing within a short period of time after speech for recognition is input. Become.

  The features of the speech recognition apparatus and speech recognition method according to the present invention are to recognize one syllable in one or both of the beginning and the end of a spoken word, narrow down the recognition target vocabulary based on the result, and then to the entire utterance. A collation process is performed to specify the vocabulary.

  Hereinafter, the present invention will be described in detail by way of embodiments.

(First embodiment)
In this embodiment, the present invention is applied to facility search of a vehicle navigation system. Hereinafter, the contents of the present invention will be described with reference to the drawings.

  The present embodiment is a voice recognition device that is installed on a steering wheel and has the following components as shown in FIG. That is, the present embodiment includes a voice input start instructing unit 101 that instructs to start voice input, a voice input unit 102 that includes a microphone as a component, converts input voice into a voice signal, and outputs, and voice input. An initial and ending speech recognition means 103 capable of recognizing one syllable at the beginning and the ending syllable in the speech signal converted by the means 102; and a word speech recognition means 104 for recognizing the vocabulary in the speech signal. A speech recognition dictionary 105 in which the recognition target vocabulary is expanded, a speech recognition information database 106 in which the recognition target vocabulary is stored, response generation means 107 for generating a response output by voice or screen display, and text as a response output Visual information display means 108 for outputting information and map information to a monitor, and auditory information provision for outputting a generated response sentence as a response output. Recognition using means 109, navigation system control means 110 for controlling the navigation system according to the input content, and one or both of the first syllable and the last syllable recognized by the beginning / ending speech recognition means 103 Vocabulary narrowing means 111 for narrowing down the range of the target vocabulary.

  The present embodiment is a voice input device that can operate a navigation system installed in a vehicle by voice. Here, the operation of the voice input device according to the present invention will be described by taking as an example the case of performing facility name search by voice.

Hereinafter, the flow of action will be described in association with the flowchart shown in FIG.
(A) The voice input start instruction unit 101 is installed on the steering wheel of the vehicle. This is used for a voice input start instruction.
(B) As the voice input unit 102, a microphone is installed in the vicinity of the driver's mouth, such as in the vicinity of a vehicle rearview mirror or a steering column.
(C) When the voice input start button of the voice input start instructing unit 101 is pressed, the voice input unit 102 is activated and transitions to a speech input enabled state (S201). At this time, the guidance information for urging the utterance is output from the auditory information presentation unit 109, such as “Please give me a request” (S202). At the same time, a display prompting voice input is output on the monitor screen through the visual information display means 108. At this time, the recognition target vocabulary is transferred and set from the speech recognition information database 106 to the speech recognition dictionary 105.
(D) A beep sound is output after the guidance message “Please give me a request”, and when the beep sound ends, a voice input standby state is entered. At this time, an icon indicating that voice input is possible is displayed on the screen. Here, the operator utters the command “facility search” (S203).
(E) The user's utterance “facility search” is input to the voice input means 102, converted into a voice signal by the voice input means 102, and the voice signal is converted to the voice recognition dictionary 105 by the word voice recognition means 104. The collation is performed and the command “facility search” is recognized (S204).
(F) A guidance voice “Please name the facility” is output from the auditory information presentation unit 109 (S205). At the same time, through the visual information display means 108, a display prompting the facility name input is output on the monitor screen. At this time, one character data of 50 sounds is set in the speech recognition dictionary 105.
(G) A beep sound is output after the guidance sound “Please name the facility”, and when the beep sound ends, a voice input standby state is entered. At this time, an icon indicating that voice input is possible is displayed on the screen. Here, the operator speaks the facility name “Tokyo Disneyland” (S206). This utterance information (including the utterance of the operator converted into an audio signal by the audio input means 102) is temporarily stored.
(H) The beginning / ending speech recognition means 103 recognizes one syllable of each of the beginning and end of the spoken word. In this example, the beginning of the word is recognized as “to” and the end of the word is recognized as “do” (S207).
(I) Information whose facility name information stored in the speech recognition information database 106 is “to” and whose ending is “do” is selected (S207). The vocabulary range is narrowed down under the narrowing condition that the ending is “DO”, and the facility name stored in the speech recognition information database 106 within the narrowed range is the recognition target vocabulary. Transferred to and set in the speech recognition dictionary 105. As a result, the vocabulary to be recognized is narrowed down (S208).
(J) In the word speech recognition means 104, recognition processing of the information previously spoken in (g) is executed, and the vocabulary “Tokyo Disneyland” in the input speech signal is recognized (S209). At this time, the utterance information (input speech signal) temporarily stored is transferred to the speech recognition dictionary 105, and the vocabulary with the word beginning “to” and the word ending “do” (recognition after narrowing down) is recognized. This is done by collating with the target vocabulary. In this way, the input speech signal that has already been recognized by the beginning / ending speech recognition means 103 is recognized again.
(K) The guidance information “Search for Tokyo Disneyland” is output from the auditory information presentation means 109 (S210), and the map of the searched point is displayed on the monitor screen through the visual information display means 108. The
(L) Thereafter, according to the instructions on the screen, operations such as “register the displayed point” and “set it as the destination” can be performed. This is performed by the navigation system control means 110. At this time, a map of Tokyo Disneyland is displayed (S211), and the above operation is executed (S212).

