JP2004117771A - Method and device for preparing dictionary for speech recognition, method and device for speech recognition, speech recognition program, and voice recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and a device for preparing a dictionary for speech recognition for realizing a higher recognition rate by simplifying selection of words and sentences to be stored in an erroneous recognition countermeasure dictionary necessary for speech recognition and selecting words and sentences being apt to be erroneously recognized, from word groups being commonly used and to provide a speech recognition device and a speech recognition system which are provided with these method and device. <P>SOLUTION: In the speech recognition device, terms having a prescribed reading length or shorter except terms of which the speech recognition results are decided to be correct are extracted from a database where terms in everyday use consisting of words and sentences are enumerated, and the erroneous recognition countermeasure dictionary is prepared. When a term entered in this dictionary is obtained as a result of speech recognition, this term is excluded from the recognition result. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech recognition device and a speech recognition system, and more particularly to creation of a dictionary used for speech recognition.
[0002]
[Prior art]
There is a voice recognition technology as a technology for recognizing an operation instruction by voice and controlling a device. Speech recognition controls devices based on unstable conditions such as human voices, and does not control them using electrically stable signals such as hitting a keyboard or turning on an operation switch. In one example, there is a large individual difference in speech so that it is possible to determine who it is just by listening to the voice. Furthermore, even if the same person utters the same word, the sound may be different depending on the day.
[0003]
For this reason, in the related art, a feature is extracted from a word to be speech-recognized, parameterized, and compared with an input speech signal, thereby adopting a speech recognition method in which such an unstable element is eliminated as much as possible ( For example, see Patent Document 1). However, there is a case where the speech recognition device using the method of Patent Document 1 cannot recognize the word yet, and depending on the manner of utterance and the surrounding noise, it is not rare that the word is recognized as a different word from the uttered word.
[0004]
For example, a method of providing a dictionary in which words to be excluded from recognition results are registered, that is, a dictionary for countering erroneous recognition, in addition to a speech recognition dictionary in which words to be recognized are registered, as in Patent Literature 2, for example. By preliminarily registering a word that is easily recognized by mistake but that cannot be approximately an operation instruction, a word that is incorrectly recognized is detected in advance and is excluded from the recognition result.
[0005]
However, in the case of Patent Literature 2, a word that is easily misrecognized is extracted by the method described in the paragraph [0014]. First, vowels are extracted for each character at the time of reading a word to be recognized, and the number of vowels is accumulated for each character appearance position. Words that have vowels that do not appear at the appearance position should not be words that you want to recognize when viewed at the appearance position of the character, so create a word that has such a character at the appearance position of each vowel, and It is stored as a misrecognition prevention dictionary.
[0006]
This method has a problem that if the number of words to be recognized increases, it takes time to extract and examine vowels, and the processing becomes complicated. In addition, words to be recognized by voice are words that can be used as operation instructions for operating a certain device, and the number of words should be naturally limited. Since the words to be registered in the misrecognition countermeasure dictionary are generated based on the words to be recognized, the variations of the words to be generated are naturally limited.
[0007]
However, since speech recognition is originally used in the life of the user, in order to effectively eliminate words that are easily misrecognized, the general use of words by the user must be Words to be registered in the misrecognition measures dictionary should be selected.
[0008]
Also, the method of Patent Document 2 can cover a wide range of words by generating a large number of words to be registered in the misrecognition countermeasure dictionary. However, it is generally desired to reduce the storage amount required by the misrecognition countermeasure dictionary as much as possible. is there. When a large number of words are generated, if only words that are unlikely to be used usually are used, the amount of storage increases only uselessly.
[0009]
[Patent Document 1] JP-A-11-143485
[0010]
[Patent Document 2] Japanese Patent Application Laid-Open No. 2001-184085
[0011]
[Problems to be solved by the invention]
The present invention simplifies the selection of terms to be stored in the misrecognition countermeasure dictionary necessary for speech recognition, and selects a term that is easily misrecognized from a group of commonly used terms, thereby achieving a higher recognition rate for speech recognition. An object of the present invention is to provide a dictionary creation method, a creation device, a speech recognition method, and a speech recognition device and a speech recognition system using the same.
