JP3825526B2 - Voice recognition device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition apparatus that recognizes an input speech by comparing a feature parameter series obtained by acoustic analysis of the input speech with a keyword speech model that constitutes each recognition vocabulary prepared in advance. The present invention relates to a speech recognition apparatus suitable for reducing recognition errors in recognition vocabulary readings and misrecognition due to reading memory differences when the apparatus is used.
[0002]
[Prior art]
In general, in a speech recognition device that recognizes input speech, by inputting a reading of a vocabulary (recognition vocabulary) to be recognized by the device, a speech model of a keyword constituting a corresponding recognition vocabulary is preliminarily obtained from the reading. Created and stored for recognition of input speech. Recognition of input speech by this type of speech recognition apparatus is performed as follows.
[0003]
First, the input speech is acoustically analyzed to obtain a feature parameter series. Next, the input speech is recognized by comparing the obtained feature parameter series of the input speech with a speech model of a keyword constituting each recognition vocabulary prepared in advance.
[0004]
Conventionally, such a speech recognition apparatus has a problem that if a reading of a recognized vocabulary is mistakenly registered, it cannot be correctly recognized even if the correct reading is uttered during use. In addition, there is a problem that even if a correct reading is registered at the time of registration of the recognition vocabulary, it is not correctly recognized if an incorrect reading is made at the time of use.
[0005]
[Problems to be solved by the invention]
As described above, in the conventional speech recognition device, if the reading of the recognized vocabulary is registered incorrectly, it cannot be recognized correctly even if the correct reading is uttered at the time of use, and conversely, the correct reading is registered at the time of registering the recognized vocabulary. However, there was a problem that it would not be recognized correctly if a wrong reading was made during use.
[0006]
The present invention has been made in consideration of the above circumstances, and its purpose is to make a mistake in registering the reading of the recognized vocabulary and the memory of reading when using the device., Wrong readingIt is an object of the present invention to provide a speech recognition apparatus that can prevent a reduction in recognition performance caused by the above.
[0012]
[Means for Solving the Problems]
  Of the present inventionFrom one perspective,A speech recognition device that recognizes an input speech by collating with a speech model of each recognition vocabulary including a speech model of another name of the recognition vocabularyIs provided. This voice recognition deviceOf the recognized vocabulary to registerFirstIn addition to readingSecondAnother name registration means to register the reading as another name, and registered as another nameSecondReadingRegisteredRecognition vocabularyFirstA registered vocabulary similarity determining unit that determines whether or not the reading is similar, and a warning output unit that outputs a warning to that effect when the registered vocabulary similarity determining unit determines that the reading is similar, By the above nickname registration meansAs another nameRegisteredSecondIf the input speech is recognized by matching the reading speech model,SecondIt has a recognition result output means for recognizing a vocabulary whose reading is another name.Here, when the recognition result output means recognizes the second reading registered as another name, the first reading of the vocabulary nicknamed the second reading is used by the recognition result output means by the user. The configuration presented in
[0013]
  In such a configuration, the recognition vocabulary registered by inputting from the recognition vocabulary input means is stored.The reading (second reading) is different from the reading (first reading).When registering with another name registration means,thisAliasSecond reading registered asOther vocabularyFirst readingSimilar toHaveIn the registered vocabulary similarity judgment meansIf it is judged and similar,By notifying the user from the warning output means, it is possible to prevent a reduction in recognition performance due to nickname registration. At this time, it is preferable to cancel the corresponding alias registration.When the input speech is recognized by collation with the second reading speech model registered as another name, for example, “Kobe” is recognized by recognizing the vocabulary with the second reading as another name. Even if the speaker utters “Kobe”, which is a general reading, with “Kobe” registered as another name, the correct “Kobe” is obtained as a recognition result. be able to. At this time, by presenting to the speaker the reading of the vocabulary “Kobe” (first reading) “Kobe”, which is nicknamed “Kobe” (second reading), This makes it possible for the speaker to remember the correct reading of the word that was remembered under another name, and to input the correct reading thereafter.
