JP4769223B2 - Text phonetic symbol conversion dictionary creation device, recognition vocabulary dictionary creation device, and speech recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a text phonetic symbol conversion dictionary creator which creates a text phonetic symbol conversion dictionary for accurately creating phonetic symbol strings having high possibility of being pronounced when words are pronounced. <P>SOLUTION: The text phonetic symbol conversion dictionary creator 100 includes: acquiring the words, segment division information for dividing the words into segments and phonetic symbols of each segment from learning data; calculating occurrence probability of phoneme segment pairs in the learning data and connection probability of connection phoneme segment pair system; and creating the text phonetic symbol conversion dictionary including these calculated probability. <P>COPYRIGHT: (C)2009,JPO&amp;INPIT

Description

  The present invention relates to a text phonetic symbol conversion dictionary creation device, a text phonetic symbol conversion dictionary creation program, a text phonetic symbol conversion dictionary creation method, a recognition vocabulary dictionary creation device using the created text phonetic symbol conversion dictionary, a recognition vocabulary dictionary creation program, The present invention relates to a recognition vocabulary dictionary creation method, a speech recognition device, a speech recognition program, and a speech recognition method.

  In a speech synthesizer that converts input (text) of an arbitrary word or sentence into speech and outputs it, spelling-phonetic symbol conversion is performed to convert the sentence input into a phonetic symbol string. As a conventional spelling-phonetic symbol conversion method, there is a method based on a word dictionary. For example, each character string such as a word may be associated with a phonetic symbol string and stored as a word dictionary. An example of a word dictionary is shown in FIG. As shown in the figure, the word dictionary can search for the word “abaca” and specify the phonetic symbol string / abaka / corresponding to it. In this specification, a character surrounded by / is used hereinafter as a notation of a phonetic symbol.

As a second conventional spelling-phonetic symbol conversion method, there is a rule-based method. For example, rules regarding the arrangement of characters are used. As an example of a rule, as shown in FIG. 11, “a” at the beginning of a character string is pronounced as “/ a /”, and “a” when a character string is arranged as “abe” is pronounced as “/ e /”. Is stipulated. Further, it is defined that “a” when “aben” and a character string are arranged is pronounced as “/ o /”.
Non-patent document 1 describes a technique called Graphe To Phoneme (G2P) that converts written characters into a minimum unit (phoneme) of speech as a conventional technique of these spelling-phonetic symbol conversion methods.

An Introduction to Text-to-Speech Synthesis: (KLUWER ACADWMIC PABLISHERS: by Thierry Dutoit)

The spelling-phonetic symbol conversion method may be used not only in a speech synthesizer but also in a speech recognition device. As an example of such a voice recognition device, there is a mobile phone or the like having a function that allows a user to select a destination name registered in phone book data by voice recognition. In order to use the other party name as the recognition target vocabulary, it is necessary to generate a phonetic symbol string from the spelling of the other party name.
In the spelling-phonetic symbol conversion method based on the word dictionary described above, a phonetic symbol string cannot be obtained if a word that matches the recognition target vocabulary is not registered in the word dictionary. In addition, in order to obtain a phonetic symbol string for many words, it is necessary to increase the number of words registered in the word dictionary, resulting in a problem that the word dictionary is enlarged.

On the other hand, in the spelling-phonetic symbol conversion method based on the rules described above, some phonetic symbols can be obtained for any input word, but in order to increase the accuracy of the output phonetic symbol string, complicated rules are used. It is necessary to use this, and the storage area for storing this rule increases. In addition, since the rules are complicated, there is a problem in that the amount of processing for outputting a phonetic symbol string from an input word also increases.
Furthermore, since there is no index indicating the certainty of the phonetic symbol string output for the input word, a predetermined number of phonetic symbol strings from among a plurality of phonetic symbol string candidates obtained for one word There is also a problem that it is difficult to select the most probable phonetic symbol string.

  The present invention has been made in view of these points, and creates a text phonetic symbol conversion dictionary for accurately generating a phonetic symbol string that is likely to be pronounced when a word is pronounced. Provided are a text phonetic symbol conversion dictionary device, a recognition vocabulary dictionary creation device that creates a recognition vocabulary dictionary using the text phonetic symbol conversion dictionary, and a speech recognition device that recognizes speech at a high recognition rate using the recognition vocabulary dictionary. For the purpose. Another object is to output an index indicating the certainty along with the phonetic symbol string.

In order to solve the above problems, a text phonetic symbol conversion dictionary creating apparatus according to claim 1 of the present invention creates a text phonetic symbol conversion dictionary for creating a text phonetic symbol conversion dictionary used for text-to-phonetic symbol conversion. A learning data storage means for storing data including a word , segment division information obtained by dividing the word into segments, and a phonetic symbol for each segment as learning data; from the learning data, the word Learning data acquisition means for acquiring the segment division information and the phonetic symbol for each segment, the segment division information acquired by the learning data acquisition means, and the phonetic symbol for each segment, Generates phoneme segment pairs that are pairs of names and phonetic symbols corresponding to the segment names An occurrence probability calculation means for calculating an occurrence probability based on the frequency of occurrence of the phoneme segment pair in the learning data, and a connected phoneme segment pair sequence that is a sequence of phoneme segment pairs connected in a word in the learning data. A connection probability calculating means for calculating a connection probability based on the appearance frequency; an occurrence probability for each phoneme segment pair calculated by the occurrence probability calculating means; and for each connected phoneme segment pair sequence calculated by the connection probability calculating means A text phonetic symbol conversion dictionary storing means for storing a text phonetic symbol conversion dictionary including the connection probability of the word, and the learning data storage means is a word obtained by connecting a plurality of character strings in which the word can be regarded as one word. In this case, segment division information obtained by dividing the word into segments corresponding to the character string , And to store the phonetic symbols of each segment corresponding to the character string, the.

The recognition vocabulary dictionary creation device according to claim 2 is a recognition vocabulary dictionary creation device for creating a recognition vocabulary dictionary used for recognizing speech, and the text vocabulary symbol conversion dictionary creation according to claim 1 A text phonetic symbol conversion dictionary storage means for storing a text phonetic symbol conversion dictionary created by the apparatus; a recognition vocabulary dictionary storage means for storing a recognition vocabulary dictionary in which a recognition vocabulary subject to speech recognition is registered; and the recognition vocabulary A recognition vocabulary acquisition means for acquiring a recognition vocabulary registered in a dictionary, and a segment sequence for generating a segment sequence that is a segment sequence by dividing the acquired recognition vocabulary into segments by referring to the text phonetic symbol conversion dictionary A phoneme segment which is a sequence of phoneme segment pairs from the segment sequence generated by the generation unit and the segment sequence generation unit Generating a pair sequence and referring to the text phonetic symbol conversion dictionary to calculate a cumulative cost for each phoneme segment pair sequence; and the phoneme based on the cumulative cost calculated by the cumulative cost calculation unit A speech candidate selection means for selecting a higher phoneme segment pair sequence as a phoneme segment pair sequence candidate from the segment pair sequence; and a phonetic symbol sequence corresponding to the phoneme segment pair sequence candidate selected by the speech candidate selection means, Speech candidate registration means for registering in a recognition vocabulary dictionary, and the accumulated cost calculating means includes the occurrence probability of the phoneme segment pair in the phoneme segment pair sequence and the connection probability of the connected phoneme segment pair sequence. The accumulated cost is calculated based on this.