[0012]
[Means for Solving the Problems]
According to the method of the present invention,
A speech recognition dictionary creation method for creating a misrecognition countermeasure dictionary used to determine that a speech recognition result is misrecognition,
From the database that lists everyday terms, extract terms whose reading is less than a predetermined length,
As a result of the speech recognition, using a speech recognition dictionary listing the predetermined terms to determine that correct speech recognition has been performed if a term comprising a predetermined word or sentence is obtained,
From the extracted terms, select a term that differs in the character arrangement or character length of the reading of the terms listed in the speech recognition dictionary.
A method for creating a dictionary for speech recognition, comprising creating an erroneous recognition countermeasure dictionary using the selected term.
Is provided.
Also,
A speech recognition method using a speech recognition dictionary listing terms that are determined to be correct as the speech recognition result, and an erroneous recognition countermeasure dictionary used to determine that the speech recognition result is erroneous recognition,
From a database that lists everyday terms consisting of words and sentences, extract terms whose reading is less than a predetermined length,
From the extracted terms, select a term that differs from the character arrangement or length of the reading of the terms listed in the speech recognition dictionary.
Using the selected term to create the misrecognition measures dictionary,
When performing voice recognition,
If the speech recognition result indicates the same term as that listed in the misrecognition countermeasure dictionary, this result is excluded from the recognition result.
Voice recognition method characterized by the following:
Is provided.
Further, a speech recognition dictionary creating device, a speech recognition device, a speech recognition program, and a speech recognition system that realize these methods are provided.
[0013]
BEST MODE FOR CARRYING OUT THE INVENTION
FIG. 1 shows an example of a speech recognition system according to an embodiment of the present invention. FIG. 1 shows a sound collection device 101, a speech recognition device 102, an output device 103, a dictionary creation device 104, and a network 105. The sound collection device 101, the voice recognition device 102, and the dictionary creation device 104 are connected so as to be able to communicate via a network 105.
[0014]
The sound collection device 101 has a function of collecting sound information such as surrounding sounds with a microphone or the like and transmitting the sound information as a sound signal to the speech recognition device 102 via a network.
[0015]
The voice recognition device 102 receives the voice signal transmitted by the sound collection device 101, and analyzes the voice included in the signal. When it is determined from the result of the analysis that the voice is a voice, it has a function of recognizing what word the voice is, and outputting a signal based on the recognition result.
[0016]
The output device 103 is a device that receives an output signal of the speech recognition device 102 and performs an operation according to a speech recognition result. For example, it corresponds to a display device such as a display device or a buzzer, a home appliance such as a television, or a switch such as a relay.
[0017]
The dictionary creation device 104 has a function of creating a dictionary used by the speech recognition device 102 for speech recognition. The created dictionary is transmitted to the voice recognition device 102 via the network 105.
[0018]
The network 105 realizes communication among the sound collection device 101, the voice recognition device 102, and the dictionary creation device 104. This network includes, for example, a wired LAN such as Ethernet (R), a wireless LAN defined by IEEE802.11b, and Bluetooth developed for short-range communication. ^(TM) Alternatively, IrDA (Infrared Data Association), which is an infrared communication standard, is also applicable. The form of communication is not limited to those described above, and any network may be used as long as they can communicate with each other.
[0019]
FIG. 2 shows an example of a block diagram of the speech recognition apparatus 102 according to one embodiment of the present invention. FIG. 2 shows a voice processing unit 201, a recognition unit 202, an output unit 203, a communication processing unit 204, a dictionary update unit 205, a recognition dictionary 206, and a misrecognition countermeasure dictionary 207.
[0020]
The voice processing unit 201 has a function of shaping a voice signal obtained from the sound collection device 101 into a signal suitable for voice recognition. First, the audio signal transmitted by the sound collection device 101 is input to the audio processing unit 201 by the communication processing unit 204. This audio signal contains distortion and noise that have been introduced during communication. The voice processing unit 201 corrects such distortion and removes noise unnecessary for voice recognition via a filter circuit. Thereafter, the output signal level and the like are adjusted and output to the recognition unit 202.
[0021]
The recognition unit 202 has a function of receiving a voice signal output from the voice processing unit 201 and recognizing voice included in the voice signal. There are many known signal analysis methods performed to determine what word or sentence signal is included in an audio signal. In the present embodiment, for example, an analysis method based on a feature parameter obtained by extracting a feature included in an audio signal when a word is uttered as described in Patent Literature 1 may be used. In the recognition unit 202 of the present embodiment, the method of analyzing the uttered word or sentence included in the audio signal from the audio signal is not limited to the above-described method, and may be any method. With respect to the spoken word or sentence included in the audio signal, any method can be used as long as the method can extract the word or sentence.