[0020]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings.
[First Embodiment]
FIG. 1 is a block diagram showing a schematic configuration of a speech recognition apparatus according to the first embodiment of the present invention.
[0021]
In the speech recognition apparatus of FIG. 1, speech input from the speech input unit 101 is converted into feature parameters by the acoustic analysis unit 102. As typical feature parameters used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient obtained by LPC (linear prediction) analysis, etc. are often used. The type of feature parameter does not matter.
[0022]
The acoustic analysis unit 102 outputs the obtained feature parameter series of the input speech to the model matching unit 103. The model matching unit 103 generates each keyword (keyword constituting the recognition vocabulary) to be recognized, which is created in advance according to the recognition vocabulary input from the recognition vocabulary input unit 105 and stored in the speech model creation / storage unit 104. Calculation is performed to obtain the similarity or distance between the speech model and the feature parameter series of the input speech.
[0023]
As a matching method of the model matching unit 103, a voice model is also expressed by a feature parameter series, and a method of obtaining a distance between the feature parameter series of the voice model and the feature parameter series of the input voice by a DP (dynamic programming) method, A method of expressing a speech model using an HMM (Hidden Markov Model) and calculating the probability of each speech model when a feature parameter sequence of the input speech is input is widely used. It doesn't matter.
[0024]
The recognition vocabulary input unit 105 includes a recognition vocabulary, division information for each keyword of the recognition vocabulary (keyword division information), and reading information for each keyword necessary for creating a speech model of each keyword constituting the recognition vocabulary. Can be realized with a keyboard or a file. The recognized vocabulary input from the recognized vocabulary input unit 105 is registered in the recognized vocabulary storage unit 109.
[0025]
The keyword conversion unit 106 extracts each keyword from the keyword division information input from the recognized vocabulary input unit 105, and is selected based on the sound similarity between keywords and attributes such as part of speech. A keyword conversion table (keyword conversion table) 106a having the same part of speech is created and stored.
[0026]
The keyword expansion unit 107 causes the keyword conversion unit 106 to convert the keyword (with similarity or distance) obtained by the model matching unit 103 into another keyword having sound similarity according to the keyword conversion table 106a. Perform expansion.
[0027]
The recognition result output unit 108 sets a similarity (or distance) to a keyword string in which each voice model obtained by the model matching unit 103 and a keyword obtained by expansion by the keyword expansion unit 107 are combined with a certain condition (for example, similarity) Of the recognition vocabulary stored in the recognition vocabulary storage unit 109 and outputs the recognition target category having the maximum similarity (or the minimum distance) as a recognition result. . Note that, as the sorting restriction, for example, a person's name may be given priority over a company name.
[0028]
The specific operation of the speech recognition apparatus having the configuration shown in FIG. 1 described above is as follows. The vocabulary to be recognized by the speech recognition apparatus, that is, the recognition vocabulary is “Sato Shoten”, “Katan Shokudo”, “Tanaka Shoten” A case where there are three types will be described as an example.
[0029]
In this case, the above three types of recognition vocabulary are input from the recognition vocabulary input unit 105. In this embodiment, the speech model of each keyword constituting the recognition vocabulary is stored in the speech model creation / storage unit 104. For example, “Sato-Shop”, “Katani-Shokudo”, “Tanaka-Bookstore”, and so on, can be created by entering the keyword division symbol (keyword division information) “-” in the recognition vocabulary. Is done.
[0030]
The speech model creation / storage unit 104 receives the recognition vocabulary “Sato-shop”, “Kato-shokudo”, and “Tanaka-bookstore” with the keyword division symbol “-” inserted from the recognition vocabulary input unit 105. Each recognized vocabulary is divided into keywords constituting the vocabulary according to the keyword division symbol “-” inserted in the vocabulary. Here, the above three types of recognition vocabulary are divided into six keywords: “Sato”, “Katan”, “Tanaka”, “Shop”, “Restaurants”, and “Bookstore”. The voice model creation / storage unit 104 creates and stores voice models for the six keywords “Sato”, “Katan”, “Tanaka”, “Store”, “Dining room”, and “Bookstore”.