According to a third aspect of the present invention, there is provided a speech recognition apparatus for recognizing a voice based on the recognition vocabulary dictionary created by the recognition vocabulary dictionary creation apparatus according to the second aspect.
A text phonetic symbol conversion dictionary creation program according to claim 4 is a text phonetic symbol conversion dictionary creation program for causing a computer to create a text phonetic symbol conversion dictionary used for conversion from text to phonetic symbols. In addition, data including a word , segment division information obtained by dividing the word into segments, and a phonetic symbol for each segment is stored as learning data, and a plurality of character strings that can be regarded as one word are connected. If it is a word, learning data that further stores segment division information obtained by dividing the word connected to a plurality of character strings into segments corresponding to the character strings, and phonetic symbols for each segment corresponding to the character strings from the storage means, and said word, said segment division information, and phonetic symbols of each of the segments, A phoneme that is a set of a segment name and a phonetic symbol corresponding to the name of the segment from the learning data acquisition step to be acquired, the segment division information acquired by the learning data acquisition step, and the phonetic symbol for each segment An occurrence probability calculating step of generating a segment pair and calculating an occurrence probability based on a frequency of occurrence of the phoneme segment pair in the learning data; and a connected phoneme segment pair sequence that is a sequence of phoneme segment pairs connected in a word. A connection probability calculation step of calculating a connection probability based on a frequency of appearance in the learning data, an occurrence probability of each phoneme segment pair calculated by the occurrence probability calculation step, and a concatenation calculated by the connection probability calculation step The connection probability for each phoneme segment pair sequence Characterized in that it comprises a text phonetic conversion dictionary storage step of storing the non-text phonetic conversion dictionary, the

Further, the recognition vocabulary dictionary creation program according to claim 5, the computer, a recognition vocabulary dictionary creating program for executing the creation of the recognition vocabulary dictionary used to recognize the speech, according to claim 4 A text phonetic symbol conversion dictionary acquisition step for acquiring the text phonetic symbol conversion dictionary from the text phonetic symbol conversion dictionary storage means for storing the text phonetic symbol conversion dictionary created by the text phonetic symbol conversion dictionary creation program; The recognition vocabulary dictionary storage step for storing the recognition vocabulary dictionary in which the recognition vocabulary is registered, the recognition vocabulary acquisition step for acquiring the recognition vocabulary registered in the recognition vocabulary dictionary, and the text phonetic symbol conversion dictionary acquisition step The acquired recognition vocabulary is segmented by referring to the text phonetic symbol conversion dictionary. A segment sequence generation step of generating a segment sequence that is a sequence of segments, and generating a phoneme segment pair sequence that is a sequence of phoneme segment pairs from the segment sequence generated by the segment sequence generation means, A cumulative cost calculating step of calculating a cumulative cost for each phoneme segment pair sequence by referring to the text phonetic symbol conversion dictionary, and a higher rank from the phoneme segment pair sequence based on the cumulative cost calculated in the cumulative cost calculating step A speech candidate selecting step for selecting a phoneme segment pair sequence of the phoneme segment pair sequence as a candidate for a phoneme segment pair sequence, and a speech for registering a phonetic symbol sequence corresponding to the phoneme segment pair sequence candidate selected in the speech candidate selection step in the recognition vocabulary dictionary A candidate registration step, and calculating the accumulated cost Step is characterized in that it comprises a step of calculating the accumulated cost based the an occurrence probability of the phoneme segment pair in the phoneme segment pair sequence, in said connection probabilities of the articulated phoneme segment pair sequence.

According to a sixth aspect of the present invention, there is provided a speech recognition program for executing a process including a step of recognizing speech based on a recognized vocabulary dictionary created by a recognized vocabulary dictionary creating program according to claim 5. .
The text phonetic symbol conversion dictionary creating method according to claim 7 is a text phonetic symbol conversion dictionary creating method for creating a text phonetic symbol conversion dictionary used for text-to-phonetic symbol conversion. When data including segment division information obtained by dividing a word into segments and a phonetic symbol for each segment is stored as learning data, and the word is a word obtained by connecting a plurality of character strings that can be regarded as one word, From the learning data storage means for further storing segment division information obtained by dividing the word connected with a plurality of character strings into segments corresponding to the character strings, and phonetic symbols for each segment corresponding to the character strings, the words and A learning data acquisition step for acquiring the segment division information and a phonetic symbol for each segment; A phoneme segment pair that is a set of a segment name and a phonetic symbol corresponding to the name of the segment is generated from the segment division information acquired by the data acquisition step and the phonetic symbol for each segment. An occurrence probability calculating step for calculating an occurrence probability based on the frequency of occurrence of the phoneme segment pair, and a connected phoneme segment pair sequence that is a sequence of phoneme segment pairs connected in a word based on the frequency of occurrence in the learning data A connection probability calculating step for calculating a connection probability, an occurrence probability for each phoneme segment pair calculated by the occurrence probability calculation step, and a connection probability for each connected phoneme segment pair sequence calculated by the connection probability calculation step. Text to save text phonetic symbol conversion dictionary Characterized in that it comprises a preparative pronunciation symbols conversion dictionary storage step.

The recognition vocabulary dictionary creation method according to claim 8 is a recognition vocabulary dictionary creation method for creating a recognition vocabulary dictionary used for recognizing speech, and the text vocabulary symbol conversion dictionary creation according to claim 7 A text phonetic symbol conversion dictionary acquisition step for acquiring the text phonetic symbol conversion dictionary from the text phonetic symbol conversion dictionary storage means for storing the text phonetic symbol conversion dictionary created by the method, and a recognition vocabulary that is a target of speech recognition are registered. A recognition vocabulary dictionary storage step for storing the recognized vocabulary dictionary, a recognition vocabulary acquisition step for acquiring a recognition vocabulary registered in the recognition vocabulary dictionary, and the text phonogram conversion dictionary acquired in the text phonogram conversion dictionary acquisition step By dividing the acquired recognition vocabulary into segments, a segment system that is a sequence of segments Generating a phoneme segment pair sequence that is a sequence of phoneme segment pairs from the segment sequence generated by the segment sequence generating means and referring to the text phonetic symbol conversion dictionary. A cumulative cost calculating step for calculating a cumulative cost for each segment pair sequence, and a phoneme segment pair sequence that is higher in the phoneme segment pair sequence based on the accumulated cost calculated in the cumulative cost calculating step as a phoneme segment pair sequence candidate A speech candidate selection step of selecting, and a speech candidate registration step of registering a phonetic symbol sequence corresponding to the phoneme segment pair sequence candidate selected in the speech candidate selection step in the recognized vocabulary dictionary, and calculating the accumulated cost The step includes the step in the phoneme segment pair sequence. Said occurrence probability of the rhyme segment pair, and calculates the accumulated cost on the basis of said connection probabilities of the articulated phoneme segment pair sequence.
A speech recognition method according to claim 9 includes a step of recognizing speech based on the recognition vocabulary dictionary created by the recognition vocabulary dictionary creation method according to claim 8 .