[0022]
The extracted words or sentences are matched with the information stored in the misrecognition countermeasure dictionary 207. If no match is found as a result of the matching, the information is further matched with the information listed in the recognition dictionary 206. As a result of the matching, if there is the same word or sentence listed here, it is determined that the extracted word or sentence is a correct instruction. It is determined that instructions other than the instruction based on the word or sentence determined in this way are not correct instructions.
[0023]
When it is determined that the word or the sentence of which the feature is extracted is a correct instruction, a signal is output to the output unit 203 in accordance with the processing executed by the word or the sentence.
[0024]
The output unit 203 has a function of receiving a signal of the speech recognition result of the recognition unit 202 and performing an operation based on the signal. The output unit 203 controls the output device 103 connected to the speech recognition device 102. For example, power on / off of the output device 103 is performed. If the output device 103 is a television, besides turning on and off the power, operations such as switching the image receiving channel and increasing the volume may be performed.
[0025]
The communication processing unit 204 has a function of communicating with the sound collection device 101 and the dictionary creation device 104 via the network 105. The audio signal transmitted from the sound collection device 101 is output to the audio processing unit 201. When the dictionary information created by the dictionary creating device 104 is received, the received dictionary information is output to the dictionary updating unit 205.
[0026]
The dictionary updating unit 205 has a function of updating the recognition dictionary 206 and the misrecognition countermeasure dictionary 207 based on dictionary information obtained from the dictionary creation device 104 via the network 105.
[0027]
The recognition dictionary 206 is a dictionary including a storage device such as a memory in which words and sentences as operation instructions when performing voice recognition by the recognition unit 202 are stored. It is referred to when the recognition unit 202 extracts a word or a sentence included in the voice signal, and if there is the same word or sentence listed in this dictionary, it is determined that the recognition unit 202 has given a correct instruction by voice. to decide.
[0028]
For example, when the output device 103 is a television, “loud” is uttered to increase the volume of the television. When the recognition unit 202 receives the voice signal “loud” issued by the user, when the word or the sentence is in the recognition dictionary 206, the recognition unit 202 sends the sound to the output device 103 via the output unit 203 so as to increase the volume of the television. When a user outputs a signal and gives an instruction by voice, the operation to be performed on the output device 103 usually differs depending on the word or sentence to be uttered. Therefore, the recognition unit 202 outputs the recognition unit 202 for each listed voice or sentence. The output signals (commands) given to 203 are also listed in association with each other. Preferably, words and sentences stored in the recognition dictionary 206 and output signals should be exchanged. With such a configuration, the present invention can be applied even to the output device 103 having different words and sentences to be recognized and different output signals.
[0029]
At the time of speech recognition by the recognition unit 202, words or sentences that are not stored in the recognition dictionary 206 in advance are not recognized as correct instructions.
[0030]
The misrecognition countermeasure dictionary 207 is a list of words and sentences extracted from the features and extracted by the recognizing unit 202 that are not clearly an operation instruction to the output device 103. The words and sentences stored here are words and documents that are to be first eliminated by the recognition unit 202 after the feature extraction of the words and sentences by the recognition unit 202, assuming that they are not operation instructions to the output device 103. Selection of a word or a sentence to be stored in the misrecognition countermeasure dictionary 207 will be described in the following dictionary information creation method. Preferably, words and sentences stored in the misrecognition countermeasure dictionary 207 should be replaced. With such a configuration, the present invention can be applied to a case where a word or a sentence to be determined to be erroneously recognized is changed for each output device 103 or arbitrarily.
[0031]
FIG. 3 shows an example of a block configuration diagram of the dictionary creation device 104 according to an embodiment of the present invention. FIG. 3 shows a word dictionary 301, an erroneous recognition countermeasure dictionary creation unit 302, a recognition dictionary creation unit 303, a communication control unit 304, recognition dictionary information 305, and erroneous recognition countermeasure dictionary information 306.
[0032]
The word dictionary 301 is a database that lists words and sentences used in daily conversation. The words and sentences stored here can be, for example, items listed in a commercially available Japanese language dictionary. Commercially available Japanese language dictionaries contain many words and sentences frequently used in ordinary conversation. It is suitable for covering words and sentences uttered in daily conversation and the like. It is also possible to create and use a word dictionary 301 excluding items that are considered to be uttered less frequently as spoken words from items listed in a commercially available Japanese language dictionary. In this way, the capacity of the database constituting the word dictionary 301 can be reduced. The word dictionary 301 can be configured by any storage device such as a magnetic disk device, a memory, and a CD-ROM.