  In the above, one syllable at the beginning of the word and one syllable at the end of the word are recognized, and both are used to narrow down the vocabulary. You may narrow down the vocabulary. In that case, the burden of the one syllable recognition process can be further reduced. However, there is a possibility that the number of vocabulary to be recognized after that may increase, so that the total number of recognition vocabularies is taken into account and both the beginning and ending are recognized and used for narrowing down the vocabulary only. Then, it may be determined whether to use it for narrowing the vocabulary or to recognize only the ending and use it for narrowing the vocabulary. That is, the number of words to be recognized within the range narrowed down by the vocabulary narrowing means 111 using the result of the recognition of only one of the initial syllable and the ending syllable of the speech signal in the speech signal. However, when the number exceeds the predetermined number, the initial / ending speech recognition means 103 recognizes the other one of the initial syllable and the final syllable in the speech signal, and also uses the result of the recognition. Thus, the vocabulary narrowing means 111 may narrow down the range of the recognition target vocabulary.

  In this embodiment, the syllable at the beginning or the end of the word is recognized when the facility name is recognized. However, this may be applied at the time of command utterance of “facility search”. In that case, if the number of commands to be recognized is small, the command may be narrowed down to just one by recognizing the sound at the beginning or end of the word. In such a case, the recognition process can be speeded up. is there. In addition, the facility name may be input by an operator using an input method other than voice (for example, selection by key operation).

  The facility name may be input after the vocabulary is narrowed down and before the speech recognition by the word speech recognition means 104.

  In the above, the recognition target vocabulary range, which is the narrowing down of the recognition target vocabulary range by category classification, is used as the facility name before the beginning / ending speech recognition means 103 performs syllable recognition (before (h)). The limited range is narrowed down ((c) to (e)), but the range of the recognition target vocabulary by such category classification is narrowed down after the syllable recognition is performed by the beginning / ending speech recognition means 103 (( After h), it may be performed before the word speech recognition means 104 starts vocabulary recognition (before (j)).

  As described above, in the present embodiment, only one or both of the first syllable and the last syllable in the spoken vocabulary are recognized, and the candidate vocabulary is narrowed down based on the result. The recognition process is performed again, and the result is specified.

  According to this method, even if there are a large number of recognition target words, it is sufficient to perform collation with the candidates after narrowing down, so that recognition can be performed in a short time. That is, it is possible to specify the spoken vocabulary by completing the collation process within a short time after the input of recognition speech.

  In addition, when reading a vocabulary from an external storage device such as a disk, only the narrowed vocabulary needs to be read, so that the reading time can be shortened.

  In addition, since the above configuration is used, an effect that a memory capacity for expanding the recognition target vocabulary is small can be obtained.

  Also, recognition of one syllable at the beginning / end of the word is easier than recognition of one syllable other than the beginning / end of one word, thereby enabling highly accurate speech recognition.

(Second embodiment)
In this embodiment, the present invention is applied to facility search of a vehicle navigation system. Hereinafter, the contents of the present invention will be described with reference to the drawings.

