JP3797003B2 - Audio output device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an audio output device used when interacting with a speaker or when recognizing and outputting the content spoken by the speaker.
[0002]
[Prior art]
The voice output device used when interacting with the speaker recognizes the input voice and outputs a recognized voice for confirmation, or is generated on the device side for questions or answers Output audio. In this case, emphasis processing such as increasing the volume is performed on important parts, etc., or difficult words are input in order to ensure dialogue efficiency and reliable communication between the speaker and the device. In such a case, a method of replacing the voice with a plain expression and outputting the voice for confirmation is adopted. For example, the apparatus disclosed in Japanese Patent Laid-Open No. 10-171485 analyzes text information to identify an important part, and performs a modification such as emphasis to output it. An apparatus disclosed in Japanese Patent Application Laid-Open No. 7-121537 is an apparatus that identifies a difficult word / phrase from a sentence to be read out and replaces it with a plain expression to read out.
[0003]
[Problems to be solved by the invention]
However, with the above devices,
-Since the process according to the past output situation is not performed, when the same output is given, the transformation process is always performed, and as a result, the voice output that the user feels annoying may be obtained.
[0004]
-Since the occurrence of an error in the speech recognition result on the device side with respect to the user's utterance is not considered, it is difficult for the user to detect the recognition error of the device.
[0005]
・ In order to replace homonyms, etc., with plain expressions, it is necessary to prepare explanatory expressions for all the words used.
[0006]
There were problems such as.
[0007]
An object of the present invention is to provide an audio output device that can perform smooth audio output most desirable for a user by dynamically controlling the output audio in accordance with the user's situation.
[0008]
[Means for Solving the Problems]
According to the present invention, the user's situation such as user's confusion, occurrence of an interrupt, and the certainty of recognition of input voice from the user is known, and the output voice is dynamically controlled according to the situation at that time.
[0009]
In the present invention, the following terms are used.
[0010]
Voice recognition: Recognizing each phoneme by breaking the user's input voice into phonemes. Dialogue status: Refers to each stage of dialogue such as “greeting”, “confirmation”, and “answer” in a series of dialogues between the user and the device.
[0011]
・ Expression expressing confusion: Expression of voice or attitude expressed by the user when the user cannot recognize the output of the device.
[0012]
-Deformation of words incorporated in output data: Refers to changing the manner of output of words incorporated in output data output from the apparatus to the user.
[0013]
User interrupted voice: Voice that the user utters in order to give some interrupt during output from the device side.
[0014]
-Input speech recognition certainty: Refers to the certainty of recognition when the user's input speech is recognized.
[0015]
Additional words: Terms that are added immediately before or after a word or character to be incorporated in output data from the apparatus.
The present invention is configured as follows.
[0016]
(1) Voice inputBe doneA voice input unit;
Input voiceofA voice recognition unit for performing voice recognition;
An output data table for storing output data to be output from the device according to the state of interaction with the user;
An output data selection means for determining a conversation status from a voice recognition result and selecting output data to be output at that time from the output data table;
An audio output unit for outputting the selected output data as audio;
Means for detecting voice representing the user's confusion with respect to voice output from the apparatus from the voice recognition result;
Deformation means for deforming a word incorporated in output data when detecting a voice representing the confusion and causing the sound output unit to re-output the word,
WithThe
[0017]
The voice input unit is constituted by a microphone, for example. A speech recognition unit that performs speech recognition from input speech detects phonemes of speech input by the user. For example, when the user speaks “I”, the voice recognition unit recognizes “wa”, “ta”, and “shi” as phonemes. In this speech recognition method, a feature pattern such as a linear prediction coefficient and a Fourier spectrum coefficient is extracted from speech data, and this is matched with a pattern stored in the dictionary for recognition. A known method such as a speech recognition method using a Hidden Markov method or a syntax recognition method is used. In the pattern matching technique, a DP (dynamic programming) pattern matching technique that absorbs differences in utterance speeds depending on users is generally employed. In speech recognition, words and parts of speech can be recognized by performing morphological analysis together with recognition of each phoneme. In the present invention, the speech recognition unit may or may not recognize the word or part of speech.