[0033]
The misrecognition countermeasure dictionary creation unit 302 has a function of selecting a word or a sentence from the words or sentences stored in the word dictionary 301 according to a predetermined rule and storing the selected word or sentence in the misrecognition countermeasure dictionary information 306. The rules for extracting words and sentences from the word dictionary 301 performed by the misrecognition countermeasure dictionary creation unit 302 will be described in the following dictionary information creation method.
[0034]
The recognition dictionary creation unit 303 has a function of storing words and sentences to be recognized by the voice recognition performed by the voice recognition device 102 in the recognition dictionary information 305. The word or sentence to be recognized is a word or sentence issued by the user for controlling the output device 103 by the speech recognition device 102. As a method of inputting a word or a sentence to the recognition dictionary information 305 by the recognition dictionary creating unit 303, a method of typing from a keyboard (not shown) provided in the dictionary creating device 104 can be considered. Once the output device 103 to be controlled is determined, words and sentences suitable for control by voice are determined. Taking the above-mentioned television as an example, it is sufficient for basic control to determine words and sentences such as “turn on”, “turn off”, “o channel”, “large” and “small”.
[0035]
Once the words or sentences to be recognized by voice recognition according to the control target and what output signals are to be output to the output device 103 when they are given as voice signals are determined, the user or the developer Hit the keyboard to enter information. The input information is processed by the recognition dictionary creation unit 303 and stored in the recognition dictionary information 305.
[0036]
Alternatively, the recognition dictionary creation unit 303 acquires a voice signal of a word or a sentence stored in the recognition dictionary information 305 from the sound collection device 101 via the network 105, and stores a speech-recognized speech signal in the recognition dictionary information 305. May be configured.
[0037]
The communication processing unit 304 has a function of communicating with the sound collection device 101 and the voice recognition device 102 via the network 105. The information to be communicated includes a voice signal transmitted from the sound collection device 101 and dictionary information for transmitting the recognition dictionary information 305 and the misrecognition countermeasure dictionary information 306 to the voice recognition device 102.
[0038]
The recognition dictionary information 305 is a database in which words and sentences that the speech recognition device 102 wants to recognize as a control instruction to the output device 103 are listed by the recognition dictionary creation unit 303. The recognition dictionary information 305 can be configured by any storage device such as a magnetic disk, a memory, and a CD-ROM.
[0039]
The words and sentences listed in the recognition dictionary information 305 are exchanged according to the output device 103 to be controlled. With this configuration, the present embodiment can be applied even if the words and sentences uttered for controlling the control target and the output signal are different.
[0040]
The communication processing unit 304 transmits the word and sentence information stored in the recognition dictionary information 305 via the network 105 as dictionary information for updating the recognition dictionary 206 included in the voice recognition device 102.
[0041]
The misrecognition countermeasure dictionary 306 is a database in which readings of words and sentences to be excluded from the recognition result by the speech recognition device 102 by the misrecognition countermeasure dictionary creation unit 302 are listed. The misrecognition countermeasure dictionary 306 can be configured by any storage device such as a magnetic disk, a memory, and a CD-ROM.
[0042]
The words and sentences listed in the misrecognition dictionary information 306 can be rewritten as needed. With this configuration, even when a word or a sentence that the user does not want to recognize when controlling the control target changes, the present embodiment can be applied.
[0043]
Like the recognition dictionary information 305, the information on the words and sentences stored in the misrecognition countermeasure dictionary information 306 is used by the communication processing unit 304 to update the misrecognition countermeasure dictionary 207 included in the speech recognition apparatus 102 via the network 105. Sent as dictionary information.
[0044]
FIG. 4 is a diagram illustrating an example of a block configuration diagram of the sound collection device 101 according to an embodiment of the present invention. FIG. 4 shows a microphone 401, a voice processing unit 402, and a communication processing unit 403.
[0045]
The microphone 401 converts the surrounding voice into a voice signal and outputs the voice signal to the voice processing unit 402. The audio signal includes an audio signal of a word or a sentence issued by the user to control the output device 103.
[0046]
The voice processing unit 402 corrects distortion from the voice signal obtained from the microphone 401 and removes noise unnecessary for voice recognition via a filter circuit. Thereafter, the signal level of the audio signal output of the signal from which noise has been removed is adjusted, and the adjusted signal is output to the communication processing unit 403.