[0031]
At the same time, the keyword conversion unit 106 receives the above six words obtained from the recognition vocabulary “Sato-shop”, “Katan-shokudo”, and “Tanaka-bookstore” with keyword division symbols inputted from the recognition vocabulary input unit 105. The keywords “Sato”, “Katan”, “Tanaka”, “Store”, “Restaurants”, and “Bookstore” are examined for voice similarity, similar keywords are extracted, and a keyword conversion table 106a is created. . Here, as a rule for creating the keyword conversion table 106a, keywords having different syllable numbers with different keyword readings are registered in the table 106a as having similar sounds, for example, keywords having one syllable or less. In this case, “Sato” and “Kato”, “shop” and “bookstore” are extracted as similar, and a keyword conversion table 106a as shown in FIG. 2 is created.
[0032]
Then, at the time of speech recognition, for example, when “Sato Shoten” is input, and the matching results in the model matching unit 103 are “Sato” and “Bookstore”, the keyword expansion unit 107 sets “Sato” and “ For the “bookstore”, the keyword conversion unit 106 performs keyword conversion of “Sato → Katan” and “bookstore → shop” according to the keyword conversion table 106 a, and the matching result of the model matching unit 103 is “Sato” and “bookstore”. In addition, the keywords are expanded as if there are “katan” and “shop”.
[0033]
Among the combinations of keywords expanded by the keyword expansion unit 107, “Sato Shoten” is one that matches the recognized vocabulary stored in the recognized vocabulary storage unit 109. Therefore, the recognition result output unit 108 can correctly recognize and output “Sato Shoten” even though the matching results in the model matching unit 103 are “Sato” and “Bookstore”.
[0034]
On the other hand, when there is no keyword conversion unit 106 and keyword expansion unit 107 and keyword expansion is not performed, “Sato” composed of “Sato” and “bookstore” which are the collation results in the model collation unit 103 Since “bookstore” does not exist in the recognized vocabulary storage unit 109, “Sato Shoten” cannot be correctly recognized.
[0035]
It should be noted that the similarity may be reduced by a certain value or a certain rate in the collation with the speech model of the keyword obtained by keyword conversion.
As described above, even when the speaker utters “Sato Shoten” but the model matching unit 103 obtains an incorrect matching result of “Sato” and “Bookstore”, the keyword The example that can recognize "Sato Shoten" correctly by extension was described. Similarly, in the present embodiment, even when the speaker mistakenly says “Sato Shoten” as “Sato Shoten”, it is possible to correctly recognize “Sato Shoten” by extending the keyword focusing on sound similarity. it can.
[0036]
As described above, in the present embodiment, the keyword is expanded by paying attention to the similarity as a sound, so that it is possible to effectively prevent a decrease in recognition performance due to a keyword recognition error or a speaker's error.
[Second Embodiment]
FIG. 3 is a block diagram showing a schematic configuration of a speech recognition apparatus according to the second embodiment of the present invention.
[0037]
In the speech recognition apparatus of FIG. 3, the speech input from the speech input unit 201 is converted into feature parameters by the acoustic analysis unit 202. As typical feature parameters used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient obtained by LPC (linear prediction) analysis, etc. are often used. The type of feature parameter does not matter.
[0038]
The acoustic analysis unit 202 outputs the obtained feature parameter series of the input speech to the model matching unit 203. The model collation unit 203 performs a calculation to obtain the similarity or distance between the speech model of the syllable string of any length stored in the speech model storage unit 204 and the feature parameter series.