[0018]
The output data table stores data to be output as a sound from the device according to the state of interaction with the user. By dialogue state means, for example, in a series of dialogues, there is a greeting first, then a voice uttered by the user is confirmed, and further, a question from the user is recognized and answered. By storing in advance in the output data table for each dialog state, it becomes possible to deal with any scene by updating the contents of this table. The output data is combined with fixed data and embedded data (words or the like). Fixed data is voice pattern data fixed in advance on the system, and embedded data is classified into voice pattern data for returning voice input from a user as it is and voice pattern data dynamically generated in the system. . Normally, when the conversation status is “Confirm” type, the embedded data is a voice pattern that returns the input voice from the user as it is. When the conversation status is “Answer” type, the embedded data is dynamically generated. Sound pattern.
[0019]
The output data selection means determines the conversation status from the voice recognition result, and selects output data to be output at that time from the table. The output means outputs the selected output data from a speaker or the like. At this time, built-in data in the output data is generated according to the conversation status.
[0020]
When the user detects a voice representing a confusion, such as “E”, with respect to the output voice output by the output means, the embedded data incorporated in the output data is transformed and output again. Toddler means an expression uttered as a result of the user's failure to recognize the output of the device. In addition to “Eh” and the like, “Nani” and other expressions are a form of misrepresentation. When this confusion is detected, for example, the output is re-output in response to the deformation of “large volume” and “low speed” than the previous time. If no confusion is recognized, the next dialogue is entered.
[0021]
(2) Voice inputBe doneA voice input unit;
Input voiceofA voice recognition unit for performing voice recognition;
An output data table for storing output data to be output from the device according to the state of interaction with the user;
An output data selection means for determining a conversation status from a voice recognition result and selecting output data to be output at that time from the output data table;
An audio output unit for outputting the selected output data as audio;
Means for detecting the user's interrupted voice for the voice output from the apparatus from the voice recognition result;
Deformation means for transforming a word incorporated in output data when the interrupted speech is detected and re-outputting the speech output unit;
WithThe
[0022]
In the present invention, if the interrupt request from the user is a voice, the data (such as a word) incorporated in the output data is transformed and re-output. For example, the interruption sound includes a sound such as “another time” or “more detailed”.
[0023]
(3) Voice inputBe doneA voice input unit;
Input voiceofA voice recognition unit for performing voice recognition;
An output data table for storing output data to be output from the device according to the state of interaction with the user;
An output data selection means for determining a conversation status from a voice recognition result and selecting output data to be output at that time from the output data table;
An audio output unit for outputting the selected output data as audio;
Imaging means for imaging a user;
Means for detecting a user's confused state with respect to audio output from the device from the user's image captured by the imaging means;
Deforming means for deforming a word incorporated in output data when the confused state is detected from a user's image and causing the voice output unit to re-output the word.The
[0024]
In this invention, the user's confused state is detected from the user image imaged by the imaging means. For example, when the user takes a pose that spreads both hands and does not understand the meaning of the sound output from the device, it is determined that the user is stuck. In the image processing in this case, attention is paid to an image corresponding to both hands of the user, and it is determined that the image is in a confusing state when it is detected that both hands are spread out to the left and right.
[0025]
(4) Voice inputBe doneA voice input unit;
Input voiceofA voice recognition unit for performing voice recognition;
An output data table for storing output data to be output from the device according to the state of interaction with the user;
An output data selection means for determining a conversation status from a voice recognition result and selecting output data to be output at that time from the output data table;
An audio output unit for outputting the selected output data as audio;
Deformation means for detecting the recognition certainty of the input speech, transforming a word incorporated in the output data when the certainty is equal to or less than a threshold value, and causing the speech output unit to re-output.The
[0026]
In the present invention, when the user's input voice cannot be reliably recognized due to surrounding noise conditions or the like, the word incorporated in the output data is transformed. The present invention is applied when confirming the content of voice uttered by a user. For example, when the user inputs voice as “Osaka no Meisho”, the device outputs output data confirming that it is “Okay,” but at this time, the certainty of recognition of the input voice from the user is threshold. If the value is less than or equal to the value, “Osaka”, which is a word incorporated in the output data, is output slowly and reliably. Alternatively, increase the volume and output. As a result, the user can know the degree of recognition by the device side recognition means without fail, and if it is not wrong, a positive response such as “Yes” is given, and if it is wrong, a negative response such as “No” is given. Will do. When the device recognizes the positive response voice input, it enters the next dialog, and when it recognizes the negative response voice input, it further decreases the speed and repeats the voice output, or further increases the volume. Then, a modified output operation such as repeating voice output is performed.