[0047]
The communication processing unit 403 has a function of transmitting the voice signal output by the voice processing unit 402 to the voice recognition device 102 and the dictionary creation device 104 via the network 105.
[0048]
Next, FIG. 5 shows an example of an operation flow showing a dictionary information creation method of the dictionary creation device 104 according to an embodiment of the present invention.
[0049]
Prior to creating the misrecognition countermeasure dictionary information 306, first, the recognition dictionary information 305 is created. A user or a developer determines a word or a sentence to be used in the voice for controlling the output device 103, and determines what output signal should be output to the output device 103 based on the word or the sentence. When these are determined, the user or the developer inputs information of words, sentences, and output signals to the recognition dictionary information 305 via the recognition dictionary creating unit 303 (step S01).
[0050]
Next, words that are not recognized as operation instructions by the voice recognition device 102 or misrecognized words that are sentences are extracted from the word dictionary 301 (step S02).
[0051]
The words and sentences finally listed in the misrecognition countermeasure dictionary information 306 are extracted in consideration of the following points.
a) Those excluding words and sentences used as operation instructions.
b) A word or sentence that is not an operation instruction but is recognized as an operation instruction by the voice recognition device 102.
c) Because the word length is short, there is little feature information when uttered.
[0052]
a) means that the word or sentence to be recognized is not included in the misrecognition countermeasure dictionary information 306 and the misrecognition countermeasure dictionary 207 regardless of the circumstances.
[0053]
b) If the uttered word or sentence is a word or sentence that is not an operation instruction but is easily recognized as an operation instruction by the voice recognition device 102, these words and sentences are individually extracted. . By doing so, words and sentences that are likely to be erroneously recognized by the voice recognition device 102 can be positively excluded.
[0054]
In c), when a certain word or sentence is uttered, a speech signal that has few features and is difficult to be distinguished from other words or sentences during speech recognition is extracted as misrecognition countermeasure dictionary information. Meaning.
[0055]
In step S02, words and sentences are extracted by focusing on the point c). In the present embodiment, a criterion that there are few feature points included in the audio signal is a word or a sentence described by one to four characters. FIG. 6 shows an example of selecting an extraction candidate from a part of a word or a sentence listed in the word dictionary 301. The table of FIG. 6 shows a column of registered words indicating the character arrangement of the reading of words listed in the word dictionary 301, a character length of each of the registered words, and a column of misrecognition determination which is a determination result of a misrecognized word. ing. Hereinafter, words and sentences that are determined to have few features included in the audio signal are referred to as misidentified words.
[0056]
There are eight registered items that are determined to be misrecognized words: "Asu", "Aichi", "Ai", "Aikagi", "Aiken", "Aokai", and "Aogu". The misrecognition countermeasure dictionary creation unit 302 indicates that the eight words have less vocal characteristics when uttered, and the speech recognition device 102 is likely to erroneously recognize other words or sentences including an operation instruction to the output device 103. Judge as a word. On the other hand, seven words, "I can't accept it", "I don't like it", "Aienka", "Aikido", "Aikiyou", "Aisho" and "Aozare" are judged as words that are hard to be misrecognized. It is not stored in the recognition countermeasure dictionary information 306.
[0057]
When a word or a sentence that is likely to be erroneously recognized as another word or a sentence is simply extracted based on the number of characters as described above, the creation of the erroneous recognition countermeasure dictionary 207 is facilitated. In addition, if a word group or a sentence used by a user who instructs an operation by voice is extracted from the word dictionary 301 in which the words or sentences are listed in a daily manner, the dictionary 207 is a misrecognition countermeasure dictionary 207 covering words and sentences with a high probability of being uttered. It becomes possible.
[0058]
In the present embodiment, the criterion for determining a word or a sentence having a small number of feature points is expressed by one to four characters when read. For example, if the speech recognition device 102 has a high capability of extracting the features of words and sentences included in the speech signal and a high speech recognition rate, the length of the words and sentences to be extracted as the misrecognition countermeasure dictionary information 306 is one to three characters. Alternatively, the number of words or sentences prepared as one or two characters and prepared as an erroneous recognition word can be reduced.
[0059]
Depending on the speech recognition rate required by the user for the speech recognition device 102, the number of words and sentences listed in the misrecognition countermeasure dictionary information 306 can be reduced by reducing the character length corresponding to the misrecognized word. By reducing the number of storages, the storage capacity required for the misrecognition countermeasure dictionary information 306 and the misrecognition countermeasure dictionary 207 can be made smaller.