According to the text phonetic symbol conversion dictionary creating apparatus according to claim 1, the text phonetic symbol conversion dictionary creating program according to claim 4 , and the text phonetic symbol conversion dictionary creating method according to claim 7 according to the present invention, Text phonetic symbol conversion based on the occurrence probability calculated based on the appearance frequency of each phoneme segment pair and the connection probability of the connected phoneme segment pair sequence representing the likelihood of the phoneme segment pair connection pattern from large-scale learning data Set in dictionary. As a result, the occurrence probability and connection probability based on the statistical value are obtained, so in the conversion from text to phonetic symbol, this text phonetic symbol conversion dictionary can be used to pronounce the word when it is pronounced It is possible to generate a phonetic symbol string with high accuracy. Further, by selecting a higher occurrence probability value and connection probability value according to the size of the text phonetic symbol conversion dictionary, a statistical value having a high importance can be set in the text phonetic symbol conversion dictionary.
Further, according to the text phonetic symbol conversion dictionary creating apparatus according to claim 1, since the phoneme is used as the segment, the phoneme can be easily extracted as the learning data from the phonetic symbol used in a general dictionary. . Furthermore, since not only a segment corresponding to one phoneme but also a segment corresponding to a plurality of connected phonemes is used as a segment, an occurrence probability value and a connection probability value for a segment such as a word are set in the text phonetic symbol conversion dictionary. be able to. Thus, in the conversion from text to phonetic symbols, by using this text phonetic symbol conversion dictionary, phonetic symbols based on phonetic symbols of known words can be generated. Therefore, for a word whose pronunciation is unknown to the user and the word is considered to be connected to a plurality of known words, there is a high possibility that the word is pronounced in consideration of this known word. The accuracy of conversion from to phonetic symbols increases.

According to the recognized vocabulary dictionary creating apparatus according to claim 2 , the recognized vocabulary dictionary creating program according to claim 5 , and the recognized vocabulary dictionary creating method according to claim 8 according to the present invention, the appearance for each phoneme segment pair The occurrence probability calculated based on the frequency and the connection probability of the concatenated phoneme segment pair sequence representing the likelihood of the concatenation pattern of the phoneme segment pair are obtained from large-scale learning data and set in the text phonetic symbol conversion dictionary. Therefore, using this text phonetic symbol conversion dictionary, a phonetic symbol string that is likely to be pronounced when a word is pronounced can be accurately generated and registered in the recognition vocabulary dictionary. In addition, since the probability of the phonetic symbol string can be determined by the value of the accumulated cost corresponding to the generated phonetic symbol string, the phonetic symbols to be flexibly registered in the recognition vocabulary dictionary from among the phonetic symbol string candidates A column can be selected.

Speech recognition apparatus according to claim 3 of the present invention, the speech recognition program according to claim 6, and according to the speech recognition method according to claim 9, the statistical value obtained from a large training data By using the text phonetic symbol conversion dictionary created based on this, the phonetic symbol strings that are likely to be pronounced when a word is pronounced are accurately generated and registered in the recognition vocabulary dictionary. Recognition accuracy can be improved.

[First Embodiment]
Hereinafter, a first embodiment of the present invention will be described with reference to the drawings. 1 to 3 are diagrams showing embodiments of a text phonetic symbol conversion dictionary creation device, a text phonetic symbol conversion dictionary creation program, and a text phonetic symbol conversion dictionary creation method according to the present invention.
First, the structure of the text phonetic symbol conversion dictionary creation apparatus according to the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration of a text phonetic symbol conversion dictionary creation apparatus 100 according to the present invention.

  The learning data storage unit 10 stores, as learning data, data including words represented by character strings, segment division information obtained by dividing each word into segments, and pronunciation symbols corresponding to the segments. Here, the segment means a part of a word text (character string) corresponding to one or more continuous phonetic symbols. In the present embodiment, the word “abaca” and the character corresponding to one phonetic symbol such as “a”, “b”, “c”, “u”, “le” constituting the word “abacule” are segmented. . In addition, phoneme notation is used as a phonetic symbol.

The learning data acquisition unit 11 acquires words, segment division information, and pronunciation symbols corresponding to each segment from the learning data stored in the learning data storage unit 10.
The appearance frequency counter 12 counts the frequency at which a pair of phonemes and segments corresponding to a segment (hereinafter referred to as a phoneme segment pair) appears in the learning data, and stores it in the memory 16. The frequency at which each segment appears in the learning data is counted and stored in the memory 16. When the processing for all the words in the learning data is completed, the appearance frequency of each phoneme segment pair and the total frequency that is the appearance frequency of each segment are determined and stored in the memory 16.
The occurrence probability calculation unit 13 calculates the occurrence probability for each phoneme segment pair and stores it in the memory 16 by dividing the appearance frequency of the phoneme segment pair by the total frequency corresponding to the segment included in the phoneme segment pair.

The connection probability calculation unit 14 counts the frequency at which a connected phoneme segment pair sequence, which is a sequence of phoneme segment pairs connected in a word, appears in the learning data, and stores it in the memory 16. When the processing for all the words in the learning data is completed, the appearance frequency for each connected phoneme segment pair sequence is determined and stored in the memory 16. Then, the connection probability is calculated for each connected phoneme segment pair sequence by dividing the appearance frequency of the connected phoneme segment pair sequence by the total appearance frequency of the preceding phoneme segment pair in the connected phoneme segment pair sequence. save.
The text phonetic symbol conversion dictionary storage unit 15 reads out the occurrence probability calculated for each phoneme segment pair and the connection probability calculated for each connected phoneme segment pair sequence from the memory 16 and stores them as a text phonetic symbol conversion dictionary.

Next, the flow of processing for creating a text phonetic symbol conversion dictionary in the text phonetic symbol conversion dictionary creating device 100 having such a configuration will be described with reference to FIG. FIG. 2 is a flowchart for explaining a method for creating a text phonetic symbol conversion dictionary executed by the text phonetic symbol conversion dictionary creating device 100.
As shown in the flowchart of FIG. 2, the text phonetic symbol conversion dictionary creation process first proceeds to step S201, where the learning data acquisition unit 11 uses words and segments from the learning data stored in the learning data storage unit 10. The division information and the phonetic symbols corresponding to each segment are acquired, and the process proceeds to step S202.

In step S202, the appearance frequency counter 12 determines whether or not the processing for all the words in the learning data has been completed. If it is determined that the processing has been completed (Yes), the process proceeds to step S205. (No) moves to step S203.
When the process proceeds to step S <b> 203, the appearance frequency counter 12 counts the frequency at which each phoneme segment pair appears in the learning data and stores it in the memory 16. Further, the frequency at which each segment appears in the learning data is counted and stored in the memory 16, and the process proceeds to step S204.

In step S204, the connection probability calculation unit 14 counts the frequency at which the connected phoneme segment pair sequence, which is a sequence of phoneme segment pairs connected in the word, appears in the learning data, stores it in the memory 16, and proceeds to step S201. .
When the process proceeds to step S205, the occurrence probability calculation unit 13 divides the appearance frequency of the phoneme segment pair by the total frequency corresponding to the segment included in the phoneme segment pair, thereby calculating the occurrence probability for each phoneme segment pair. The calculated value is stored in the memory 16, and the process proceeds to step S206.

In step S206, the connection probability calculation unit 14 divides the appearance frequency of the connected phoneme segment pair sequence by the total appearance frequency of the preceding phoneme segment pair in the connected phoneme segment pair sequence, thereby obtaining each connected phoneme segment pair sequence. The connection probability is calculated and stored in the memory 16, and the process proceeds to step S207.
In step S207, the text phonetic symbol conversion dictionary storage unit 15 reads out the occurrence probability calculated for each phoneme segment pair and the connection probability calculated for each connected phoneme segment pair sequence from the memory 16, and the text phonetic symbol conversion dictionary. Save as to finish the process.

Example 1
The operation of the text phonetic symbol conversion dictionary creation device 100 will be specifically described below with reference to FIGS. 3 (a), (b), and (c).
The learning data stored in the learning data storage unit 10 is generated based on data extracted from a general dictionary including the spelling of the word and the phonetic symbol of the word. In the present embodiment, as data included in the learning data, as shown in FIG. 3A, for example, segment division information for the word “abaca” is “(a) baca”, “a (b) aca”, “ It consists of information such as “ab (a) ca”, “aba (c) a”, and “abac (a)”. The word “abaca” is divided into five segments such as “a”, “b”, “a”, “c”, “a”, and the word “abacule” is divided into “a”, “b”, “a”. ”,“ C ”,“ u ”,“ le ”, and so on.