[0027]
(5) An additional word table for each word for storing additional words for explaining words to be incorporated into the output data is provided, and the deformation means deforms by outputting a sound obtained by adding the additional words to the words.The
[0028]
In the modification of the voice incorporated in the output data, it is possible to output the voice by adding an additional word in order to make the contents of the output voice easier to understand. An additional word is a term for explaining a word incorporated in output data. For example, when the target word is “hashi”, “bridge” is added to the word as an additional word. Therefore, in this case, “bridge bridge” is output. The case where the additional word is added is a case where the user receives a voice input indicating a confusion such as “E”.
[0029]
(6) An additional word table for storing character-specific additional words for explaining each character of the word to be incorporated in the output data is provided, and the deforming means generates a sound obtained by adding the additional word to each character constituting the word. Transform by outputThe
[0030]
In addition to the additional word for the word as in (5) above, there are additional words for each character constituting the word. For example, when performing a voice output of “Tokyo”, “Higashi no To” is added as an additional word. Therefore, the voice is output as “Higashi no Tokyo” as a whole. Such a modification is performed, for example, when a user inputs a voice such as “Eh” even if an additional word is added to the word and the voice is output in (5) above. That is, when it seems that it is easier for the user to understand the voice output by adding the additional word for each character constituting the word than by adding the additional word to the word and outputting the voice, this (6) Transformation is performed.
[0031]
(7) Voice inputBe doneA voice input unit;
Input voiceofA voice recognition unit for performing voice recognition;
Means for generating the result of voice recognition as output data in order to repeat the input voice on the device side;
An audio output unit for outputting output data as audio,
A word-by-word additional word table for storing additional words for explaining words to be incorporated into output data, and a transforming means for transforming the built-in words into those obtained by adding additional words and re-outputting them by the output means. AndThe
[0032]
The present invention is applied to a voice input system that recognizes a full sentence of a user's utterance. For example, this apparatus is used when a user reads out a document written on paper into digitized data. That is, the user reads a certain amount of a document, and the apparatus recognizes the voice and outputs the same content as voice. At this time, if there is a homonym, an additional word is added to the word output by the apparatus. When a word expression that can be understood by adding an additional word, that is, a kanji is different, the user inputs a correction by voice. By repeating this, digitized data can be created only by voice input.
[0033]
(8) After the output data in which the modified word is incorporated is output, when a word added with an additional word is input as a corrected word by the user, the corresponding word is converted into the input corrected word. And when the additional word input by the user in the additional word table for each word is not registered, the word with the same sound and the additional word are sequentially output as a candidate and the correct additional word from the user. Table updating means for adding the corresponding word to the word-specific additional word table as a candidate additional word at that time.The
[0034]
In the present invention, in the device of (7) above, when a word added with an additional word is input as a corrected word by the user, the corresponding word is converted into the input corrected word and output again. In addition, if the additional word input by the user at this time is not registered, the homonym word and the additional word are sequentially output as a candidate by voice. When a correct additional word is output as a voice, the user inputs a voice (for example, “Yes”) indicating that the correct additional word is a correct additional word, and at this time, the candidate additional word is newly stored in the table. By this series of operations, the word-by-word additional word table is automatically learned, and a voice output device having high ability can be obtained as it is used.
[0035]
(9) Notification means for performing notification by means of external means such as a lamp, a buzzer, and a vibrator when outputting a word to be incorporated when the word incorporated in the output data is transformed by the transformation means and output by the voice output unit. PreparationThe
[0036]
In the present invention, when a word incorporated in output data is deformed and re-outputted, the user is notified by means of external means such as a lamp, a buzzer or a vibrator at the time of outputting the incorporated word. On the other hand, when important information or user's voice recognition is not certain, it is possible to easily inform the user of points to be emphasized.
[0037]
DETAILED DESCRIPTION OF THE INVENTION
FIG. 1 shows an image diagram of a scene using an audio output device according to an embodiment of the present invention. This audio output device represents an information terminal installed in a street, a government office, a hotel, or the like. Instead of such an information terminal, the voice output device can be applied to a voice dialogue system using a telephone or the like.