[0060]
Next, it is determined whether the erroneous recognition countermeasure dictionary creating unit 302 has extracted all erroneous words from the word dictionary 301 (step S03). When all the misrecognized words have been extracted, the information of the words and sentences listed in the recognition dictionary information 305 and the misrecognition countermeasure dictionary information 306 is transmitted to the speech recognition device 102 (step S06).
[0061]
On the other hand, if the word dictionary 301 still contains a misrecognized word, it is determined whether the extracted misrecognized word is included in the recognition dictionary information 305 (step S04). As described above in step S02, this is a process to be performed for a) in which an operation instruction is not included in the misrecognition countermeasure dictionary. At this time, if the extracted misrecognized word is found in the recognition dictionary information 305, the misrecognition countermeasure dictionary creating unit 302 stops the processing of the misrecognized word and extracts a new misrecognized word (step S02).
[0062]
In the determination of the misidentified word in step S04, when the words and sentences are read in terms of the character arrangement and character length, it is determined that the two words are the same if they are the same word or sentence. Thus, the identity of words and sentences is determined simply by whether or not the reading is the same.
[0063]
If it is determined that the misrecognized word is a word or a sentence that is not in the recognition dictionary information 305, the misrecognized word is stored in the misrecognition countermeasure dictionary information 306 together with its reading (step S05). After storing, a new misidentified word is extracted from the word dictionary 301 again (step S02).
[0064]
Next, FIG. 7 shows an operation flow of the speech recognition apparatus 102 according to an embodiment of the present invention. It is assumed that, prior to speech recognition, the recognition dictionary 206 and the erroneous recognition countermeasure dictionary 207 have been normally updated based on the dictionary information transmitted from the dictionary creation device 104.
[0065]
The recognition unit 202 receives the audio signal transmitted by the sound collection device 101, and extracts a word or a sentence from the features of the audio included in the audio signal (Step S11).
[0066]
Next, the recognizing unit 202 determines whether or not the uttered word or sentence included in the audio signal has been normally extracted (step S12). If the result of the determination is that words or sentences cannot be extracted normally, speech recognition processing of the next sent speech signal is performed (step S11).
[0067]
When the words and sentences of the input voice have been successfully extracted, the recognition unit 202 determines whether or not the extracted words and sentences are listed in the misrecognition countermeasure dictionary 207 (step S13). In this determination, the identity of the words and sentences is determined based on whether the character arrangement and the character length of the reading of the words and sentences are the same. If it is listed, the word or sentence included in this audio signal is judged to be a misrecognized word, that is, it is judged that there is a high possibility that it has been erroneously recognized. Processing is performed (step S11).
[0068]
When the recognizing unit 202 determines that the extracted word or sentence is not listed in the misrecognition countermeasure dictionary 207, it determines whether the next extracted word or sentence is listed in the recognition dictionary 206 (step S14). ). If the word or document is not listed in the recognition dictionary 206, the processing is stopped because it is not an operation instruction to the output device 103, and the voice recognition processing of the next transmitted voice signal is performed (step S11).
[0069]
If any of the words or sentences listed in the recognition dictionary 206 includes the same word or sentence as the extracted word or sentence, output signal information associated with the extracted word or sentence is extracted from the recognition dictionary 206. Then, the extracted output signal is output to the output device 103 via the output unit 203 (Step S15).
[0070]
With the above configuration, it is easy to create an erroneous recognition countermeasure dictionary used in the speech recognition system. In addition, since words and sentences commonly used by the user who instructs the operation by voice are extracted from the word dictionary, misidentified words are extracted. Can be. Therefore, it is possible to reduce the cost of creating an erroneous recognition countermeasure dictionary and improve the speech recognition rate.
[0071]
(Modification 1 of Embodiment)
The functions of the sound collection device 101, the voice recognition device 102, and the dictionary creation device 104 shown in FIG. 1 can be realized by one housing. At this time, the misrecognition countermeasure dictionary 207 and the misrecognition countermeasure dictionary information 306, and the recognition dictionary 206 and the recognition dictionary information 305 may be realized by one storage device. The misrecognition countermeasure dictionary creation unit 302 directly updates the misrecognition countermeasure dictionary 207, and the recognition dictionary creation unit 303 similarly updates the recognition dictionary 206 directly.