[0039]
As a matching method of the model matching unit 203, a voice model is also expressed by a feature parameter series, and a method of obtaining a distance between the feature parameter series of the voice model and the feature parameter series of the input voice by a DP (dynamic programming) method, A method of expressing a speech model using an HMM (Hidden Markov Model) and calculating the probability of each speech model when a feature parameter sequence of the input speech is input is widely used. It doesn't matter.
[0040]
Based on the collation result in the model collation unit 203, the recognition result output unit 205 is similar to, for example, a syllable string in which the first syllable and the last syllable match according to the constraint condition stored in the constraint condition storage unit 206. The degree (or distance) is sorted under a certain condition, and the leading syllable of the syllable string having the maximum similarity (or the minimum distance) is output as a recognition result.
[0041]
The specific operation of the speech recognition apparatus having the configuration shown in FIG. 3 described above will be described by taking, for example, the case of saying “Asahi (Asahi)” as an example.
First, as a result of the speaker uttering “Asahi (Asahi no Aa)”, the model collation unit 203 uses a syllable string “A” “SA” having a similarity to the syllable string as shown in FIG. It is assumed that “hi”, “no”, “a” and syllable strings “u”, “sa”, “hi”, “no”, “a” having a similarity of 92 are obtained.
[0042]
In this case, if the first syllable of the syllable string having the maximum similarity is output as the recognition result from the recognition result output unit 205 without providing the restriction that the first syllable and the last syllable coincide with each other, An incorrect syllable “U” different from the first syllable of the input speech is output.
[0043]
On the other hand, in the present embodiment, the restriction that the first and last syllables coincide with each other due to the restriction condition stored in the restriction condition storage unit 206. Therefore, the recognition result in the recognition result output unit 205 is the syllable. It becomes “A” and the leading syllable of the input speech can be recognized correctly. Moreover, since the first syllable of the input speech is recognized under the constraint that the first and last syllables match, the speech uttered when inputting the syllable `` a '' as in this example is `` Asahi The first syllable and the last syllable need only be the same, such as not only “Noah” but also “Ahirunoa” and “Ajia (Asia)”.
[0044]
As described above, in this embodiment, it is possible to realize highly accurate syllable recognition by recognizing the leading syllable of the input speech under the restriction that the leading syllable and the last syllable of the input speech match. . Moreover, since it is not necessary to memorize the words uttered when inputting each syllable, anyone can use it immediately.
[Third Embodiment]
FIG. 5 is a block diagram showing a schematic configuration of a speech recognition apparatus according to the third embodiment of the present invention.
[0045]
In the speech recognition apparatus of FIG. 5, the speech input from the speech input unit 301 is converted into feature parameters by the acoustic analysis unit 302. As typical feature parameters used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient obtained by LPC (linear prediction) analysis, etc. are often used. The type of feature parameter does not matter.
[0046]
The acoustic analysis unit 302 outputs the obtained feature parameter series of the input speech to the model matching unit 203. The model collation unit 303 is configured according to the recognition vocabulary input from the recognition vocabulary input unit 305 and stored in the speech model creation / storage unit 304 as a recognition target keyword and the characteristics of the input speech. Calculation to obtain similarity or distance to the parameter series.
[0047]
As a matching method of the model matching unit 303, a speech model is also expressed by a feature parameter sequence, and a method of obtaining a distance between the feature parameter sequence of the speech model and the feature parameter sequence of the input speech by a DP (dynamic programming) method, A method of expressing a speech model using an HMM (Hidden Markov Model) and calculating the probability of each speech model when a feature parameter sequence of the input speech is input is widely used. It doesn't matter.
[0048]
The recognition result output unit 308 outputs a vocabulary having the maximum similarity (or the minimum distance) for each recognized vocabulary obtained by the model matching unit 303 as a recognition result.
The recognition vocabulary input unit 305 is for registering a vocabulary to be recognized and its reading, and can be realized by a keyboard or a file.