[0038]
In the figure, when the user approaches the audio output device, the audio output device outputs a voice message “Welcome. It is a tourist information center.” Since this audio output is not important information, it is output quickly and gently. When the user listens to the guidance, the user asks, for example, “What is the recommended place in Nagaokakyo City?”. After recognizing the voice output from the user, the voice output device determines the conversation status, selects a voice output pattern (output data) to be output at that time from the table, and outputs the voice. At this time, if the user's voice of “Nagaokakyo City” is not sufficiently recognized, “Nagaokakyoshi” is memorized slowly and surely, and “Are you sure?” Is output quickly and swiftly. That is, when the recognition certainty of voice recognition is below a certain threshold value, it is assumed that there is no confidence in voice recognition, and important information (as will be described later, this information is a word in which the entire output data is incorporated). Is transformed and output so as to be surely confirmed by the user.
[0039]
“Yes” is uttered as an affirmative voice from the user in response to the above “Nagaokakyoshi” (slow and reliable output) “Is it OK” (fast and smooth output) output from the device, and this is recognized When it is done, the device will output “Koryouji” (output slowly and reliably as important information) but “Recommend” (output quickly and quickly as unimportant information). At this time, if the user does not know “Komyouji”, he / she utters “Eh” as a voice indicating confusion. When the device recognizes this “Eh”, it considers that “Koumyoji” was not understood, treats it as an obfuscated word, adds an additional word, and outputs it again. The additional words include an additional word for explaining the word itself and an additional word for explaining each character constituting the word. In the example shown in FIG. 1, the characters constituting the word “Komoji” are “Kou”, “Myo”, and “Ji”, and each of these characters is described by an additional word. In other words, the device outputs “Hikaru Akari no Tera no Koji” (slow and reliable output) “I recommend” (fast and light output).
[0040]
In the above example, “Komoji” is treated as an obfuscated word, and this is re-output by adding an additional word for each character. Similarly, when outputting a homonym rather than an obfuscated word, a word or an additional word of each character constituting the word is added and output again. In this way, the user interacts with the device side, and the device side selects and outputs the voice data stored in the table in advance according to the situation, and the words embedded in the voice data are displayed at that time. The output is transformed based on. Note that, for example, words that are incorporated in the entire output data include those that return user input as they are and those that are dynamically generated by the device itself. In the example illustrated in FIG. 1, “Nagaoka Kyoshi” is a word that returns input from the user, and “Koukyo” is a word that is dynamically generated on the device side.
[0041]
As will be described later, whether the word incorporated in the entire output data is a word that returns input from the user or a word that is dynamically generated by the apparatus is stored in advance for each conversation situation. For example, if the dialogue status is the first “greeting” type, the output data is fixed and there are no words embedded in it. In the above example, “Welcome. Tourist Information Center” is the output data when this type is “greeting”. Further, when the conversation status is of the “confirmation” type, the word incorporated in the entire output data is a word that returns input by the user. In the example shown in FIG. 1, “Nagaoka Kyoushi” is a word that incorporates “Nagaoka Kyoshi”. When the conversation status is of the “simple answer” type, the words incorporated in the entire output data are dynamically generated words on the device side. In the example shown in FIG. 1, the word “Komyoji” is included in the output data of “Komyoji is recommended.” This is a portion that is dynamically generated on the device side.
[0042]
FIG. 2 is a system configuration diagram of the audio output device. An input device 1 composed of a microphone is used as an audio input unit, and an output device 2 composed of a speaker is used as output means for outputting output data that is an audio output pattern. The input device 1 and the output device 2 exchange data with an operating system 3 that controls the dialog system. The operating system 3 includes a dialog processing application program 4 that performs dialog processing based on a voice output program. Is implemented. The external recording medium 5 provides the interactive processing application 4 and data for this system. Further, by employing a high-speed recording medium such as a hard disk, it can be provided as a system resident area or user area.
[0043]
FIG. 3 is a schematic block diagram constituting the dialog processing application 4. Here, the interactive processing application 4 has a concept including a data storage area together with a program.
