JP2006259641A - Voice recognition device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To simply add additional information consisting of letters of special characters, signs and/or images to the voice recognized letter string. <P>SOLUTION: This voice recognition device 12 is provided with a sound detector 15 to detect various sound data about the voice input from a microphone 11, the main converter 17 to create a letter string matching the input voice from the sound data detected by this sound detector 15, an additional information converter 18 to create additional information added to the letter string from the sound data detected by this sound detector 15, and a converter connector 19 to create the output information made by adding the additional information to the letter string. The additional information converter 18 is provided with a corrector 21 to correct the rhythm information detected by the sound detector 15, and an additional information selector 22 to choose the additional information showing the feeling by estimating the talker's feeling contained in the input voice from the corrected rhythm information. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a speech recognition device and a speech recognition program, and more particularly, a speech recognition device that can easily add additional information such as emoticons and special characters of pictograms when inputting characters by speech. And a speech recognition program.

  Recently, with the widespread use of personal computers and mobile phones, communication using characters such as e-mail has been actively performed. Accordingly, the diversity of communication has been pursued. As a result, not only simple sentences but also emoticons and pictograms composed of various characters and symbols (hereinafter “ An expression using a combination of “special characters” is used. These special characters can be input by using a keyboard, a numeric keypad, etc., like normal characters, due to recent improvements in word processor conversion dictionary functions.

By the way, a voice recognition technology for converting human voice into a character string (text) has been developed, and the voice recognition technology enables character input to an electronic mail or the like by voice. As such speech recognition technology, a speech recognition system is known in which a color or font is partially changed from a change in input speech information to a character string converted based on the speech information (Patent Document). 1).
JP 2001-134290 A

  However, although the voice recognition system can change the color and font of the converted character string, there is a disadvantage that additional information such as the special character cannot be input. Further, if the additional information is added to the converted character string using the current speech recognition technology, there is a problem that troublesome input is required on the user side. That is, when inputting a special character by voice, a voice meaning the type of the character, for example, a voice meaning special characters such as “emoticon 1”, “emoticon, joy”, “emoticon, smiling” In addition, it is necessary for the user to input a voice or the like meaning the content of the input character in a complex manner, and a troublesome input operation is required.

  The present invention has been devised by paying attention to such inconveniences, and its purpose is to add additional information consisting of characters such as special characters, symbols and / or figures, etc. to a character string that has been voice-recognized. An object of the present invention is to provide a speech recognition apparatus and a speech recognition program that can be easily added.

  (1) In order to achieve the above object, the present invention includes an additional information conversion unit that creates additional information to be added to a character string converted from input speech, and the additional information conversion unit includes the prosody of the input speech. It is configured to be created based on information.

  (2) Moreover, it is preferable that the said additional information conversion part takes the structure of estimating the emotion embedded in the said input audio | voice from the said prosodic information, and creating the additional information expressing the said emotion.

(3) Furthermore, the additional information conversion unit is configured to add the additional information based on the prosodic information of the entire target portion of the input speech to which the additional information is added and the prosodic information of the final mora portion in the target portion. It is preferable to adopt a configuration in which the content of information is determined.
In the present specification, “mora” is a unit corresponding to beat sensation. For example, an utterance of “onsei” is counted as 4 mora.

  (4) In addition, the additional information conversion unit may also employ a configuration in which it is determined whether or not the additional information should be added for each section divided by a pause of a predetermined time or more.

  (5) Further, the additional information conversion unit may create the additional information using the normalized prosodic information.

(6) The present invention is a speech recognition program for causing a computer to execute speech recognition processing,
A configuration is adopted in which the computer is caused to create additional information added to the character string converted from the input speech based on the prosodic information of the input speech.

  According to the present invention, since additional information is created based on the prosody information of the speaker who is the user, the emotion of the speaker can be estimated from the prosodic information, and additional information corresponding to the emotion can be automatically created. Therefore, when creating the additional information, it is not necessary for the speaker to input the voice that means the content of the additional information independently, and the additional information is added to the character string with high accuracy by a simple input method that is easy for the user to become familiar with. be able to.

  Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 shows a schematic configuration diagram of a speech recognition system to which a speech recognition apparatus according to the present embodiment is applied. In this figure, a voice recognition system 10 is incorporated in a terminal such as a personal computer, a portable information terminal, a mobile phone, etc., and a microphone 11 to which a voice of a speaker who is a user is input, and an input voice to the microphone 11 The voice conversion device 12 for creating output information composed of characters, symbols and / or figures, and the display device 13 such as a liquid crystal display for displaying the output information created by the voice conversion device 12 are provided. Here, the microphone 11 and the display device 13 of the present embodiment employ a known configuration, and a detailed description of each configuration is omitted here.