  The present embodiment is a voice recognition device that is installed in a steering and has the following components as shown in FIG. That is, the speech recognition apparatus includes a speech input start instructing unit 301 that instructs to start speech input, a speech input unit 302 that converts a speech input into a speech signal, and includes a microphone or the like, and a speech input unit. Initial and ending speech recognition means 303 capable of recognizing one syllable at the beginning and one ending in the speech signal converted by 302, word speech recognition means 304 for recognizing the vocabulary in the speech signal, A speech recognition dictionary 305 in which the recognition target vocabulary is expanded, a recognition probability analyzing means 306 for calculating a probability of the recognition result of the beginning / end of one syllable, a speech recognition information database 307 storing the recognition target vocabulary, , Response generation means 308 for generating a response output by voice or screen display, and outputting text information and map information to the monitor as a response output Visual information display means 309; auditory information presentation means 310 that outputs a generated response sentence as a response output; navigation system control means 311 that controls the navigation system in accordance with the input content; A vocabulary narrowing means 312 for narrowing down the range of the vocabulary to be recognized using one or both of one syllable at the beginning and one syllable at the end of the word recognized by 303.

An example of the likelihood of the recognition result of the first ending / ending syllable calculated by the recognition probability analyzing unit 306 is a recognition likelihood. The recognition likelihood is defined as, for example, the degree of pattern matching between the input speech waveform and the collation syllable waveform, that is, the absolute value of the correlation coefficient between the two waveforms. In this case, the correlation coefficient between the two waveforms is f (t) and g (t) (t is time) for each waveform, and S (x) is x in the time interval where both waveforms exist. Is a value obtained by integrating with respect to time t,

S (f (t) · g (t)) / [(S (f (t) ² ) · S (g (t) ² )] ^1/2

Is defined as This recognition likelihood is 1 when both waveforms are the same (when k is a constant and expressed as f (t) = k · g (t)), and the recognition likelihood becomes smaller as the similarity between both waveforms is lower. Become.

The present embodiment is a voice input device that can operate a navigation system installed in a vehicle by voice. Here, the operation of the present invention will be described by taking the case of performing facility name search by voice as an example. Hereinafter, the flow of action will be described in association with the flowchart of FIG.
(A) The voice input start means 301 is installed on the steering wheel of the vehicle. This is used to start voice input or instruct a cancel operation.
(B) As the voice input means 302, a microphone is installed in the vicinity of the driver's mouth, such as in the vicinity of a vehicle rearview mirror or a steering column.
(C) When the voice input start button of the voice input start unit 301 is pressed, the voice input unit 302 is activated and transitions to a speech input enabled state (S401). At this time, the guidance information prompting the utterance is output from the auditory information presenting means 310, such as “Please give me a request” (S402). At the same time, through the visual information display means (309), a display prompting voice input is output on the monitor screen. At this time, the recognition target vocabulary is transferred and set from the speech recognition information database 307 to the speech recognition dictionary 305.
(D) A beep sound is output after the guidance message “Please give me a request”, and when the beep sound ends, a voice input standby state is entered. At this time, an icon indicating that voice input is possible is displayed on the screen. Here, the operator utters a command “facility search” (S403).
(E) The user's utterance is input to the word speech recognition unit 304 as a voice signal through the voice input unit 302, checked by the word speech recognition unit 304 against the speech recognition dictionary 305, and a command “Facilities Search” is issued. Recognized (S404).
(F) After the guidance sound “Please name the facility” is output, a beep sound is output, and when the beep sound ends, a voice input standby state is entered (S405). At this time, an icon indicating that voice input is possible is displayed on the screen. Here, the operator speaks the facility name “Tokyo Disneyland” (S406). This utterance information is temporarily stored.
(G) The beginning / end speech recognition unit 303 recognizes one syllable of each of the beginning and end of a sentence spoken and converted into a speech signal by the speech input unit 302. In this example, the beginning of the word is recognized as “to” and the end of the word is recognized as “do” (S407).
(H) At this time, in the recognition probability analysis means (306), first, the probability of the recognition result of each one syllable of the beginning and ending is calculated. In this example, it is assumed that the recognition probability of one syllable at the beginning of a word is calculated as A and the recognition probability of one syllable at the end of a word is calculated as B. Here, the certainty is a scale indicating the closeness of words prepared as a dictionary and input words when collation processing for speech recognition is performed. In this example, the recognition probability analysis means (306) assumes that the larger the value, the more likely the recognition result is. Then, in the recognition probability analysis means (306), the thresholds T and A, B of the certainty, which are predetermined criteria, are set. The magnitude relationship is analyzed (S408).