[0044]
In FIG. 3, an input / recognition unit 101 processes and analyzes voice data input from the input device 1 and performs voice recognition. As described above, well-known speech recognition methods such as DP pattern matching method and hidden Markov method are adopted for speech recognition. For example, when “I” is input, this is referred to as “wa” “ta”. It recognizes for each phoneme, like “shi”. Also, based on these recognized phonemes, syntax analysis is performed to recognize the meaning. For the parsing, so-called morphological analysis is adopted. For example, when the phonemes “wa”, “ta”, and “shi” are recognized, they are recognized as “me”.
[0045]
The dialogue processing unit 102 is a part that receives the result from the input / recognition unit 101 and determines the content to be returned to the user. Here, the contents are determined with reference to the output data table 103. The output data table 103 is a part that stores output data from the apparatus according to the state of interaction with the user. FIG. 4 shows a configuration example of this output data table. This table stores output data (audio output pattern) and type. The type indicates the state of interaction with the user. Output data is stored for each dialog state, that is, for each type. Each output data includes <xxx>, <yyy>, and a deformation attribute indicated by () following the output data. <Xxx> and <yyy> indicate words to be deformed. <Xxx> is a part in which an input by the user is returned as it is, and <yyy> is a part into which a result dynamically generated by the system side is inserted. In the example illustrated in FIG. 1, “Nagaoka Kyoshi” corresponds to <xxx>, and “Koukyo” corresponds to <yyy>. The parts of <xxx> and <yyy> are incorporated in the entire output data. The deformation attribute indicated by () will be described later.
[0046]
The transformation unit 104 that transforms the output content is a part that transforms the output data selected by the dialogue processing unit 102. In the deforming unit 104, the character-specific additional word table 105, the word-specific additional word table 106, the deformation counter 107, and the deformation content table 108 are referred to.
[0047]
FIG. 5 shows a configuration example of the modification content table 108. This table defines the deformation type (corresponding to the attribute) and specific deformation contents. For example, if the mold is deformation 0, the deformation content is “no deformation”. That is, it does not deform. When the type is deformation 1, the deformation contents are “high volume” and “low speed”. In other words, control is performed to increase the volume output of <xxx> or <yyy> as a target and to increase the volume output and decrease the voice output speed to speak slowly and slowly. As shown in the figure, the modification contents include a change in volume, a change in speed, the presence / absence of word repetition, the presence / absence of additional words for explanation, and the like.
[0048]
FIG. 6 shows a configuration example of the character-specific additional word table 105.
[0049]
This table is composed of one kanji character and additional words for explaining the character. In this embodiment, as will be described later, this character-by-character additional word table is used when words that are not in the word-by-word additional word table are described and each word is divided into individual characters.
[0050]
FIG. 7 shows a configuration example of the word-by-word additional word table 106. This table includes a character string of words and additional words for explaining the words.
[0051]
FIG. 8 shows a configuration example of the deformation counter 107. The counter includes a word and a count unit that stores the number of times the word has been transformed in the past. The value of the counter is used to determine whether to transform the specified word. The larger this value is, the higher the possibility that the deformation is not performed, and the smaller this value is, the higher the possibility that the deformation is performed.
[0052]
Next, a schematic operation will be described with reference to FIG.
[0053]
ST101 Clear the deformation counter 107 to zero. The deformation counter 107 is used to determine whether to perform deformation by referring to the value. If the value is small, the deformation is performed, and if the value is large, the deformation is not performed. In ST101, this counter is cleared to 0, so that initialization is performed so that deformation is performed.
[0054]
ST102 Speech recognition is performed from input speech data from the user, and morphological analysis is performed to obtain a word string. Also, the certainty of recognition for each word is obtained. A technique for obtaining a certainty of recognition when performing speech recognition is common. For example, in the DP matching method, it is possible to obtain the recognition certainty numerically by obtaining a correlation value with reference data (reference data).
[0055]
ST103 The voice output unit determines whether an additional word is currently being output. The case where the input speech is recognized in ST102 while an additional word is being output is when there is a speech input expressing annoyance such as “I understand” from the user. Therefore, at this time, the output of additional words is interrupted in ST104, and the deformation counter 107 is set to infinity (ST105). In the operations of ST103 to ST105, when there is an interrupt input from the user during the output of the additional word, it is determined that the explanation of the additional word is annoying by the user, and the word is output as the additional word. Do not do it.
