JP2014206642A - Voice recognition device and voice recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce erroneous recognition among keywords.SOLUTION: A voice recognition device (1) includes: a similarity word database (DB) (203) for pre-storing similarity word information related to keywords mutually similar acoustically; a determination part (106) for referring to the similarity word information to determine whether there is a second word similar to a first word estimated by a first recognition process part (100); an extraction part (108) for extracting a partial voice signal in a specific section predetermined from a voice signal if determined that there is the second word; and a second recognition part (110) that includes a determination part (114) for performing determination processing in which a keyword having a higher degree of likelihood between the first and second words is regarded as a recognition result, on the basis of: a feature amount of the partial voice signal; and a second model parameter (202) corresponding to the respective specific sections of the first and second words.

Description

  The present invention relates to a speech recognition apparatus and a speech recognition program, and more particularly to a speech recognition apparatus and speech recognition program that perform speech recognition using an isolated word recognition method.

  Conventionally, there is a technique for recognizing words that are acoustically similar to each other.

  Japanese Patent Laid-Open No. 10-274994 (Patent Document 1) discloses a process for recognizing similar words such as “Nanji” and “Nando” in a speech recognition technique using a DRNN (Dynamic Recurrent Neural Networks) word model. It is disclosed. Specifically, for example, among the DRNN output corresponding to “Nanji”, a section t1 including a DRNN output corresponding to a characteristic part of a similar word (phonetic part of “ji” and “do”) is set, and a section t1 is set. Describes what kind of vowels exist.

JP-A-10-274994

  In the above document, when a special acoustic model called DRNN word model is used, a similar word that is similar to a recognition target word (keyword) and is not a recognition target may have a certain degree of certainty. To deal with, the vowels of the word features are examined.

  On the other hand, in speech recognition using a general acoustic model such as HMM (Hidden Markov Model), the recognition rate of a certain keyword (registered word) is registered as a word that is acoustically similar to the keyword. There is a tendency for the former to be lower in cases where it is not registered and cases where it is not registered. Therefore, in speech recognition using a general acoustic model, reducing misrecognition between keywords leads to an improvement in the overall recognition rate.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a speech recognition apparatus and a speech recognition program capable of reducing erroneous recognition between keywords.

  A speech recognition apparatus according to an aspect of the present invention executes a recognition process based on a feature amount of a speech signal and a first model parameter for each of a plurality of keywords, thereby obtaining a first word from the plurality of keywords. By referring to the first recognition processing means for estimating, the storage unit for storing similar word information about acoustically similar keywords in advance, and the similar word information, the first recognition processing means When a determination means for determining whether or not a second word that is a keyword similar to the estimated first word exists, and when the determination means determines that the second word exists, an audio signal The recognition process is executed using the extraction means for extracting a partial speech signal in a specific section determined in advance and the feature amount of the partial speech signal extracted by the extraction means. And a second recognition processing means. The second recognition processing means, based on the feature amount of the partial speech signal and the second model parameter corresponding to the specific section for each of the first word and the second word, A determination unit configured to execute a determination process using a keyword having a higher likelihood of the two words as a recognition result;

  Preferably, the similar word information includes identification information about a similar keyword for each specific keyword that may be erroneously recognized as an acoustically similar keyword, and the storage means includes a specific keyword for each specific keyword. Section information defining the section is further stored. The extraction unit divides the audio signal into a plurality of sections using a predetermined algorithm, and extracts the partial sound signal based on the plurality of divided sections and the section information.

  Preferably, the first recognition processing unit includes a first analysis unit that calculates a feature amount of the voice signal by cutting out the voice signal in units of frames of the first time length and performing analysis for each frame. The second recognition processing means extracts a partial speech signal in units of frames having a second time length shorter than the first time, and performs analysis for each frame to calculate a feature amount of the partial speech signal. Means are further included.

  Preferably, an output means for outputting the keyword determined as the recognition result by the determination means is further provided, and the output means performs the first recognition when the determination means determines that the second word does not exist. The first word estimated by the processing means is output as a recognition result.

  Preferably, when the determination unit determines that there are a plurality of keywords similar to the first word, the determination unit selects the keyword with the higher likelihood in the recognition process in the first recognition processing unit. The determination process is executed as the second word.

  Preferably, when the determination unit determines that there are a plurality of keywords similar to the first word, the determination unit uses the second model parameter for each of the first word and the plurality of similar keywords. The keyword having the highest likelihood among the first word and a plurality of similar keywords is determined as the recognition result.