[0049]
On the other hand, for example, when registering “Kobe” from the recognition vocabulary input unit 305 as a place name to be registered, “Kobe” is generally used as a reading of “Kobe” and is misread as “Kobe”. If it is determined that there is a high possibility of misreading, “Kobe” is easily misread, that is, “Kobe” is registered as an alternative name from the alternative name registration unit 306. The nickname registration unit 306 may be configured with a keyboard, for example, so that the nickname may be registered (input) by the user's operation of the keyboard, or a table in which the vocabulary can be retrieved. May be created in advance and provided in the alias registration unit 306, and when multiple readings exist, the alias may be automatically generated (input) in the alias registration unit 306 from the table.
[0050]
The registered vocabulary similarity determination unit 307 determines the similarity between the readings of all the vocabularies registered from the recognized vocabulary input unit 305 and the readings registered by the alias registration in the alias registration unit 306. If there is a vocabulary similar to a nickname-registered reading (a vocabulary is registered), the registered vocabulary similarity determination unit 307 warns the user by the warning output unit 309, Cancel registration. For the determination of the similarity of reading in the registered vocabulary similarity determination unit 307, for example, a condition that the difference in reading syllables is one syllable or less can be applied.
[0051]
As described above, in the present embodiment, when the alias of the recognized vocabulary registered by inputting from the recognized vocabulary input unit 305 is registered by the alias registration unit 306, whether the alias is similar to other vocabularies is registered. By determining the similarity in the similarity determination unit 307 and informing the user from the warning output unit 309, it is possible to prevent a reduction in recognition performance due to nickname registration. For example, when “Kobe” is registered as an alternative term for “Kobe” even though “Kobe” is registered as a recognition vocabulary instead of “Kobe” as described above. “Kobe” and “Kobe” cannot be distinguished from each other, but the speech recognition apparatus shown in FIG. 5 can avoid such a problem.
[0052]
Note that if there is no similar recognition vocabulary to the nickname registered by the nickname registration unit 306, for example, when “Kobe” is registered as another name of the recognition vocabulary “Kobe”, the registered vocabulary similarity determination unit 307 is registered. It is judged that there is no similar vocabulary, and nickname registration is permitted, and the speech model creation / storage unit 304 stores the recognition vocabulary “Kanbe” separately from the speech model of the recognition vocabulary “Kobe”. A speech model of another name “Kobe” is stored. In this case, even if “Kobe” is mistakenly uttered as “Kobe”, the model matching unit 303 (in the speech model creation / storage unit 304) is also called “Kobe” (in the speech model creation / storage unit 304). "Kobe" is recognized by collating with the voice model.
[Fourth Embodiment]
FIG. 6 is a block diagram showing a schematic configuration of a speech recognition apparatus according to the fourth embodiment of the present invention.
[0053]
In the speech recognition apparatus in FIG. 6, the speech input from the speech input unit 401 is converted into a characteristic parameter by the acoustic analysis unit 402. As typical feature parameters used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient obtained by LPC (linear prediction) analysis, etc. are often used. The type of feature parameter does not matter.
[0054]
The acoustic analysis unit 402 outputs the obtained feature parameter series of the input speech to the model matching unit 403. The model matching unit 403 recognizes each of the keywords to be recognized (keywords constituting the recognition vocabulary) created in advance according to the recognition vocabulary input from the recognition vocabulary input unit 405 and stored in the speech model creation / storage unit 404. Calculation is performed to obtain the similarity or distance between the speech model and the feature parameter series of the input speech.
[0055]
As a matching method of the model matching unit 403, a speech model is also expressed by a feature parameter sequence, and a method of obtaining a distance between the feature parameter sequence of the speech model and the feature parameter sequence of the input speech by a DP (dynamic programming) method, A method of expressing a speech model using an HMM (Hidden Markov Model) and calculating the probability of each speech model when a feature parameter sequence of the input speech is input is widely used. It doesn't matter.