When A ≧ T and B ≧ T (S409)
By the vocabulary narrowing means 111, only the information whose facility name information stored in the speech recognition information database 307 has “to” and ending with “do” is selected, transferred to the speech recognition dictionary 305, and set. The That is, the recognition target vocabulary is narrowed down (S413).

When A ≧ T and B <T (S410)
The vocabulary narrowing means 111 selects only information whose facility name information stored in the speech recognition information database 307 starts with “to”, and transfers and sets it to the speech recognition dictionary 305. That is, the recognition target vocabulary is narrowed down (S413).

When A <T and B ≧ T (S411)
The vocabulary narrowing means 111 selects only information whose facility name information stored in the speech recognition information database 307 has the ending “DO”, and transfers and sets it to the speech recognition dictionary 305. That is, the recognition target vocabulary is narrowed down (S413).

When A <T and B <T (S412)
Encourage recurrence. That is, a response voice output saying "Please speak again" is made, and the voice input standby state of (f) is entered (S405).
(I) In cases other than S412, the word speech recognition unit 304 compares the information previously uttered in (f) with the dictionary set in (h) to recognize “Tokyo Disneyland” (S414). ).
(J) The guidance information “Search for Tokyo Disneyland” is output from the auditory information presentation means 310 (S415), and the map of the searched point is displayed on the monitor screen through the visual information display means 309. (S416).
(K) Thereafter, in accordance with the instructions on the screen, operations such as “register the displayed point” and “set it as the destination” can be performed (S417). This is performed by the navigation system control means 311.

As described above, in the present embodiment, the narrowing range of the recognition target vocabulary is changed based on the probability of the recognition result of the beginning / ending of the word. In particular,
(1) When the accuracy of the recognition result of the beginning / ending of the word is more than a predetermined value, the recognition target vocabulary is narrowed down by the recognition result of both the beginning / ending of the word,
(2) Only when the initial recognition result is more than a certain value, the recognition target vocabulary is narrowed down by the initial recognition result.
(3) When only the ending recognition result is more than a predetermined value, the recognition target vocabulary is narrowed down by the ending recognition result,
(4) When the likelihood of the recognition result of the beginning / ending of the word is less than a predetermined value, re-speech is urged.

  In the above description, the recognition target vocabulary range, which is the narrowing down of the recognition target vocabulary range by category classification, is used as the facility name before the beginning / ending speech recognition means 303 performs syllable recognition (before (g)). Although the limited range is narrowed down ((c) to (e)), the range of the recognition target vocabulary by such category classification is narrowed down after the syllable recognition by the beginning / ending speech recognition means 303 (( After g), it may be performed before the word speech recognition means 304 starts vocabulary recognition (before (i)).

  As described above, in the present embodiment, only one or both of the first syllable and the last syllable in the spoken vocabulary are recognized, and the candidate vocabulary is narrowed down based on the result. The recognition process is performed again, and the result is specified.

  According to this method, even if there are a large number of recognition target words, it is sufficient to perform collation with the candidates after narrowing down, so that recognition can be performed in a short time. That is, it is possible to specify the spoken vocabulary by completing the collation process within a short time after the input of recognition speech.

  Further, when reading a vocabulary from an external storage device such as a disk, it is only necessary to read the narrowed vocabulary, so the reading time can be shortened.

  In addition, since the above configuration is used, an effect that a memory capacity for expanding the recognition target vocabulary is small can be obtained.

  Also, recognition of one syllable at the beginning / end of the word is easier than recognition of one syllable other than the beginning / end of one word, thereby enabling highly accurate speech recognition.

  Furthermore, in the present embodiment, since the recognition target vocabulary can be narrowed down according to the probability of the recognition result of the beginning and ending of the word, it is possible to narrow down the vocabulary according to the utterance state of the operator, resulting in recognition performance. (There is no need to perform a collation process on a vocabulary that may be mistaken by using a recognition vocabulary with a low certainty as a narrowing-down condition). For example, it is possible to eliminate a case where an erroneous narrowing down is performed based on a recognition result of an erroneous head or tail.

  The constituent requirements of the speech recognition apparatus according to claim 1 and the constituent requirements of the speech recognition method according to claim 7 are satisfied in the first embodiment, and the speech recognition apparatus according to claims 4 to 6 is satisfied. The component requirement of the voice recognition method according to the eighth aspect is satisfied in the second embodiment.