  Preferably, when the determination unit determines that there are a plurality of keywords similar to the first word, the determination unit uses a plurality of similar keywords using the second model parameter for each of the plurality of similar keywords. The keyword having the higher likelihood is determined, and the keyword having the higher likelihood is determined as the second word to execute the determination process.

  Preferably, the specific section is determined as a section having the highest recognition rate between keywords that are acoustically similar to each other by a simulation during learning.

  A speech recognition program according to another aspect of the present invention executes a recognition process based on a feature amount of a speech signal and a first model parameter for each of the plurality of keywords, so that the first word from the plurality of keywords is obtained. Whether or not there is a second word that is a keyword similar to the estimated first word by referring to the pre-stored similar word information for the keywords that are acoustically similar to each other. A step of determining whether or not, a step of extracting a partial speech signal in a predetermined specific section from the speech signal when it is determined that the second word is present, and a feature amount of the extracted partial speech signal And a second model parameter corresponding to a particular interval for each of the first word and the second word, the first word and the second And a step of executing the determination processing of the recognition result keyword higher likelihood among word to the computer.

  According to the present invention, erroneous recognition between keywords can be reduced.

It is a block diagram which shows the hardware structural example of the speech recognition apparatus which concerns on embodiment of this invention. It is a functional block diagram which shows the function structure of the speech recognition apparatus which concerns on embodiment of this invention. It is a figure which shows the data structure example of the similar word database in embodiment of this invention. It is a flowchart which shows the speech recognition process in embodiment of this invention. It is a figure which shows the specific example of a difference area. It is a figure which shows the specific example of a difference area.

  Embodiments of the present invention will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals and description thereof will not be repeated.

  The speech recognition apparatus according to the present embodiment employs an isolated word recognition method and analyzes a speech signal to estimate and output a word represented by the speech signal from a plurality of keywords. In this embodiment, the “keyword” represents a registered word, that is, a word to be recognized.

  The speech recognition apparatus according to the present embodiment uses an HMM as an acoustic model and calculates the likelihood by, for example, a Viterbi algorithm at the time of recognition. According to the Viterbi algorithm, when the likelihood is calculated, the state of the HMM (state number) is associated with the speech section. By the association by the Viterbi algorithm (hereinafter referred to as “Viterbi alignment”), the speech signal is divided into at least a phoneme section and an articulation connection section, and a state number is assigned to each section. Note that an acoustic model other than the HMM such as DTW (Dynamic Time Warping) may be used as long as it is an acoustic model capable of aligning audio signals by Viterbi alignment or the like.

  Hereinafter, the configuration and operation of the speech recognition apparatus according to the present embodiment will be described in detail.

<About configuration>
(Hardware configuration)
The speech recognition apparatus according to the present embodiment can be realized by a general-purpose computer such as a PC (Personal Computer).

  FIG. 1 is a block diagram showing a hardware configuration example of a speech recognition apparatus 1 according to an embodiment of the present invention. Referring to FIG. 1, a speech recognition apparatus 1 stores a CPU (Central Processing Unit) 11 for performing various arithmetic processes, a ROM (Read Only Memory) 12 for storing various data and programs, work data, and the like. Random Access Memory (RAM) 13, a hard disk 14 which is a nonvolatile storage device, an operation unit 15 including a keyboard, a display unit 16 for displaying various information, and data and programs from the recording medium 17a Drive device 17 capable of reading and writing data, and a communication I / F (interface) 18 for Internet communication. The recording medium 17a may be, for example, a CD-ROM (Compact Disc-ROM) or a memory card.

  The voice recognition device 1 may further include an input unit 19 for inputting a voice signal from the microphone 20. When the speech recognition apparatus 1 does not have the input unit 19, recognition processing is performed on, for example, a speech signal obtained from the communication I / F 18 or a speech signal read from the recording medium 17a.

(Functional configuration)
FIG. 2 is a functional block diagram showing a functional configuration of the speech recognition apparatus 1 according to the embodiment of the present invention. With reference to FIG. 2, the speech recognition apparatus 1 includes a primary recognition processing unit 100, primary HMM data 201, and an output unit 116, similarly to a general speech recognition apparatus. The primary HMM data 201 corresponds to each of all keywords and is used for primary recognition by the primary recognition processing unit 100. Each primary HMM is a model parameter generated from the entire keyword speech. In this embodiment, “primary HMM” is used to distinguish it from the HMM (secondary HMM data 202) used in the secondary recognition processing unit 110. Each primary HMM is associated with an identification number.