  A phonetic symbol corresponding to each segment is expressed as, for example, “(a) baca → a”. The phonemes corresponding to the five segments “a”, “b”, “a”, “c”, “a” constituting the word “abaca” are / a /, / b /, / a /, respectively. , / K /, / a /. The phonemes corresponding to the six segments “a”, “b”, “a”, “c”, “u”, “le” constituting the word “abacule” are / a /, / b, respectively. /, / A /, / k /, / y /, / l /.

  The learning data acquisition unit 11 acquires words, segment division information, and pronunciation symbols corresponding to each segment from the learning data. Divide words obtained from learning data into segments. The appearance frequency counter 12 counts the frequency at which phoneme segment pairs appear in the learning data and stores them in the memory 16. Hereinafter, the segment “y” that is pronounced as the phonetic symbol “X” is denoted as “X | y”. When only the word “abaca” is processed, the appearance frequencies of the phoneme segment pairs “a | a”, “b | b”, and “k | c” are three times, once, and once, respectively.

  When the processing for all the words in the learning data is completed, the appearance frequency of each phoneme segment pair and the total frequency that is the appearance frequency of each segment are determined and stored in the memory 16. In the example of FIG. 3A, the appearance frequencies of the phoneme segment pairs “A | a”, “a | a”, “b | b”, and “p | b” are 1223, 45142, 12372, and 267, respectively.

  The occurrence probability calculation unit 13 calculates the occurrence probability for each phoneme segment pair and stores it in the memory 16 by dividing the appearance frequency of the phoneme segment pair by the total frequency corresponding to the segment included in the phoneme segment pair. In the example of FIG. 3A, the total frequency of the segment “a” is 46365 times. There are two types of phonemes corresponding to the segment “a”, / A / and / a /, and the frequency of occurrence of the phoneme segment pair “A | a” and “a | a” is 1223 times and 45142 times, respectively. It is. At this time, the occurrence probabilities are 0.03 and 0.97, respectively. In order to reduce the amount of calculation in a recognition vocabulary dictionary creation apparatus, which will be described later, a logarithmic value of occurrence probability is used. For example, for occurrence probability values 0.03 and 0.97, −3.63 and −0.02 are logarithmic values of the occurrence probabilities, respectively.

  The connection probability calculation unit 14 counts the frequency at which a connected phoneme segment pair sequence, which is a sequence of phoneme segment pairs connected in a word, appears in the learning data, and stores it in the memory 16. In the word “abaca”, the concatenated phoneme segment pair sequence composed of two consecutive phoneme segment pairs is “a | a → b | b”, “b | b → a | a”, “a | a → k | c”. ”,“ K | c → a | a ”, and the frequency of appearance is once. When the processing for all the words in the learning data 11 is completed, the appearance frequency of each pair of phoneme segment pairs is determined and stored in the memory 16.

  Then, by dividing the appearance frequency for each connected phoneme segment pair sequence by the total appearance frequency of the preceding phoneme segment pair of the connected phoneme segment pair sequence, a connection probability is calculated for each connected phoneme segment pair sequence and stored in the memory 16. To do. As shown in FIG. 3A, the total frequency of “a | a” is 45142 times. The frequency of occurrence of “a | a → b | b” is 2487 times. At this time, the connection probability of “a | a → b | b” is 0.055, and its logarithmic value is −2.89. Hereinafter, logarithmic values are used as values of occurrence probability and connection probability.

  As shown in FIG. 3A, the information on the occurrence probability for each phoneme segment pair includes the names of phoneme segment pairs “A | a”, “a | a”, “b | b”, “p | It is a numerical value such as −3.63, −0.02, −0.02, −3.85, which is a probability of occurrence for each pair of phoneme segments and a label such as “b”. The information on the connection probability for each connected phoneme segment pair sequence is “a | a → b | b”, “a | a →”, which is the name of the connected phoneme segment pair sequence, as shown in FIG. p | b "," a | a → k | c ", labels such as" a | a → s | c "and the connection probabilities of the connected phoneme segment pairs -2.89, -8.23,- It becomes a numerical value such as 3.18, -4.69.

As described above, the present embodiment is not limited to a configuration in which characters corresponding to one phonetic symbol are segmented, and a character composed of a plurality of phonetic symbols can also be segmented. is there. The characters composed of a plurality of phonetic symbols are, for example, words, prefixes, suffixes, and the like.
For example, each of the character strings “up” and “hall” constituting the word “uphall” can be regarded as one word.

  When these character strings are segmented, the corresponding phoneme segment pairs are “ap | up” and “hol | hall”, respectively. Such a segment may be mixed in a segment that is a character corresponding to the above-mentioned one phonetic symbol to create a text phonetic symbol conversion dictionary. In this way, the occurrence probability value and the connection probability value for the segment that is a word can be set in the text phonetic symbol conversion dictionary, so that the recognition vocabulary dictionary creation device described later can generate pronunciation based on the phonetic symbol of a known word. Symbols can be generated. Therefore, for a word whose pronunciation is unknown to the user and the word is considered to be connected to a plurality of known words, there is a high possibility that the word is pronounced in consideration of this known word. The accuracy of conversion from to phonetic symbols increases.

Note that when the frequency of the segment, which is a character composed of a plurality of phonetic symbols, appears in the learning data, the occurrence probability and the connection probability may be set to predetermined values.
The text phonetic symbol conversion dictionary storage unit 15 stores information on the occurrence probability for each phoneme segment pair and information on the connection probability for each connected phoneme segment pair sequence as a text phonetic symbol conversion dictionary. In the first embodiment, the learning data storage unit 10 corresponds to the learning data storage unit described in claim 1, and the learning data acquisition unit 11 corresponds to the learning data acquisition unit described in claim 1 and appears. The frequency counter 12 and the occurrence probability calculation unit 13 correspond to the occurrence probability calculation unit according to claim 1, and the connection probability calculation unit 14 corresponds to the connection probability calculation unit according to claim 1, and is stored in a text phonetic symbol conversion dictionary. The unit 15 corresponds to the text phonetic symbol conversion dictionary storage means described in claim 1.

  In the first embodiment, step S201 corresponds to the learning data acquisition step according to claim 6 or claim 9, and steps S202 to S203 and S205 are the occurrence according to claim 6 or claim 9. Corresponding to the probability calculating step, steps S204 and S206 correspond to the connection probability calculating step according to claim 6 or claim 9, and step S207 is the text phonetic symbol conversion dictionary storing step according to claim 6 or claim 9. Corresponding to

  The text phonetic symbol conversion dictionary creation program according to the first embodiment described above can be executed by a general computer system including a storage unit. In this case, when the computer executes the text phonetic symbol conversion dictionary creation program stored in the storage unit, the text phonetic symbol conversion dictionary creation operation described above is performed. The text phonetic symbol conversion dictionary creation program may be supplied to the computer system via a communication medium. The text phonetic symbol conversion dictionary creation program may be recorded on a storage medium such as an optical disk, and the text phonetic symbol conversion dictionary creation program recorded on the recording medium may be read by a computer system.