  The voice recognition device 12 is configured by hardware and / or software, and includes a plurality of program modules and / or processing circuits such as a processor, and is installed with a program capable of executing various processes described below. .

  Specifically, the voice recognition device 12 includes a voice detection unit 15 that detects various voice data related to input voice from the microphone 11, and a character corresponding to the input voice from the voice data detected by the voice detection unit 15. A main conversion unit 17 that creates a string, an additional information conversion unit 18 that creates additional information to be added to the character string from the voice data detected by the voice detection unit 15, and a character created by the main conversion unit 17 A conversion coupling unit 19 that creates output information with the additional information created by the additional information conversion unit 18 added to the column is configured.

  The voice detection unit 15 is provided so as to be able to detect voice data required by the main conversion unit 17 and the additional information conversion unit 18, and one of them is the frequency (Hz) at which the prosody information, that is, the tone is used. , Volume (dB), and speed (number of mora / hour) are detected.

  The main conversion unit 17 expresses the input voice as text with respect to the input voice data from the microphone 11 by using any one or more of various conventionally known voice recognition techniques. ing. Note that this voice recognition technology is not the gist of the present invention, and thus detailed description thereof is omitted.

  The additional information converting unit 18 corrects the prosodic information detected by the voice detecting unit 15 and estimates the emotion included in the input speech from the corrected prosodic information to express the emotion. Additional information selection means 22 for selecting additional information is provided.

  The correction means 21 performs normalization processing (standardization processing) on the prosodic information detected by the voice detection unit 15. In this normalization process, even in the prosodic information of speech expressing the same emotion, the degree of loudness such as the pitch of the voice differs depending on individual differences such as gender differences. This is done for application to certain standards. That is, here, as will be described later, an average value and standard deviation of each prosodic information that is unique data of a speaker individual performing speech recognition is obtained in advance, and speech recognition is performed using the average value and standard deviation. Each prosodic information of the speech of the speaker input when performing is normalized.

  The additional information selection unit 22 is corrected by the correction unit 21 as a target portion to which the additional information is added to each voice segment divided by a pause of a predetermined time or more when the speaker inputs voice. Whether or not to add the additional information and its type are determined based on the value of each prosodic information after that. Here, as additional information in the present embodiment, the emphasis expression in which all the character strings of the target part are bolded, the paragraph change in which the end of the character string of the target part is broken, and the end of the character string of the target part are added. There are five types of emoticons (angry), emoticons (joy), and emoticons (sad). The additional information is not limited to these five types, and various information such as characters, symbols, and / or figures including other emoticons and pictograms may be used in association with each prosodic information. Is possible.

  Here, the selection criteria for each additional information is whether each prosodic information of the entire target part and each prosodic information of the final mora that is the final beat part of the utterance are parameters, and whether or not those values exceed a predetermined reference value Specifically, the determination is made by the decision tree method shown in the flowchart of FIG. This decision tree is constructed and set in advance by data accumulated by causing a plurality of speakers to utter the same sentence according to emotions designated in advance. This decision tree may be constructed by learning by utterance on the user side.

  In the additional information selection means 22 in the present embodiment, at the time of determination by the decision tree of FIG. 2, at least the average value of the frequency and the average value of the volume are obtained as the prosodic information of the entire target portion, while the prosodic information of the final mora is At least the phrase end-of-frequency increase, the average value of the volume, and the speed are obtained. The frequency end-of-phrase rise is the slope of a straight line obtained by approximating the frequency curve of the entire final mora to a straight line by the method of least squares. Regarding the frequency and volume, it is better to determine the maximum value, the minimum value, the range that is the width of the entire target portion and the final mora, and to determine additional information using a more complicated decision tree. In this case, the recognition accuracy is further improved.

  Next, a voice recognition procedure by the voice recognition device 12 will be described.

  First, initial setting is performed. That is, a speaker who performs voice recognition using the voice recognition device 12 utters a plurality of types of sentences several times each, and the voice is input to the microphone 11. Then, the voice detection unit 15 detects the frequency, volume, and speed that are the prosody information of the voice. Next, the correction means 21 calculates an average value and a standard deviation for the frequency, volume, and speed detected by the voice detection unit 15. As will be described later, these average value and standard deviation are used for normalization processing on the prosodic information of the input speech when converting to the output information.