  The primary recognition processing unit 100 performs recognition processing based on the feature amount of the audio signal and the primary HMM data 201. The primary recognition processing unit 100 includes an analysis unit 102 and an estimation unit 104 as its functions. The analysis unit 102 extracts the audio signal in units of frames having a first time length, and calculates the feature amount by analyzing the audio signal in units of frames. For example, the extracted audio signal is converted into an MFCC (Mel-frequency cepstral coefficient) feature quantity. The estimation unit 104 calculates a likelihood that each primary HMM generates a sequence of calculated feature values, and estimates a keyword indicated by the primary HMM having the highest likelihood as a recognition result. The keyword estimated here is hereinafter also referred to as “estimated word”.

  The output unit 116 outputs the recognition result. The output unit 116 is realized by the display unit 16, for example.

  In a general speech recognition apparatus, a recognition result (primary recognition result) in the primary recognition processing unit 100, that is, an estimated word is output as it is. However, when a keyword that is acoustically similar to the estimated word is registered, the possibility of misrecognition increases as ambient noise increases.

  Therefore, the speech recognition apparatus 1 according to the present embodiment further includes a determination unit 106, an extraction unit 108, and a secondary recognition processing unit 110 as its functions. In addition to the primary HMM data 201, for example, the hard disk 14 stores secondary HMM data 202 and a similar word database (DB) 203.

  The secondary HMM data 202 corresponds to each of a plurality of keywords that may be erroneously recognized, and is a model parameter generated from the speech in a specific section of the speech of the keyword. Each secondary HMM is associated with an identification number. This “specific section” is a section in which the similarity between keywords that are acoustically similar to each other is high, that is, a section in which a feature amount is likely to be different. In that sense, a specific section can also be expressed as a “difference section”. During learning, the state (state number) of the HMM constituting the section from which the speech signal is extracted is changed in various ways, and the section with the highest recognition rate between similar keywords is determined as a difference section by calculation simulation. . A secondary HMM is generated from the audio signal in the difference section. In the present embodiment, the audio signal in the difference section is referred to as “partial audio signal”.

  The similar word database 203 stores information on keywords that are acoustically similar to each other (hereinafter referred to as “similar word information”). The similar word information is also referred to as information about a keyword that may be erroneously recognized. The similar word information includes, for example, an identification number of the primary HMM as identification information for similar keywords at least for each keyword that may be erroneously recognized. Further, the similar word database 203 stores section information about a difference section to be used for secondary recognition and secondary HMM identification information for each keyword that may be erroneously recognized. An example of the data structure of the similar word database 203 will be described later.

  The determination unit 106 refers to the similar word information in the similar word database 203 to determine whether or not there is a keyword similar to the estimated word (hereinafter referred to as “similar word”). That is, it is determined whether or not the estimated word as the primary recognition result has a possibility of erroneous recognition.

  When the determination unit 106 determines that a similar word exists, the extraction unit 108 refers to the section information in the similar word database 203 to extract a partial speech signal in the difference section from the entire speech signal. Specifically, the voice signal is divided into a plurality of sections by the Viterbi algorithm, and the partial voice signal is extracted based on the plurality of divided sections and section information.

  The secondary recognition processing unit 110 performs recognition processing using the feature amount of the partial speech signal extracted by the extraction unit 108. The secondary recognition processing unit 110 includes an analysis unit 112 and a determination unit 114 as its functions. The analysis unit 112 cuts out the partial audio signal in units of frames having a second time length shorter than the first time, and calculates the feature amount by analyzing the partial audio signals in units of frames. Based on the calculated feature amount of the partial speech signal and the secondary HMM for each of the estimated word and the similar word, the determination unit 114 uses the keyword having the higher likelihood of the estimated word and the similar word as a recognition result. judge. Specifically, each secondary HMM calculates the likelihood of generating a sequence of calculated feature values, and recognizes the keyword that is the source of the secondary HMM with the higher likelihood (secondary recognition result). Judge as.