[Second Embodiment]
Next, a second embodiment of the present invention will be described with reference to the drawings. 4 to 8 are diagrams showing embodiments of a recognized vocabulary dictionary creating apparatus, a recognized vocabulary dictionary creating program, and a recognized vocabulary dictionary creating method according to the present invention.
First, the configuration of the recognized vocabulary dictionary creating apparatus according to the present invention will be described with reference to FIG. FIG. 4 is a block diagram illustrating a recognized vocabulary dictionary creating apparatus 400 according to the present invention. The first recognized vocabulary dictionary storage unit 40 stores a first recognized vocabulary dictionary in which phonetic symbol strings of the recognized vocabulary are registered in advance for each recognized vocabulary to be speech-recognized. The second recognized vocabulary dictionary storage unit 41 stores a second recognized vocabulary in which a recognized vocabulary registered for speech recognition by the user is stored. In the second recognition vocabulary dictionary and the first recognition vocabulary dictionary, the phonetic symbol string of the recognition vocabulary is registered in advance in the first recognition vocabulary dictionary, whereas in the second recognition vocabulary dictionary, the recognition vocabulary dictionary according to the present invention is used. It differs in that it is newly registered by the creation device.

The text phonetic symbol conversion dictionary storage unit 42 stores the text phonetic symbol conversion dictionary created by the text phonetic symbol conversion dictionary creation device described above. The recognized vocabulary acquisition unit 43 acquires the recognized vocabulary stored in the second recognized vocabulary dictionary. The segment series generation unit 44 divides the acquired recognition vocabulary into segments and generates a segment series. In this embodiment, the segment part of the phoneme segment pair stored in the text phonetic symbol conversion dictionary is used as the segment label to be referred to for division, but a table in which segment labels are registered may be used. .
Taking the word “abaca” as an example, a segment sequence consisting of five segments {“a”, “b”, “a”, “c”, “a”} and a segment sequence consisting of four segments { Two segment sequences “a”, “b”, “ac”, “a”} are generated.

The accumulated cost calculation unit 45 generates a phoneme segment pair sequence from the segment sequence by assigning all possible phonetic symbols to each segment of the segment sequence generated by the segment sequence generation unit 44. The phoneme segment pairs stored in the text phonetic symbol conversion dictionary are referenced to assign phonetic symbols to the segments. Next, for each series of phoneme segment pairs, an accumulated cost is calculated with reference to the occurrence probability and the connection probability stored in the text phonetic symbol conversion dictionary 43.
The speech candidate selection unit 46 selects a speech candidate based on the accumulated cost from a series of phoneme segment pairs for which the accumulated cost is to be calculated. The speech candidate registration unit 47 registers the speech candidate selected by the speech candidate selection unit 46 in the second recognition vocabulary dictionary as a pronunciation symbol string of the corresponding recognition vocabulary.

Next, the flow of a recognition vocabulary dictionary creation process in the recognition vocabulary dictionary creation apparatus 400 configured as described above will be described with reference to FIG. FIG. 5 is a flowchart for explaining a recognition vocabulary dictionary creation method executed by the recognition vocabulary dictionary creation apparatus 400 described above.
As shown in the flowchart of FIG. 5, the recognition vocabulary dictionary creation processing first proceeds to step S501, and the recognition vocabulary acquisition unit 43 inputs a word included in the second recognition vocabulary dictionary 41, and then proceeds to step S502. . In step S502, the segment series generation unit 44 divides the input word into segments to generate a segment series, and the process proceeds to step S503. In step S503, the accumulated cost calculation unit 45 generates all possible phonetic symbols for each segment in the segment series, thereby generating a series of phoneme segment pairs from the segment series, and proceeds to step S504.

  In step S504, the accumulated cost calculation unit 45 calculates the accumulated cost for each of the phoneme segment pair sequences based on the occurrence probability and the accumulated probability stored in the text phonetic symbol conversion dictionary, and the process proceeds to step S505. In step S505, the speech candidate selection unit 46 compares the calculated accumulated cost with the boundary likelihood. Then, a speech candidate having an accumulated cost equal to or higher than the boundary likelihood is selected, and the process proceeds to step S506. In step S506, the speech candidate registration unit 47 registers the speech candidate selected by the speech candidate selection unit 46 in the second recognized vocabulary dictionary as the phonetic symbol string of the corresponding recognized vocabulary, and the process ends.

Example 1
Next, the operation of the accumulated cost calculation unit 45 will be specifically described with reference to FIG. FIG. 6 shows an example in which the accumulated cost calculation unit 45 calculates the accumulated cost of the speech candidate of the word “abaca”. For example, the accumulated cost calculation unit 45 calculates the accumulated cost of the speech candidate / abaka / of the word “abaca” as follows. That is, in the first segment “a” of the word “abaca”, there are two speech candidates of phonetic symbol / a / and phonetic symbol / A /. The accumulated cost calculation unit 45 accumulates the occurrence probability of the phonetic symbol /a/−0.02, the probability of connection between the phonetic symbol / a / and the phonetic symbol /b/-2.89, and the probability of occurrence of the phonetic symbol / b /. Then, the accumulated cost of the character string “ab” is calculated.

Since the present embodiment uses the logarithmic value of the occurrence probability and the connection probability as described above, the calculation amount is reduced by using addition instead of multiplication as the calculation for calculating the accumulated cost. can do.
Further, the accumulated cost calculation unit 45 accumulates the connection probability that continues from the phonetic symbol / b / to the phonetic symbol / a / and the occurrence probability that the speech candidate of the third segment “a” is the phonetic symbol / a /. . Further, the connection probability that continues from phonetic symbol / a / to phonetic symbol / k / and the occurrence probability that the speech candidate of segment “c” is phonetic symbol / k / are accumulated, and phonetic symbol / k / to phonetic symbol / The connection probability consecutive to a / and the occurrence probability that the speech candidate of the segment “a” is the phonetic symbol / a / are accumulated. Through the above calculation, the accumulated cost calculation unit 45 obtains the accumulated cost of 10.03 of the phonetic symbol string / abaka /.
Similarly, the accumulated cost calculation unit 45 calculates the accumulated cost for the phonetic symbol strings / abAka /, / abakA /, / Abaka /, / abasa /, etc., which are other speech candidates.

  Next, the speech candidate selection unit 46 selects a speech candidate based on the accumulated cost from the phonetic symbols for which the accumulated cost has been calculated. The speech candidate selection unit 46 of the present embodiment selects a speech candidate whose accumulated cost is greater than a predetermined threshold value. The threshold value may be a predetermined value. Further, it may be a value obtained by subtracting a predetermined value from the largest value among the calculated accumulated costs.

FIG. 7 is a diagram for explaining an example in which a threshold value is set by subtracting a predetermined value from the largest value among the calculated accumulated costs. In the illustrated example, the accumulated cost is calculated for the phonetic symbol strings / abaka /, / abAka /, / abakA /, / Abaka /, / abasa /. In the present embodiment, the phonetic symbol strings / abaka /, / abAka /, / abakaA /, / Abaka /, / abasa / are assigned from the first candidate to the fifth candidate in descending order of the accumulated cost. Shall be given the ranking. The accumulated cost of each phonetic symbol is as follows.
First candidate phonetic symbol string / abaka / accumulated cost -10.03
Second candidate phonetic symbol string / abAka / accumulated cost -12.70
Third candidate phonetic symbol string / abakA / accumulated cost -14.25
Fourth candidate phonetic symbol string / Abaka / accumulated cost -16.53
5th candidate phonetic symbol string / abaSa / accumulated cost -17.64

  In the example shown in FIG. 7, the predetermined value is 5.00. The speech candidate selection unit 46 sets a threshold value (boundary likelihood) -15.03 by subtracting 5.00 from -10.03 which is the largest accumulated cost. The speech candidate selection unit 46 selects a phonetic symbol string / abaka /, / abAka /, / abakA / that is a speech candidate whose accumulated cost is equal to or higher than the boundary likelihood. The speech candidate registration unit 47 stores the selected speech candidate in the second recognized vocabulary dictionary. On the other hand, phonetic symbol strings / Abaka / and / abaSa /, which are speech candidates whose accumulated cost is equal to or lower than the boundary likelihood, are excluded.