  After the above initial setting is performed, a voice for voice recognition is input to the microphone 11 on the speaker side as a user. For example, suppose a speaker speaks with anger, "Why didn't you come yesterday?" At this time, because the speaker wants to add an emoticon (angry) at the end of the sentence, is the pose before and after "I didn't come" and the pronunciation of the part "I didn't come" really expresses anger? As a whole, it is assumed that the final mora “TANO” is uttered early and gradually shifted from a low sound to a high sound (from a low frequency to a high frequency).

  Then, the main conversion unit 17 converts a character string (text) that matches the content of the input voice. That is, in the previous example, the input voice is converted into the character string “Why did not come yesterday”.

  Further, the additional information conversion unit 18 creates additional information for each part delimited by the pause. In other words, in the previous example, the determination of addition of additional information and the selection thereof are performed for the part “Why yesterday” in the previous stage and the part “Did not come” in the subsequent stage.

First, the detection value detected by the voice detection unit 15, that is, each prosodic information (frequency, volume, speed) of the input voice is normalized by the correcting means 21, and a correction value of each prosodic information is obtained. This correction value is obtained by the following equation using the average values and standard deviations of the frequency, volume and speed of the individual voice of the speaker obtained at the time of the initial setting.
Correction value = (Detected value-Average value) / Standard deviation (1)

  Then, the additional information is selected by the decision tree method of FIG. 2 using the obtained correction value for the target portion to which the additional information is added. In the case of the previous example, from each correction value, the “Why yesterday” part, which is normally pronounced by the decision tree of FIG. 2, is judged as “no additional information”, while the emotional “not come” An emoticon (angry) is selected as additional information for the “Tano” part.

  Then, the conversion coupling unit 19 adds the additional information created by the additional information conversion unit 18 to the character string created by the main conversion unit 17. In the previous example, after the character string “Why did not come yesterday” created by the main converter, the emoticon (angry) created by the additional information converter 18 is added. The output information thus completed is displayed on the display device 13.

  Therefore, according to such an embodiment, additional information expressing the emphasis or emotion is determined with reference to the prosodic features when the human actually expresses emphasis or emotion. When adding special characters such as emoticons, voice input with corresponding emotions can provide an effect of easily adding special characters without performing special operations or inputs.

  In addition, when adding additional information, the prosody information in which the characteristic of emotion expression appears, that is, the prosodic information of the entire target part and the prosodic information of the final mora, the identification rate for the additional information corresponding to the user's emotion is set. It becomes possible to greatly improve.

  Note that the voice recognition system 10 is configured by incorporating the entire system into one piece of hardware such as a terminal as in the above embodiment, and the configuration of the microphone 11 and the display device 13 is left on the user terminal side, and the voice The recognition device 12 may be incorporated in a computer such as a server in a remote location, and remote voice recognition processing may be performed using a network communication network such as the Internet in a remote location.

  The configuration of each part of the apparatus in the present invention is not limited to the illustrated configuration example, and various modifications are possible as long as substantially the same operation is achieved.

1 is a schematic configuration diagram of a voice recognition system according to an embodiment. The flowchart for demonstrating a decision tree system.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speech recognition system 11 Microphone 12 Speech recognition apparatus 13 Display apparatus 15 Voice detection part 17 Main conversion part 18 Additional information conversion part 19 Conversion coupling | bond part 21 Correction means 22 Additional information selection means

Claims (6)

  A speech recognition system comprising: an additional information conversion unit that creates additional information added to a character string converted from an input speech, wherein the additional information conversion unit is created based on the prosodic information of the input speech apparatus.   The speech recognition apparatus according to claim 1, wherein the additional information conversion unit estimates an emotion embedded in the input speech from the prosodic information, and creates additional information expressing the emotion.   The additional information conversion unit determines the content of the additional information based on the overall prosodic information of the target portion of the input speech to which the additional information is added and the prosodic information of the final mora portion in the target portion. The speech recognition apparatus according to claim 1 or 2, wherein   The speech recognition apparatus according to claim 1, 2 or 3, wherein the additional information conversion unit determines whether or not the additional information should be added for each section divided by a pause of a predetermined time or more.   The speech recognition apparatus according to claim 1, wherein the additional information conversion unit creates the additional information using the normalized prosodic information. A speech recognition program for causing a computer to execute speech recognition processing,
A speech recognition program that causes the computer to create additional information added to a character string converted from input speech based on prosodic information of the input speech.

Images