  Here, the 1st time length used for primary recognition is predetermined in the range of 20-30 ms similarly to the general speech recognition method, for example. The second time length used for the secondary recognition may be, for example, about half of the first time length and is predetermined in the range of 10 to 15 ms. In general, the first time length is determined so that several vowel periods can be included. On the other hand, keywords similar to each other often have different features in the consonant part other than the vowel and the articulation combination part, and the time length of these parts is shorter than the time length of the vowel. By making it shorter than the first time length, the recognition rate between similar keywords can be improved. Note that, in both cases of primary recognition and secondary recognition, the positions of the frames are cut out by shifting so that adjacent frames overlap each other. As for the time length for shifting the frame, it is desirable to shorten the time for secondary recognition than for primary recognition.

  The output unit 116 outputs either the primary recognition result or the secondary recognition result. When there is no keyword similar to the estimated word, the primary recognition result, that is, the estimated word is output. On the other hand, when a keyword similar to the estimated word exists, the secondary recognition result, that is, the estimated word or similar word is output.

  The functional blocks other than the output unit 116 shown in FIG. 2 may be realized by the CPU 11 shown in FIG. 1 executing software stored in the ROM 12, for example, and at least one of these may be implemented It may be realized by hardware.

  Here, a data structure example of the similar word database 203 will be described.

(Data structure example)
FIG. 3 is a diagram showing an example of the data structure of the similar word database 203 in the embodiment of the present invention. Referring to FIG. 3, the similar word database 203 includes a plurality of rows and a plurality of columns 31 to 45. In the present embodiment, it is assumed that each of a plurality of rows corresponds to all keywords.

  In columns 31 and 32, keywords (reading pseudonyms) and primary HMM numbers are recorded as information of each keyword itself. In columns 33 to 36 and 41, as similar word information, the number of similar words (column 33), reading pseudonyms of similar words A and B (columns 34 and 35), and primary HMMs of similar words A and B, respectively. Numbers (columns 36 and 41) are stored. For example, when the keyword is “kyu”, it is stored that there are two similar words “ju” (similar word A) and “chu” (similar word B). Of these, the columns 31, 34, and 35 relating to the reading of the keyword are provided for easy understanding, and these may be omitted.

  Further, in the similar word database 203, for each similar word to be compared, the start state number (columns 37 and 42) and the end state number (columns 38 and 43) are stored as section information, and as identification information of the secondary HMM, Two secondary HMM numbers (columns 39 and 40 and columns 44 and 45) are stored. The data 360 in the columns 37 to 40 is used when determining the estimated word and the similar word A. The data 410 in the columns 42 to 45 is used when determining the estimated word and the similar word B.

  In the present embodiment, the similar word database 203 includes not only the similar word information but also the section information and the identification information of the secondary HMM, but these may be stored in another database. Also, similar word information for all keywords is included in the similar word database 203, and the presence or absence of similar words is also stored in the column 33 (number of similar words). However, only keywords that may be erroneously recognized are stored. Similar word information may be stored.

<About operation>
(Voice recognition processing)
FIG. 4 is a flowchart showing voice recognition processing in the embodiment of the present invention. The processing procedure shown in the flowchart of FIG. 4 is stored in advance in the ROM 12 as a program, and the function of voice recognition processing is realized by the CPU 11 reading and executing the program.

  Referring to FIG. 4, when an audio signal is input from input unit 19 (step S (hereinafter abbreviated as “S”) 2), the input audio signal is stored in RAM 13 in time series, for example. It is assumed that the voice signal input in S2 includes a human voice. The analysis unit 102 of the primary recognition processing unit 100 cuts out a frame from the stored voice signal (S4). That is, the audio signal is cut out in units of 25 ms frames, for example. The frames are cut out by shifting, for example, by 10 ms so that adjacent frames overlap each other.

  When the frame is cut out, the analysis unit 102 calculates the feature amount of the audio signal for each frame (S6).

  Next, the estimation unit 104 estimates a word (keyword) represented by the audio signal based on the primary HMM data 201 from the feature amount calculated in S6 (S8). Specifically, first, each primary HMM determines the likelihood of generating a sequence of calculated feature values. Thereafter, the likelihood values of the respective primary HMMs are compared, and a keyword corresponding to the primary HMM having the maximum likelihood is set as a primary recognition result.