[0056]
ST106 When the input voice recognized in ST102 is a voice expressing a confusion such as “E” or “Nani” or a voice indicating a detailed request such as “Details”, the deformation counter 107 is set in ST107. Clear 0. This action indicates that the user did not recognize the audio output from the device or is seeking detailed explanation. Therefore, in this case, the deformation counter is cleared to 0, and the deformation is performed based on the deformation content described in the output data table.
[0057]
ST108 Select data to be output by the system from the output data table 103 based on the current dialog status. In general, the output data is selected in order from the lowest position in the table. In the example shown in FIG. 4, the output data of the “greeting” type is selected first, and then the output data of the “confirmation” type is sequentially selected from the top. Not only the method of selecting the output data in order from the top, but also the output data can be selected dynamically according to the dialogue situation at that time.
[0058]
ST109-ST111 Here, the transformation process described later is performed on each of the transformation target words incorporated in the selected output data.
[0059]
ST112 Speech-synthesizes the transformed data and outputs it.
[0060]
ST113 It is determined whether or not the current output is simply a repetition of the previous output. If it is not a repetition, the transformation counter 107 is incremented for all the transformation target words incorporated in the output. This operation means that, as the number of times of deformation is increased in the past, it is assumed that the user correctly understands the content of the word and that the deformation is not performed.
[0061]
After the above operation is completed, the process returns to ST102 and waits for a new input.
[0062]
FIG. 10 shows a control flow of the deformation processing unit in ST110.
[0063]
ST201, ST202 When the transformation target word is input by the user (corresponding to <xxx>) and the certainty of speech recognition is lower than a predetermined value, the transformation attribute corresponding to the word is set. Turn on (ST204). In the example shown in FIG. 1, when there is a user utterance of “Nagaokakyo City” and the certainty of the voice recognition is low, the process proceeds to ST204, and a deformation attribute is assigned to the corresponding word. The deformation attribute at this time is “deformation 1” because the “confirmation” type dialogue state in FIG. That is, in the actual modification, “volume is high” and “speed is low” from FIG.
[0064]
ST203 When the deformation target word is a part dynamically generated on the system side (corresponding to <yyy>) and the deformation counter 107 does not exceed a certain value, the corresponding deformation attribute for the word (ST204). In the example shown in FIG. 1, when output is performed in the dialog state of “Koryo is recommended” and “Simple answer” type from the system side, if the deformation counter 107 does not exceed a certain value, “Deformation 2”. "Is added as an attribute. Therefore, “Koryouji” (slow and reliable output) “is recommended” (fast and smooth output).
[0065]
FIG. 11 is a control flow of the speech synthesizer in ST112 of FIG.
[0066]
ST301 If there is “additional word” in the content of the transformation, the process proceeds to ST302. Here, it is determined whether or not the corresponding target word is included in the word-specific additional word table 106. If it is included, a corresponding additional word is generated (ST303), and this additional word is used as the target word. Insert immediately before (ST304). Otherwise, for each character included in the word, an additional word for each character is obtained using the character-specific additional word table 105, and the particle “ni” or the like is inserted between the additional words. One additional word string is generated (ST305). This additional word is inserted immediately after the target word (ST306). It is also possible to set a tone color change attribute for the additional word portion. In ST307, this process is performed so that the additional word and the corresponding target word can be heard and distinguished.
[0067]
ST308 Performs an acoustic process according to the content of the deformation and performs speech synthesis.
[0068]
FIG. 12 shows an image diagram of a system using an audio output device of another embodiment of the present invention.
[0069]
This system is a voice input system for recognizing a whole sentence uttered by a user. For example, it is used when a user reads a document written on paper and converts it into digitized data accurately. In this voice input system, in order to create correct digitized data, a dialogue process is performed with the user in order to prevent erroneous recognition when there is a homonym. The example shown in FIG. 12 shows an image of a scene in which the user first uttered “Please watch out for that bridge”. In this case, it is assumed that the speech recognition result of the system is “Be careful about that end”. That is, the system incorrectly recognizes “bridge” as “end”. In this case, the system adds an additional word “Hashikko” to distinguish between “bridge” and “end”. In other words, the system outputs a voice saying "Take care about the end of that shit." At this time, since the user knows that the system is making a recognition error, the user speaks a control word such as “repeat” or “correction” and enters a dialog for correction. For example, when trying to correct, the user specifies a word using an additional word such as “correction XXX of XXX”. In the example shown in FIG. 12, “correction ○ bridge chop” is uttered. At this time, the system determines whether or not “bridge” is registered in the system as an additional word specifying “bridge”. If registered, the correction process is performed and the sound is output again. That is, it outputs a voice saying “Please watch out for that bridge”. If “Bridge” is not registered in the system as an additional word that identifies “Bridge”, possible additions are added to the possible recognition candidates and output. If the output candidate is incorrect, the user utters a word expressing “no” denial. If the system recognizes negation, it outputs the next recognition candidate. As a result of this repetition, when the word requested by the user is output as a voice, the user has the opportunity to speak an affirmative expression “yes”. Then, the system corrects the recognition result, and additionally adds and updates the additional word “bridge” used by the user to the system. From then on, “bridge” can be used as an additional word to identify “bridge”.