  When configured in this way, a threshold value can be set depending on the cumulative cost value of the first candidate of each speech candidate, so that the cumulative cost value of the first candidate speech candidate is used as a reference. Thus, speech candidates that are unlikely to be input can be excluded from registration targets of the second recognition vocabulary dictionary. Therefore, it is possible to increase the recognition rate of the input speech while suppressing an increase in the size of the second recognition vocabulary dictionary.

Further, the present embodiment is not limited to the configuration for selecting speech candidates as described above, and a predetermined number of speech candidates having a higher accumulated cost may be selected. That is, for example, when registering three speech candidates for one word in the second recognition vocabulary dictionary, the speech candidate selection unit 46 determines, for the word “abaca”, three speech candidates with higher accumulated costs, phonetic symbol strings / Select abaka /, / abAka /, / abakA /.
When configured in this manner, the present embodiment can predict the number of speech candidates registered in advance and the capacity of the second recognition vocabulary dictionary necessary for the registration.

Further, as described above, the present embodiment is not limited to the configuration in which the character corresponding to one phonetic symbol is segmented, and the character constituted by a plurality of phonetic symbols may be segmented. Is possible. The characters composed of a plurality of phonetic symbols are, for example, words, prefixes, suffixes, and the like. Furthermore, it is not limited to calculating the total cost of the occurrence probability and the connection probability, and it is not limited to calculating the accumulated cost, and the sum of these can be calculated by multiplying each of the occurrence probability and the connection probability by a weighting factor. Is possible.
When the recognized vocabulary acquired from the second recognized vocabulary dictionary is divided into segments by the segment series generation unit 44, it is determined whether or not the acquired recognized vocabulary is registered in the first recognized vocabulary dictionary. The phonetic symbol string registered in the first recognized vocabulary dictionary may be registered in the second recognized vocabulary dictionary.

  FIG. 8 is a diagram for explaining an example in which a character composed of a plurality of phonetic symbols is used as a segment. In the present embodiment, a segment corresponding to one phonetic symbol is also referred to as a normal segment, and a segment corresponding to a plurality of connected phonetic symbols is also referred to as an extended segment. The text phonetic symbol conversion dictionary used in this case is created by a text phonetic symbol conversion dictionary creation device in which a normal segment and an extended segment are mixed.

In the example shown in FIG. 8, a segment composed of a plurality of connected phonetic symbols is set for the word “uphall”. A range indicated by reference numeral 82 indicates a series of phoneme segment pairs generated from segments corresponding to one phonetic symbol. A range indicated by reference numeral 81 indicates a series of phoneme segment pairs generated from segments corresponding to a plurality of connected phonetic symbols.
In the example indicated by reference numeral 81, an extended segment is formed by dividing the word “upall” into a plurality of words such as the character strings “up” and “hall”. As shown in the figure, in the recognized vocabulary dictionary creating apparatus of the present embodiment, normal segments and extended segments are mixed.

  By introducing an extended segment, an occurrence probability value and a connection probability value for a segment that is a word can be set in the text phonetic symbol conversion dictionary. Thus, in the conversion from text to phonetic symbols, by using this text phonetic symbol conversion dictionary, phonetic symbols based on phonetic symbols of known words can be generated. Therefore, for a word whose pronunciation is unknown to the user and the word is considered to be connected to a plurality of known words, there is a high possibility that the word is pronounced in consideration of this known word. The accuracy of conversion from to phonetic symbols increases.

  In the second embodiment, the second recognition vocabulary dictionary 41 corresponds to the recognition vocabulary dictionary storage means described in claim 4, and the text phonetic symbol conversion dictionary storage unit 42 includes the text phonetic symbol conversion described in claim 4. Corresponding to the dictionary storage means, the recognized vocabulary acquisition unit 43 corresponds to the recognized vocabulary acquisition means described in claim 4, and the segment sequence generation unit 44 corresponds to the segment sequence generation means described in claim 4, and calculates the accumulated cost. The unit 45 corresponds to the accumulated cost calculating means described in claim 4, the speech candidate selecting unit 46 corresponds to the speech candidate selecting means described in claim 4, and the speech candidate registering unit 47 corresponds to the speech described in claim 4. Corresponds to candidate registration means.

  In the second embodiment, step S501 corresponds to the recognized vocabulary acquisition step according to claim 7 or claim 10, and step S502 corresponds to the segment sequence generation step according to claim 7 or claim 10. Correspondingly, steps S503 to S504 correspond to the accumulated cost calculation step according to claim 7 or claim 10, and step S504 corresponds to the text phonetic symbol conversion dictionary acquisition step according to claim 7 or claim 10, Step S505 corresponds to the speech candidate selection step according to claim 7 or claim 10, and step S506 corresponds to the speech candidate registration step according to claim 7 or claim 10.

  The recognized vocabulary dictionary creation program according to the second embodiment described above can be executed by a general computer system including a storage unit. In this case, the recognized vocabulary dictionary creating operation described above is performed by the computer executing the recognized vocabulary dictionary creating program stored in the storage unit. The recognized vocabulary dictionary creation program may be supplied to the computer system via a communication medium. Alternatively, the recognition vocabulary dictionary creation program may be recorded on a storage medium such as an optical disk, and the recognition vocabulary dictionary creation program recorded on the recording medium may be read by a computer system.

[Third Embodiment]
Next, a third embodiment of the present invention will be described with reference to the drawings. FIG. 9 is a diagram showing an embodiment of a speech recognition apparatus according to the present invention. The speech recognition apparatus 900 of the present embodiment includes the recognition vocabulary dictionary 93 created by the recognition vocabulary dictionary creation apparatus 400 of the present embodiment described above and performs speech recognition.
As illustrated in FIG. 9, the speech recognition apparatus 900 includes a speech input unit 90 that inputs input speech, a feature amount extraction unit 91 that extracts a time series of feature amounts from the input speech, and an acoustic model storage that stores an acoustic model. Unit 92, recognition vocabulary dictionary 93 for storing speech candidates, matching unit 94 for pattern matching, recognition result output unit 95 for outputting input speech recognition result candidates, and operation unit 96 for inputting control signals to the speech recognition device It has.

  The voice input unit 90 performs A / D conversion on voice input by the user with a microphone (not shown) or the like. The feature amount extraction unit 91 extracts a time series of feature amounts from the input voice. The acoustic model storage unit 92 stores an acoustic model expressed by, for example, a continuous distribution type HMM (Hidden Markov Model). The acoustic model is created for every phoneme using speech data of a large number of speakers.

  The recognized vocabulary dictionary 93 is created by a recognized vocabulary dictionary creating device in which normal segments and extended segments are mixed as shown in FIG. For example, the extension segment includes character strings “up” and “hall”. For each speech candidate stored in the recognized vocabulary dictionary 93, the matching unit 94 generates a speech pattern model by connecting the acoustic model according to the phoneme notation (phonetic symbol) of the speech candidate, while the feature amount extracting unit 91 When the feature amount time series is received, the acoustic likelihood of the speech candidate is obtained by pattern matching between the feature amount time series and speech pattern models related to a plurality of speech candidates generated in advance. A plurality of speech candidates are set as input speech recognition result candidates.

  The recognition result output unit 95 outputs a plurality of words corresponding to a plurality of speech candidates to a display (not shown) or the like when a recognition result candidate is obtained by pattern matching of the matching unit 94. The operation unit 96 instructs to select one of the plurality of output words or to start or stop the voice recognition to the voice recognition apparatus. Or it is used for operation for a user to perform the next instruction | indication based on the result of the arithmetic processing performed based on the recognition result and the recognition result.