[0072]
In addition, since updating via the network 105 is unnecessary, in this case, the speech recognition system according to the present invention can be configured without including the communication control units 204, 304, and 403 and the dictionary updating unit 205 in the configuration.
[0073]
In addition, the effects of the present invention can be obtained even when the components shown in FIG. 1 are appropriately combined, such as including the function of the speech recognition device 102 in the output device 103.
[0074]
For example, if the sound collecting devices 101 are installed at a plurality of places in the home and the sound signals transmitted by the sound collecting devices 101 are processed by one voice recognition device 102, an environment in which operation instructions can be given from anywhere in the home can be provided. Can be provided to users. Alternatively, if the functions of the sound collection device 101 and the voice recognition device 102 are housed in a single housing, a portable device having a voice recognition function, for example, a remote control device can be provided.
[0075]
With this configuration, an appropriate system configuration can be obtained according to the usage scene and manufacturing cost of the speech recognition system.
[0076]
(Modification 2 of Embodiment)
After the recognition dictionary information 305 and the misrecognition countermeasure dictionary information 306 created by the dictionary creation device 104 are updated by the dictionary updating unit 205 via the communication processing units 304 and 204, the This is a case where information is added to the recognition dictionary 206 and the misrecognition countermeasure dictionary 207 by the recognition device 102.
[0077]
For example, when it is desired to change a part of a word or a sentence used for an operation instruction, or when a word or a sentence not extracted by the misrecognition countermeasure dictionary creating unit 302 is to be stored in the misrecognition countermeasure dictionary 207 as a misrecognition word, the dictionary is bothersome. This is effective when it is not desired to re-create the dictionary information with the creation device 104, or when the dictionary information cannot be created again because the dictionary creation device 104 is installed in a remote place.
[0078]
FIG. 8 shows an example of a block configuration diagram of a speech recognition device 102 according to the present modification. The difference from the block configuration diagram shown in FIG. 2 is that an input unit 208 is newly added.
[0079]
The input unit 208 is, for example, a keyboard, and information input by the user from the keyboard is output to the dictionary updating unit 205 in order to update information listed in the recognition dictionary 206 and the misrecognition countermeasure dictionary 207. The dictionary updating unit 205 updates the recognition dictionary 206 and the misrecognition countermeasure dictionary 207 based on the information obtained from the input unit 208.
[0080]
Alternatively, the dictionary update unit 205 may obtain voice signals of words and sentences listed in the respective dictionaries from the sound collection device 101 via the network 105 and use the analyzed words and sentences as update dictionary information. good.
[0081]
Note that the configurations of the above-described embodiment and modifications of the present invention are not limited to the configurations described in the embodiments, and the effects of the present invention can be obtained even if some or all of the configurations are replaced with configurations having similar functions. . Further, the effects of the present invention can be similarly obtained by using a computer that implements a program code for realizing the operation flow of the speech recognition device 102 and the dictionary creating device 104 described in the embodiment of the present invention.
[0082]
【The invention's effect】
Speech recognition dictionary that realizes a higher recognition rate by simplifying the selection of words to be stored in the misrecognition countermeasure dictionary required for speech recognition, and by selecting easily misrecognized words and sentences from commonly used words. And a speech recognition apparatus and a speech recognition system having the same.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating an example of a speech recognition system according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating an example of a block configuration diagram of a speech recognition apparatus 102 according to the embodiment of the present invention.
FIG. 3 is a diagram illustrating an example of a block configuration diagram of a dictionary creation device 104 according to the embodiment of the present invention.
FIG. 4 is a diagram illustrating an example of a block configuration diagram of a sound collection device 101 according to the embodiment of the present invention.
FIG. 5 is a diagram showing an example of an operation flow of the dictionary creation device 104 according to the embodiment of the present invention.
FIG. 6 is a diagram showing an example of a method for selecting a mispermitted word in the embodiment of the present invention.
FIG. 7 is a diagram showing an example of an operation flow of the speech recognition device 102 according to the embodiment of the present invention.
FIG. 8 is a diagram illustrating an example of a block diagram of a speech recognition device 102 according to a first modification of the embodiment of the present invention.