[0070]
In the example shown in FIG.
-After the user utters "Please watch out for that bridge", know that the system is doing the wrong recognition result, utter the word "repeat" and the system recognizes it Output the same content again,
・ The user utters “Correction Bridge's chopsticks”, and the system outputs “Chopsticks to eat rice” as the first candidate.
-The user utters "No" and the system outputs as the next candidate "Is it a bridge to cross?"
The user utters “Yes” and the system outputs “Correct. Please watch out for this bridge.”
[0071]
The configuration diagram of the system is the same as the configuration shown in FIG. The main control flow is the same as the flow shown in FIG. FIG. 13 is a control flow when the user performs correction processing by uttering “correction XXX of XXX”.
[0072]
ST601 Here, it is determined whether or not the result of recognizing the input voice (ST102 in FIG. 9) corresponds to a correction command. That is, it is determined whether or not the user has uttered “correction of XXX”. In the case of a correction command, the process from ST602 is processed.
[0073]
-ST602 From the timing of the correction command uttered by the user, it is specified which word of the output data of the system the user is using for correction. In the example shown in FIG. 2, since the user uttered “correction” immediately after the system recognized the utterance of “repetition” and the system output “that hakkoko” as a voice, The word to be corrected immediately after the voice output is specified as the word.
[0074]
ST603 Extracts corrected and additional words included in the correction command. In the example shown in FIG. 12, the corrected word is “Hashi” and the additional word is “Bridge”.
[0075]
ST604 It is determined whether the corrected and additional words are already registered in the character-specific additional word table 105 or the word-specific additional word table 106 of the system. If registered, the process proceeds to ST605, and a word corresponding to the additional word is selected. If not, in ST606, recognition candidate words such as homonyms are sequentially output together with registered additional words, and confirmation is requested from the user.
[0076]
ST607 Here, it is determined whether or not there is a word requested by the user during the output of candidates that are sequentially output. This determination is made based on whether or not the user has spoken a negative or affirmative expression such as “No” or “Yes”. If there is a word requested by the user, in ST608, the additional word is newly registered in the word-specific additional word table or the character-specific additional word table. If not, the fact that it was not found is output to the user. Encourage re-entry of the original text.
[0077]
. ST609 Subsequently, the recognition content is corrected, the correction result is output as a voice (ST601), and the process returns.
[0078]
FIG. 14 is an image diagram of a system using an audio output device according to another embodiment of the present invention.
[0079]
This system has a camera attached to the apparatus shown in FIG. This camera detects the user's confusion. In the example shown in FIG. 1, when the user does not understand the contents of the audio output of the system, he / she expresses the confusion by uttering “E” or “Nani”, but when the user feels confusion, for example, There is a habit and habit of spreading both hands. Moreover, you may decide as a rule beforehand. In such a case, it is possible to detect that the user is in a confused state by detecting the spread state of the hand based on the image. Therefore, in this case, the user can notify the system of a scam by simply spreading his / her hand without uttering “E” or “Nani”.
[0080]
FIG. 15 is an image diagram of a system using an audio output device according to another embodiment of the present invention.
[0081]
In this system, in addition to the operation of transforming the output data, the lamp is turned on or blinked in accordance with the output timing when the word to be incorporated is output. In the example shown in FIG. 15, when the certainty of voice recognition is below a certain level, the lamp is blinked at the same time when the target word (where you are not confident in voice recognition) is output. In addition, the lamp is also blinked for a part that outputs important information dynamically generated on the system side. Accordingly, in the image diagram shown in FIG. 15, “Nagaoka Kyoshi” and “Kou Koji” are output as voices while blinking the lamp.