[0056]
The keyword-specific recognition result output unit 406 semantically inputs the vocabulary input from the recognition vocabulary input unit 405 and stored in the recognition vocabulary storage unit 407 according to the similarity (or distance) to each keyword obtained by the model matching unit 403. Are sorted by the same keyword, and a plurality of vocabularies having the maximum similarity (or the minimum distance) are output as recognition results.
[0057]
For example, in the recognition vocabulary storage unit 407, six types of recognition vocabulary, “Tanaka Hotel”, “Sato Hotel”, “Katan Hotel”, “Tanaka Hotel”, “Sato Hotel”, and “Tanaka Gas Station” are registered. In this case, as keywords for speech recognition, six keywords “Tanaka”, “Sato”, “Kayato”, “Hotel”, “Shuzhou”, and “Gas station” are considered.
[0058]
Here, if “Tanaka Hotel” is uttered, the recognition result and similarity obtained by the model matching unit 403 are assumed to be as shown in FIG. In this case, as shown in FIG. 8B, the keyword-specific recognition result output unit 406 is similar to a keyword string in which keywords are simply combined (matches the vocabulary stored in the recognized vocabulary storage unit 407). If a plurality of candidates are output in descending order of the sum (conventional output method), there is a problem that it is difficult to understand when selecting candidates because “hotels” and “hotels” are mixed.
[0059]
On the other hand, in this embodiment, the keyword-specific recognition result output unit 406 outputs, for example, a keyword string in which the similarity of a keyword representing a business type is equal to or greater than a predetermined threshold for each keyword representing the business type. For example, if you display by keyword that represents an industry with a similarity of 100 or more (here, “hotel” and “hotel”)FIG. 8 (a)The candidates can be displayed with high visibility.
[0060]
Thus, in this embodiment, candidate selection can be performed efficiently by outputting a plurality of recognition results for each keyword in descending order of similarity (or in ascending order of distance).
[Fifth Embodiment]
FIG. 9 is a block diagram showing a schematic configuration of a speech recognition apparatus according to the fifth embodiment of the present invention.
[0061]
In the speech recognition apparatus of FIG. 9, the speech input from the speech input unit 501 is converted into feature parameters by the acoustic analysis unit 502. As typical feature parameters used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient obtained by LPC (linear prediction) analysis, etc. are often used. The type of feature parameter does not matter.
[0062]
The acoustic analysis unit 502 outputs the obtained feature parameter series of the input speech to the model matching unit 503. The model matching unit 503 generates each speech model of the keyword to be recognized (here, the recognized vocabulary) created in advance according to the recognized vocabulary input from the recognized vocabulary input unit 505 and stored in the speech model creation / storage unit 504. (Speech model of each recognized vocabulary including a speech model of another name) and a feature parameter series of the input speech and a calculation for obtaining a similarity or distance.
[0063]
As a matching method of the model matching unit 503, a method of obtaining a distance between a feature parameter sequence of a speech model and a feature parameter sequence of an input speech by a DP (dynamic programming) method by expressing the speech model as a feature parameter sequence, A method of expressing a speech model using an HMM (Hidden Markov Model) and calculating the probability of each speech model when a feature parameter sequence of the input speech is input is widely used. It doesn't matter.
[0064]
The recognition vocabulary input unit 505 is for registering a vocabulary to be recognized and its reading, and can be realized by a keyboard or a file.
On the other hand, for example, when registering “Kobe” from the recognition vocabulary input unit 505 as a place name to be registered, “Kobe” is more commonly read as “Kobe” and is misread as “Kobe”. If it is determined that there is a high possibility of misreading, “Kobe” is easily misread, that is, “Kobe” is registered as an alternative name from the alternative name registration unit 506. The nickname registration unit 506 may be configured with a keyboard, for example, so that the nickname may be registered (input) by the user's operation of the keyboard, or a table that can retrieve the reading from the vocabulary. May be created in advance and provided in the alias registration unit 506, and when there are a plurality of readings, an alternative name may be automatically generated (input) from the table within the alias registration unit 506.