Next, this embodiment will be specifically described.
First, when the user speaks into a microphone (not shown), the voice input unit 90 performs A / D conversion on the input voice input to the microphone.
The feature amount extraction unit 91 analyzes the digital signal of the input voice and extracts a time series of feature amounts such as MFCC (Mel Frequency Cepstrum Coefficient).

  For each speech candidate stored in the recognized vocabulary dictionary 93 in advance, the matching unit 94 connects the acoustic model stored in the acoustic model storage unit 92 according to the phoneme notation (phonetic symbol) of the speech candidate, thereby generating a speech pattern. A model (speech pattern model obtained by modeling a time series pattern of feature values) is generated. When a time series of feature quantities extracted by the feature quantity extraction unit 91 is input, for example, a voice candidate is obtained by pattern matching between the feature quantity time series and a voice pattern model for each voice candidate using a Viterbi algorithm. Obtain the acoustic likelihood of. For example, when the word “uphall” is registered in the recognition vocabulary dictionary 93 as a recognition target, when the user utters the word “uphall”, the speech candidates / aphal /, / apfal / stored in the recognition vocabulary dictionary 93 , / Ufol /, / uphor /, / aphor /, and the like, and pattern matching is performed between the time series of feature amounts extracted from the input speech and the speech pattern model.

Since the speech candidates / aphor / corresponding to the character strings “up” and “hall” as the extended segments are stored in the recognition vocabulary dictionary 93, the acoustic likelihood is high even if the user pronounces the phonetic symbol / aphor /. Thus, the speech candidate / aphor / is included in the recognition result candidates.
The recognition result output unit 95 outputs the recognition result candidate obtained by the collation unit 94 to a display or the like. Regardless of whether the user pronounces the phonetic symbol / ufol / or / aphor /, the word "ufall" is a recognition result candidate.

  The user operates the operation unit 96 to select the word “uphall” from the recognition result candidates. The recognition result of the voice recognition device 900 is input to, for example, a map search function of a car navigation device. In such a case, the map search function performs control such as extracting information on the latitude and longitude of the point indicated by the word “upall”, assuming that the word “uphall” is input as the place name.

In the third embodiment, the recognized vocabulary dictionary 93 corresponds to the recognized vocabulary dictionary created by the recognized vocabulary dictionary creating device according to claim 4.
Note that the voice recognition program according to the third embodiment described above can be executed by a general computer system including a storage unit. In this case, the above-described speech recognition operation is performed by the computer executing the speech recognition program stored in the storage unit. The voice recognition program may be supplied to the computer system via a communication medium. Further, the voice recognition program may be recorded on a storage medium such as an optical disk, and the voice recognition program recorded on the recording medium may be read by a computer system.

  The present invention accurately generates a phonetic symbol string that is likely to be pronounced when a word is pronounced by using a text phonetic symbol conversion dictionary created based on statistical values obtained from large-scale learning data. Therefore, it can be used to create a recognition vocabulary dictionary for speech recognition.

It is a block diagram which shows the structure of the text phonetic symbol conversion dictionary creation apparatus 100 of the 1st Embodiment of this invention. It is a flowchart which shows the creation process of the text phonetic symbol conversion dictionary in the text phonetic symbol conversion dictionary creation apparatus 100 of the 1st Embodiment of this invention. It is a figure for demonstrating operation | movement of the text phonetic symbol conversion dictionary creation apparatus 100 shown in FIG. It is a block diagram which shows the structure of the recognition vocabulary dictionary creation apparatus 400 of this invention. It is a flowchart which shows the creation process of the recognition vocabulary dictionary in the recognition vocabulary dictionary creation apparatus 400 shown in FIG. It is a figure for demonstrating operation | movement of the accumulated cost calculation part shown in FIG. It is a figure for demonstrating the example which sets the threshold value of the accumulation cost of the 2nd Embodiment of this invention. It is a figure for demonstrating the example which set the segment corresponding to the several phonetic symbol to which the 2nd Embodiment of this invention is connected. It is a block diagram which shows the structure of the speech recognition apparatus 900 of the 3rd Embodiment of this invention. It is a figure for demonstrating the example of the word dictionary used with the conventional spelling-phonetic-symbol conversion method. It is a figure for demonstrating the example of the rule used with the conventional spelling-phonetic-symbol conversion method.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Text phonetic symbol conversion dictionary creation apparatus 400 Recognition vocabulary dictionary creation apparatus 900 Speech recognition apparatus 10 Learning data storage part 11 Learning data acquisition part 12 Appearance frequency counter 13 Occurrence probability calculation part 14 Connection probability calculation part 15 Text phonetic symbol conversion dictionary preservation | save part 16 Memory 40 First recognition vocabulary dictionary storage unit 41 Second recognition vocabulary dictionary storage unit 42 Text phonetic symbol conversion dictionary storage unit 43 Recognition vocabulary acquisition unit 44 Segment sequence generation unit 45 Cumulative cost calculation unit 46 Speech candidate selection unit 47 Speech candidate registration Unit 90 voice input unit 91 feature amount extraction unit 92 acoustic model storage unit 93 recognition vocabulary dictionary 94 collation unit 95 recognition result output unit 96 operation unit

Claims (9)