It is a block diagram of the example of 1st Embodiment. It is a flowchart which shows the flow of a process of 1st Embodiment. It is a block diagram of 2nd Embodiment. It is a flowchart which shows the flow of a process of 2nd Embodiment.

Explanation of symbols

  101: voice input start instructing means, 102: voice input means, 103: beginning / ending voice recognition means, 104: word voice recognition means, 105: dictionary for voice recognition, 106: information database for voice recognition, 107: response generation means , 108: visual information display means, 109: auditory information presentation means, 110: navigation system control means, 111: vocabulary narrowing means, 301: voice input start instruction means, 302: voice input means, 303: head / end voice recognition Means 304: Word voice recognition means 305: Speech recognition dictionary 306: Recognition probability analysis means 307: Speech recognition information database 308: Response generation means 309: Visual information display means 310: Auditory information presentation Means, 311: Navigation system control means, 312: Vocabulary narrowing means.

Claims (8)

In a speech recognition device that recognizes vocabulary input by speech,
A voice input means for converting the input voice into a voice signal and outputting it; and a head part capable of recognizing at least one of the first syllable and the last syllable in the voice signal outputted by the voice input means; Ending speech recognition means, vocabulary narrowing means for narrowing down the range of recognition target vocabulary using the recognition result in the beginning / ending speech recognition means, and the recognition target vocabulary within the range narrowed down by the vocabulary narrowing means A speech recognition apparatus comprising word speech recognition means for recognizing a vocabulary in the speech signal. The speech recognition device according to claim 1,
The speech recognition apparatus according to claim 1, wherein the beginning / ending speech recognition means recognizes only one or both of one syllable at the beginning and one syllable at the end of the speech signal. The speech recognition device according to claim 2,
The number of words to be recognized within the range narrowed down by the vocabulary narrowing means is determined by using the result of the word / word ending speech recognition means recognizing only one of the first syllable and the last syllable in the speech signal. When the number exceeds a predetermined number, the beginning / ending speech recognition means recognizes the other one of the beginning and ending syllables in the speech signal, and also uses the recognition result to A speech recognition apparatus, characterized in that the vocabulary narrowing means further narrows down the range of the recognition target vocabulary. In a speech recognition device that recognizes vocabulary input as speech,
A voice input means for converting the input voice into a voice signal and outputting it; and a head part capable of recognizing at least one of the first syllable and the last syllable in the voice signal outputted by the voice input means; Ending speech recognition means, recognition probability analysis means for calculating the probability of the recognition result of one syllable at the beginning and the probability of the recognition result of one syllable at the ending in the beginning / ending speech recognition means, and Vocabulary narrowing means for narrowing the range of vocabulary to be recognized using one or both of one syllable at the beginning and one syllable at the end of which the certainty of the recognized recognition result is equal to or greater than a predetermined standard, and the vocabulary A speech recognition apparatus comprising: word speech recognition means for recognizing a vocabulary in the speech signal using a recognition target vocabulary within a range narrowed down by the narrowing-down means. The speech recognition apparatus according to claim 4,
The speech recognition apparatus characterized in that the probability of the recognition result is a recognition likelihood when the recognition result is obtained. The speech recognition device according to any one of claims 1 to 5,
A speech characterized by narrowing down the range of recognition target vocabulary by category classification before or after the beginning / ending speech recognition means performs syllable recognition and before the word speech recognition means starts to recognize words Recognition device. In a speech recognition method for recognizing vocabulary input by speech,
The input speech is converted into a speech signal, one or both of the first syllable and the last syllable in the speech signal are recognized, and the range of the recognition target vocabulary is narrowed down by using the recognition result. A speech recognition method comprising: recognizing a vocabulary in the speech signal using a recognition target vocabulary within a predetermined range. In a speech recognition method for recognizing vocabulary input as speech,
The input speech is converted into a speech signal, and one or both of the first syllable and the last syllable of the speech signal are recognized, the probability of the recognition result is calculated, and the accuracy of the calculated recognition result is confirmed. Is narrowed down the range of the recognition target vocabulary by using one or both of the first syllable and the last syllable of the ending word that is equal to or more than a predetermined standard, and the above using the recognition target vocabulary within the narrowed range A speech recognition method characterized by recognizing a vocabulary in a speech signal.

Images