[Explanation of symbols]
101 sound collector
102 Speech recognition device
103 Output device
104 dictionary creation device
105 Network
202 Recognition unit
205 Dictionary update unit
206 recognition dictionary
207 Error recognition dictionary
208 Input unit
301 Word Dictionary
302 Error recognition dictionary creation unit
306 Misrecognition countermeasure dictionary information

Claims (8)

A speech recognition dictionary creation method for creating a misrecognition countermeasure dictionary used to determine that the speech recognition result is misrecognition,
From a database that lists everyday terms consisting of words and sentences, extract terms whose reading is less than a predetermined length,
As a result of speech recognition of the input speech, using a speech recognition dictionary listing the predetermined terms that determine that correct speech recognition has been performed when the predetermined terms are obtained,
A voice characterized by selecting, from among the extracted terms, terms having different character arrangements or character lengths of the readings of the terms listed in the speech recognition dictionary, and using the selected terms to create an erroneous recognition countermeasure dictionary. How to create a dictionary for recognition. A speech recognition method using a speech recognition dictionary listing terms that are determined to be correct as the speech recognition result, and an erroneous recognition countermeasure dictionary used to determine that the speech recognition result is erroneous recognition,
From a database that lists everyday terms consisting of words and sentences, extract terms whose reading is less than a predetermined length,
Of the extracted terms, select a term having a different character arrangement or length of the reading of the terms listed in the speech recognition dictionary, and create the misrecognition countermeasure dictionary using the selected term,
When performing voice recognition of input voice,
A speech recognition method characterized in that if the speech recognition result indicates a result indicating the same term as that listed in the misrecognition countermeasure dictionary, the result is excluded from the recognition result. In a speech recognition dictionary creation device for creating a misrecognition countermeasure dictionary used to determine that the speech recognition result is misrecognition,
Selecting means for selecting a term that is determined to be misrecognized and whose reading is equal to or less than a predetermined length and which is likely to be misrecognized in voice recognition, from a database listing daily terms composed of words and sentences,
A dictionary creating means for creating an erroneous recognition countermeasure dictionary using the selected term. 4. The apparatus for creating a speech recognition dictionary according to claim 3, wherein the predetermined length of the reading of the term extracted by the selection unit is four characters or less. In a speech recognition apparatus having a speech recognition dictionary listing terms that are determined to be correct as a speech recognition result, and an erroneous recognition countermeasure dictionary used to determine that the speech recognition result is erroneous recognition,
A database that lists everyday terms consisting of words and sentences,
From the terms listed in this database, terms whose readings of terms are equal to or less than a predetermined length are extracted, and among the extracted terms, a term having a different character arrangement or character length of the reading of the terms listed in the speech recognition dictionary is selected. Means for selecting,
Dictionary creation means for creating the misrecognition countermeasure dictionary using the selected term,
A speech recognition apparatus characterized in that if the result of speech recognition of an input speech is a result indicating the same term as that listed in the misrecognition countermeasure dictionary, the result is excluded from the recognition result. A speech recognition program that performs speech recognition using a speech recognition dictionary that lists terms that are determined to be correct as the speech recognition result and a misrecognition countermeasure dictionary used to determine that the speech recognition result is erroneous recognition,
Searching a database listing daily terms consisting of words and sentences, and extracting terms whose readings of terms are equal to or less than a predetermined length;
Search the voice recognition dictionary, determine whether there is a term with the same character sequence and character length of the reading of the extracted term, and only when there is no same term, the extracted term is in the misrecognition countermeasure dictionary Adding and storing,
When performing speech recognition of the input speech, if the word indicated by the speech recognition result is the same as the term stored in the misrecognition countermeasure dictionary, the result is excluded from the recognition result. Recognition program. From a database that lists daily terms consisting of words and sentences, selectively extract words that should be judged as misrecognition and have a reading of less than a predetermined length and that are easily misrecognized in speech recognition. A dictionary creation device for speech recognition that creates the misrecognition countermeasure dictionary using:
In a speech recognition system comprising a speech recognition device for recognizing speech using a speech recognition dictionary for input speech,
The voice recognition device, for the recognition result of the input voice, comprising means for searching the erroneous recognition countermeasure dictionary whether it corresponds to a term in the erroneous recognition countermeasure dictionary,
In the case of a speech recognition result corresponding to a term in the misrecognition countermeasure dictionary, the term is not output as a recognition result. The speech recognition dictionary creation device and the speech recognition device include a communication processing unit for communicating with each other,
The speech recognition system according to claim 7, wherein the communication processing unit transfers the misrecognition countermeasure dictionary created by the speech recognition dictionary creation device to the speech recognition device.

Images