A text phonetic symbol conversion dictionary creation device for creating a text phonetic symbol conversion dictionary used for conversion from text to phonetic symbols,
Learning data storage means for storing, as learning data, data including a word , segment division information obtained by dividing the word into segments, and pronunciation symbols for the segments;
Learning data acquisition means for acquiring the word, the segment division information, and the phonetic symbol for each segment from the learning data;
A phoneme segment pair that is a set of a segment name and a phonetic symbol corresponding to the segment name is generated from the segment division information acquired by the learning data acquisition unit and the phonetic symbol for each segment, and the learning An occurrence probability calculating means for calculating an occurrence probability based on the frequency of occurrence of the phoneme segment pair in the data;
A concatenated phoneme segment pair sequence that is a sequence of phoneme segment pairs concatenated within a word is
Connection probability calculation means for calculating a connection probability based on the frequency of occurrence in the learning data;
A text phonetic symbol for storing a text phonetic symbol conversion dictionary including the occurrence probability for each phoneme segment pair calculated by the occurrence probability calculating unit and the connection probability for each connected phoneme segment pair sequence calculated by the connection probability calculating unit A conversion dictionary storage means;
Equipped with a,
The learning data storage means, when the word is a word obtained by connecting a plurality of character strings that can be regarded as one word, segment division information obtained by dividing the word connected by a plurality of character strings into segments corresponding to the character strings And a phonetic symbol conversion dictionary creating apparatus for storing a phonetic symbol for each segment corresponding to the character string . A recognition vocabulary dictionary creating device for creating a recognition vocabulary dictionary used for recognizing speech,
A text phonetic symbol conversion dictionary storage means for storing a text phonetic symbol conversion dictionary created by the text phonetic symbol conversion dictionary creating device according to claim 1;
A recognition vocabulary dictionary storing means for storing a recognition vocabulary dictionary in which a recognition vocabulary to be speech recognition is registered;
Recognition vocabulary acquisition means for acquiring a recognition vocabulary registered in the recognition vocabulary dictionary;
Segment sequence generation means for dividing the acquired recognition vocabulary into segments by referring to the text phonetic symbol conversion dictionary, and generating a segment sequence that is a sequence of segments;
A phoneme segment pair sequence that is a sequence of phoneme segment pairs is generated from the segment sequence generated by the segment sequence generation means, and the accumulated cost is calculated for each phoneme segment pair sequence by referring to the text phonetic symbol conversion dictionary Cumulative cost calculation means to
A speech candidate selection unit that selects a higher phoneme segment pair sequence as a phoneme segment pair sequence candidate from the phoneme segment pair sequence based on the accumulated cost calculated by the accumulated cost calculation unit;
Speech candidate registration means for registering a sequence of phonetic symbols corresponding to the phoneme segment pair sequence candidate selected by the speech candidate selection means in the recognized vocabulary dictionary,
The cumulative cost calculating means calculates a cumulative cost based on the occurrence probability of the phoneme segment pair in the phoneme segment pair sequence and the connection probability of the connected phoneme segment pair sequence. Creation device.   A speech recognition apparatus for recognizing speech based on a recognition vocabulary dictionary created by the recognition vocabulary dictionary creation apparatus according to claim 2. A text phonetic symbol conversion dictionary creation program for causing a computer to create a text phonetic symbol conversion dictionary used for conversion from text to phonetic symbols,
Data including a word , segment division information obtained by dividing the word into segments, and pronunciation symbols for each segment are stored as learning data, and the word is a word obtained by connecting a plurality of character strings that can be regarded as one word. In this case, from learning data storage means for further storing segment division information obtained by dividing the word connected to a plurality of character strings into segments corresponding to the character strings, and phonetic symbols for each segment corresponding to the character strings Learning data acquisition step for acquiring the word, the segment division information, and the phonetic symbol for each segment;
A phoneme segment pair that is a set of a segment name and a phonetic symbol corresponding to the name of the segment is generated from the segment division information acquired by the learning data acquisition step and the phonetic symbol for each segment, and the learning An occurrence probability calculating step for calculating an occurrence probability based on the frequency of occurrence of the phoneme segment pair in the data;
A concatenated phoneme segment pair sequence that is a sequence of phoneme segment pairs concatenated within a word is
A connection probability calculation step of calculating a connection probability based on the frequency of appearance in the learning data;
A text phonetic symbol for storing a text phonetic symbol conversion dictionary including the occurrence probability for each phoneme segment pair calculated by the occurrence probability calculating step and the connection probability for each connected phoneme segment pair sequence calculated by the connection probability calculating step A conversion dictionary saving step;
A text phonetic symbol conversion dictionary creation program characterized by including: A recognition vocabulary dictionary creating program for causing a computer to create a recognition vocabulary dictionary used for recognizing speech,
A text phonetic symbol conversion dictionary acquisition step for acquiring the text phonetic symbol conversion dictionary from the text phonetic symbol conversion dictionary storage means for storing the text phonetic symbol conversion dictionary created by the text phonetic symbol conversion dictionary creation program according to claim 4; ,
A recognition vocabulary dictionary storage step for storing a recognition vocabulary dictionary in which recognition vocabulary to be recognized by speech recognition is registered;
A recognition vocabulary acquisition step of acquiring a recognition vocabulary registered in the recognition vocabulary dictionary;
A segment sequence generation step of dividing the acquired recognition vocabulary into segments by referring to the text phonetic symbol conversion dictionary acquired in the text phonetic symbol conversion dictionary acquisition step, and generating a segment sequence which is a sequence of segments;
A phoneme segment pair sequence that is a sequence of phoneme segment pairs is generated from the segment sequence generated by the segment sequence generation means, and the accumulated cost is calculated for each phoneme segment pair sequence by referring to the text phonetic symbol conversion dictionary A cumulative cost calculating step,
A speech candidate selection step of selecting, as a phoneme segment pair sequence candidate, an upper phoneme segment pair sequence from the phoneme segment pair sequence based on the accumulated cost calculated in the accumulated cost calculating step;
A speech candidate registration step of registering a sequence of phonetic symbols corresponding to the phoneme segment pair sequence candidate selected in the speech candidate selection step in the recognized vocabulary dictionary,
The accumulated cost calculating step includes a step of calculating an accumulated cost based on the occurrence probability of the phoneme segment pair in the phoneme segment pair sequence and the connection probability of the connected phoneme segment pair sequence. Recognition vocabulary dictionary creation program.   A speech recognition program that causes a computer to execute processing including a step of recognizing speech based on a recognized vocabulary dictionary created by the recognized vocabulary dictionary creating program according to claim 5. A text phonetic symbol conversion dictionary creation method for creating a text phonetic symbol conversion dictionary used for conversion from text to phonetic symbols,
Data including a word , segment division information obtained by dividing the word into segments, and pronunciation symbols for each segment are stored as learning data, and the word is a word obtained by connecting a plurality of character strings that can be regarded as one word. In this case, from learning data storage means for further storing segment division information obtained by dividing the word connected to a plurality of character strings into segments corresponding to the character strings, and phonetic symbols for each segment corresponding to the character strings Learning data acquisition step for acquiring the word, the segment division information, and the phonetic symbol for each segment;
A phoneme segment pair that is a set of a segment name and a phonetic symbol corresponding to the name of the segment is generated from the segment division information acquired by the learning data acquisition step and the phonetic symbol for each segment, and the learning An occurrence probability calculating step for calculating an occurrence probability based on the frequency of occurrence of the phoneme segment pair in the data;
A concatenated phoneme segment pair sequence that is a sequence of phoneme segment pairs concatenated within a word is
A connection probability calculation step of calculating a connection probability based on the frequency of appearance in the learning data;
A text phonetic symbol for storing a text phonetic symbol conversion dictionary including the occurrence probability for each phoneme segment pair calculated by the occurrence probability calculating step and the connection probability for each connected phoneme segment pair sequence calculated by the connection probability calculating step A conversion dictionary saving step;
A method for creating a text phonetic symbol conversion dictionary, comprising: A recognition vocabulary dictionary creating method for creating a recognition vocabulary dictionary used for recognizing speech,
A text phonetic symbol conversion dictionary acquisition step for acquiring the text phonetic symbol conversion dictionary from the text phonetic symbol conversion dictionary storage means for storing the text phonetic symbol conversion dictionary created by the text phonetic symbol conversion dictionary creating method according to claim 7; ,
A recognition vocabulary dictionary storage step for storing a recognition vocabulary dictionary in which recognition vocabulary to be recognized by speech recognition is registered;
A recognition vocabulary acquisition step of acquiring a recognition vocabulary registered in the recognition vocabulary dictionary;
A segment sequence generation step of dividing the acquired recognition vocabulary into segments by referring to the text phonetic symbol conversion dictionary acquired in the text phonetic symbol conversion dictionary acquisition step, and generating a segment sequence which is a sequence of segments;
A phoneme segment pair sequence that is a sequence of phoneme segment pairs is generated from the segment sequence generated by the segment sequence generation means, and the accumulated cost is calculated for each phoneme segment pair sequence by referring to the text phonetic symbol conversion dictionary A cumulative cost calculating step,
A speech candidate selection step of selecting, as a phoneme segment pair sequence candidate, an upper phoneme segment pair sequence from the phoneme segment pair sequence based on the accumulated cost calculated in the accumulated cost calculating step;
A speech candidate registration step of registering a sequence of phonetic symbols corresponding to the phoneme segment pair sequence candidate selected in the speech candidate selection step in the recognized vocabulary dictionary,
The accumulated cost calculating step calculates an accumulated cost based on the occurrence probability of the phoneme segment pair in the phoneme segment pair sequence and the connection probability of the connected phoneme segment pair sequence. How to make.   9. A speech recognition method comprising the step of recognizing speech based on a recognized vocabulary dictionary created by the recognized vocabulary dictionary creating method according to claim 8.

Images