[0082]
The blinking operation of the above lamps may be performed independently without performing the sound transformation process for those words.
[0083]
FIG. 16 shows an image diagram of a system using an audio output device according to another embodiment of the present invention.
[0084]
In this system, a vibrator is used instead of the lamp of FIG. That is, the vibrator is vibrated instead of blinking the lamp. In this case, the user has the vibrator.
[0085]
In addition, it is also possible to use a small output buzzer in place of the lamp or the vibrator.
[0092]
【The invention's effect】
  thisAccording to the invention, in the system that recognizes the whole sentence uttered by the user and converts it into computerized data, even if there is a homonym or an esoteric word, an error is caused by adding an additional word and confirming the output. It is possible to prevent electronic data from being created.
[0093]
  Also,Even if no additional word is registered on the device side, candidates are sequentially output, and the additional word is based on the voice output indicating affirmation from the user when the semantically correct additional word is output. Can be registered as a new additional word, and the additional word table can be learned and updated, so that a more accurate and smooth dialogue can be made according to the number of uses.
[Brief description of the drawings]
FIG. 1 is an image diagram of a system to which an audio output device according to an embodiment of the present invention is applied.
FIG. 2 is a diagram showing a system configuration
FIG. 3 is a schematic block diagram of the system.
FIG. 4 shows a configuration example of an output data table.
FIG. 5 shows a configuration example of a modified content table.
FIG. 6: Additional word configuration example by character
FIG. 7 shows an example of additional word structure for each word.
FIG. 8 shows a configuration example of a modified counter.
FIG. 9 shows a schematic flow.
FIG. 10 is a diagram showing a control flow of a deformation processing unit.
FIG. 11 is a diagram showing a control flow of the speech synthesizer
FIG. 12 is an image diagram of a system using an audio output device of another embodiment.
FIG. 13 is a diagram showing a control flow of a correction processing unit
FIG. 14 is an image diagram of a system using an audio output device according to another embodiment.
FIG. 15 is a conceptual diagram of a system using still another audio output device.
FIG. 16 is a conceptual diagram of a system using still another audio output device.

Claims (3)

An audio input unit for inputting audio;
And a voice recognition unit that performs speech recognition of the input speech,
Means for generating the result of voice recognition as output data in order to repeat the input voice on the device side;
An audio output unit for outputting output data as audio,
An additional word table for each word for storing additional words for explaining the words to be incorporated into the output data;
Deformation means for transforming the word to be incorporated into an added word to the word and re-output by the output means;
Means for converting a corresponding word into the inputted correction word when a word added with an additional word is inputted as a correction word by voice after the output data incorporating the modified word is outputted; ,
When an additional word inputted by voice from the user is not registered in the additional word table for each word, an input indicating that the word with the same sound and the additional word are sequentially outputted as a candidate and is a correct additional word from the user. If there exists, the audio | voice output apparatus provided with the table update means which adds an applicable word to the said additional word table classified by word as a candidate additional word at that time. When a word incorporated in the output data is transformed by the transforming means and output by the voice output unit, a notification means is provided for performing notification by an external voice means such as a lamp, a buzzer, or a vibrator when the incorporated word is output. Item 2. The audio output device according to Item 1 . A voice recognition step for performing voice recognition of the input voice input to the voice input unit;
In order to repeat the input voice on the device side, an output data generation step for generating a result of voice recognition as output data;
An audio output step for outputting the output data generated in the output data generation step as audio;
Using the word-specific additional word table storing additional words for explaining the words to be incorporated into the output data, transforming the incorporated words into those obtained by adding the additional words to the speech output step. A transformation step to re-output with
After the output data including the word transformed in the transformation step is output, when a word added with an additional word is input as a corrected word by the user, the corresponding word is converted into the input corrected word. Correction steps to
When an additional word inputted by voice from the user is not registered in the additional word table for each word, an input indicating that the homonym word and the additional word are sequentially outputted as a candidate and are a correct additional word from the user. If there is, a table update step of adding the corresponding word to the word-by-word additional word table as a candidate additional word at that time,
A recording medium on which a sound output program for causing a computer to execute is recorded.

Images