[0065]
When “Kobe” is registered as another name of the recognition vocabulary “Kobe” from the nickname registration unit 506, the recognition vocabulary is stored in the speech model creation / storage unit 504 separately from the speech model of the recognition vocabulary “Kobe”. The voice model of “Kobe”, another name for “Kobe”, is stored. Here, flag information indicating that it is another name for “Kobe” is attached to the speech model of “Kobe”.
[0066]
Therefore, even if “Kobe” is mistakenly uttered as “Kobe”, the model collation unit 503 (in the speech model creation / storage unit 504) utters “Kobe”, another name for “Kobe”. By collating with the model, “Kobe” is recognized. Here, as described above, flag information indicating that it is another name for “Kobe” is added to the speech model of “Kobe”. When the collation is performed, the flag information is attached to the collation result and is passed to the recognition result output unit 507. As a result, the recognition result output unit 507 identifies that the keyword recognized by the model matching unit 503 is not a correct reading but another name, adds a correct reading “Kanbe” to the recognition result “Kobe”, and displays or Output by voice.
[0067]
As described above, in this embodiment, when a recognition result is output, by outputting a correct reading, it is possible to cause the speaker to remember a correct reading of a word that the speaker remembered by another name. become able to.
[Sixth Embodiment]
FIG. 10 is a block diagram showing a schematic configuration of a speech recognition apparatus according to the sixth embodiment of the present invention.
[0068]
In the speech recognition apparatus of FIG. 10, the speech input from the speech input unit 601 is converted into feature parameters by the acoustic analysis unit 602. As typical feature parameters used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient obtained by LPC (linear prediction) analysis, etc. are often used. The type of feature parameter does not matter.
[0069]
The acoustic analysis unit 602 outputs the obtained feature parameter series of the input speech to the model matching unit 603. The model matching unit 603 is created in advance according to the reading of the recognized vocabulary input from the recognized vocabulary input unit 605 (the reading of the recognized vocabulary input in kana, katakana or romaji), and the speech model creation / storage unit 604. The calculation for obtaining the similarity or distance between each speech model of the keyword to be recognized (keyword constituting the recognition vocabulary) stored in the above and the feature parameter series of the input speech is performed.
[0070]
As a matching method of the model matching unit 603, a voice model is also expressed by a feature parameter sequence, and a distance between the feature parameter sequence of the speech model and the feature parameter sequence of the input speech is obtained by a DP (dynamic programming) method, A method of expressing a speech model using an HMM (Hidden Markov Model) and calculating the probability of each speech model when a feature parameter sequence of the input speech is input is widely used. It doesn't matter.
[0071]
When the recognition vocabulary reading is input from the recognition vocabulary input unit 605, the voice output unit 607 converts the reading into speech and outputs it. The voice output by the voice output unit 607 can be realized, for example, by storing voices for all Japanese syllables and connecting the stored voices in accordance with the input reading.
[0072]
For example, if “Takeshipa” is mistakenly input from the recognition vocabulary input unit 605 as a reading of the recognition vocabulary “Takeshiba”, a speech model of the erroneous reading “Takeshipa” is created and stored in the speech model creation / storage unit 604. Since it is memorized, even if “Takeshiba” is spoken during recognition, it cannot be recognized correctly.
[0073]
On the other hand, according to the present embodiment, when registering the reading of the recognition vocabulary “Takeshiba” from the recognition vocabulary input unit 605, if “Takeshipa” is mistakenly input, the voice output unit 607 reads “Takeshipa”. Since the voice is output, the speaker (recognized vocabulary registrant) can easily recognize an input error in reading, and can prevent deterioration in recognition performance due to an input error in reading.
[0074]
1, 3, 5, 6, 9, and 10 described above, the functions of each part of the speech recognition apparatus have a computer, for example, a built-in microphone or a microphone input terminal. A recording medium such as a CD-ROM, a floppy disk, or a memory card on which a program for causing a personal computer having a voice input function to function as each processing unit of the voice recognition device is recorded. This is realized by mounting on a personal computer and reading and executing the program recorded on the recording medium by the personal computer. The program is not limited to a recording medium, and may be downloaded from a communication line, for example.
[0075]
As detailed above, according to the present invention,It is possible to prevent the recognition performance from being deteriorated due to a registration error in the recognition vocabulary reading, a memory difference in reading when the device is used, a utterance in the wrong reading, or the like.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a first embodiment of the present invention.
FIG. 2 is a view showing an example of contents of a keyword conversion table 106a in FIG.
FIG. 3 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a second embodiment of the present invention.
4 is a diagram showing an example of a syllable string and similarity obtained by the model matching unit 203 in FIG. 3 when “Asahi no A” is uttered. FIG.
FIG. 5 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a third embodiment of the present invention.
FIG. 6 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a fourth embodiment of the present invention.
FIG. 7 is a diagram showing an example of each keyword and similarity obtained by the model matching unit 403 in FIG. 6 when “Tanaka Hotel” is uttered.
FIG. 8 is a diagram showing a recognition result display example based on the recognition result and similarity in FIG. 7 in comparison with a conventional recognition result display example.
FIG. 9 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a fifth embodiment of the present invention.
FIG. 10 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a sixth embodiment of the present invention.
[Explanation of symbols]
101, 201, 301, 401, 501, 601 ... voice input unit
102, 202, 302, 402, 502, 602 ... acoustic analysis unit
103, 203, 303, 403, 503, 603 ... model matching unit
104, 304, 404, 504, 604 ... voice model creation / storage unit
105,305,405,505,605 ... Recognition vocabulary input part
106: Keyword conversion unit
107 ... Keyword expansion part
108, 205, 308, 507, 606... Recognition result output unit
109,407 ... Recognition vocabulary storage unit
204 ... Voice model storage unit
206: Constraint storage unit
306, 506 ... Alternative name registration unit
307: Registered vocabulary similarity determination unit
309: Warning output section
406 ... Keyword-specific recognition result output section
607 ... Audio output unit

Claims (3)

A speech recognition device that recognizes an input speech by collating with a speech model of each recognition vocabulary including a speech model of another name of a recognition vocabulary,
In addition to the first reading that is the original reading of the recognized vocabulary to be registered, another name registration means for registering a second reading that is a different reading from the original reading of the recognized vocabulary as a different name,
Registered vocabulary similarity determining means for determining whether or not the second reading registered as an alternative name by the alternative name registration means is similar to the first reading of the registered recognized vocabulary;
A warning output means for outputting a warning to that effect when it is determined that the registered vocabulary similarity determination means is similar;
A recognition result output means for recognizing a vocabulary having the second reading as a different name when the input voice is recognized by collation with the second reading voice model registered as a different name by the different name registration means; A speech recognition apparatus characterized by that.   The recognition result output means, when recognizing a second reading registered as a different name, presents the user with a first reading of a vocabulary having the second reading as a different name. The speech recognition apparatus according to Item 1. Apply to the speech recognition device that recognizes the input speech by collating it with the speech model of each recognition vocabulary including the speech model of another name of the recognition vocabulary that has been created beforehand by the feature parameter series obtained by acoustic analysis of the input speech Speech recognition method,
When the speech model of the recognized vocabulary is created in advance, in addition to the first reading that is the original reading of the recognized vocabulary, the second reading that is different from the original reading of the recognized vocabulary As another name,
When registering the second reading as the alias, determine whether the second reading is similar to the first reading of the registered recognition vocabulary,
If it is determined that they are similar, a warning indicating an alias registration error will be output,
If the input speech is recognized by collation with the second reading speech model registered as an alias,
A speech recognition method characterized by recognizing a vocabulary nicknamed the second reading.

Images