  When the primary recognition processing is completed, the determination unit 106 refers to the similar word database 203 and determines whether there is a similar word in the estimated word as the recognition result (S10). Specifically, the determination unit 106 refers to the row of the primary HMM number (column 32) of the keyword having the maximum likelihood in S8 in the similar word database 203, and the column 33 of “number of similar words”. It is determined whether or not “1” or “2” is recorded. If it is determined that a similar word exists (YES in S10), the process proceeds to S12. On the other hand, when it is determined that there is no similar word (NO in S10), the output unit 116 outputs the keyword estimated in S8 as a formal recognition result (S20).

  In S <b> 12, the extraction unit 108 reads the section information from the similar word database 203, and extracts the voice signal in the difference section, that is, the partial voice signal from the voice signal stored in the RAM 13. When there is one similar word, the extraction unit 108 reads the starting state number A (column 37) and the ending state number A (column 38) because the keyword is “similar word A” in the similar word database 203. When there are two similar words, the start state number and the end state number of the similar word having the higher likelihood in the primary recognition process are read.

  The extraction unit 108 performs Viterbi alignment of the entire speech signal with the primary HMM of the estimated word. And the partial audio | voice signal of the difference area divided by the read start end state number and termination | terminus state number is extracted. In this embodiment, the speech signal is Viterbi-aligned with the primary word HMM of the estimated word. However, during learning, a secondary HMM can be generated by Viterbi alignment with the primary HMM of the similar word. For example, Viterbi alignment may be performed using a similar primary HMM.

  When the partial audio signal is extracted, the analysis unit 112 cuts out a 10 ms frame, for example, from the partial audio signal (S14). Also in this case, the frames are cut out with a shift of 5 ms, for example, so that adjacent frames overlap each other.

  When the frame is cut out, the analysis unit 112 calculates the feature amount of the audio signal for each frame (S16). The determination unit 114 determines a recognition result based on the secondary HMM data 202 of the estimated word and the similar word based on the feature amount (S18). That is, it is determined whether the keyword represented by the input voice signal is an estimated word or a similar word. Specifically, the determination unit 114 obtains a likelihood that the two secondary HMMs generate the feature amount series calculated in S16. Then, the keyword that is the basis of the secondary HMM with the higher likelihood is determined as a formal recognition result. The official recognition result is output by the output unit 116 (S20). Thereby, the voice recognition process is terminated.

  The above speech recognition processing will be described in more detail with a specific example.

(Concrete example)
For example, it is assumed that the estimated word obtained by the primary recognition process (S4 to S8) is “kyu”. Further, it is assumed that the primary HMM of “Kyu” has 11 states.

  Since the primary HMM number of “kyu” is “9”, the determination unit 106 accesses the row in which “9” is recorded in the column 32 of the similar word database 203 shown in FIG. Referring to column 33 in that row, since there are two similar words (YES in S10), secondary recognition processing is executed. Here, although there are two similar words “10” and “chu”, in this embodiment, secondary recognition processing is performed for similar words and estimated words having higher likelihood in the primary recognition processing. Run. In this specific example, it is assumed that “10” has a higher likelihood than “chu”.

  When the extraction unit 108 searches for a column that matches the HMM number of “10”, the column 36 of “the primary HMM number A of similar words” is known to be the column. Therefore, “2” and “6” are read as the start state number A (column 37) and the end state number A (column 38), respectively. The extraction unit 108 performs Viterbi alignment of the speech signal with the primary HMM of the estimated word “Kyu”, and extracts the speech signal of the difference section delimited by states 2 to 6 among states 1 to 11 (S12). .

  Here, with reference to FIG. 5 and FIG. 6, a difference area is demonstrated concretely. In the upper part of FIG. 5, an example in which the “Kyu” audio signal V <b> 1 is Viterbi-aligned with the “Kyu” primary HMM is shown. The lower part of FIG. 5 shows an example in which the partial audio signal V2 in the difference section is extracted from the entire audio signal of “kyu”. In the upper part of FIG. 6, an example of Viterbi alignment of the voice signal VA <b> 1 of “10” with the primary HMM of “Kyu” is shown. The lower part of FIG. 6 shows an example in which the partial audio signal VA2 in the difference section is extracted from the entire “10” audio signal.

  Each of the audio signals V1 and VA1 shown in FIGS. 5 and 6 is divided into 11 states by the “Kyu” primary HMM. When determining “10” and “10”, only the partial audio signals V2 and VA2 in the difference sections 50 and 60 of the state numbers 2 to 6 are used for the secondary recognition.

  When the partial speech signal in the difference section is extracted, the analysis unit 112 analyzes the partial speech signal and obtains the feature amount in the difference section (S14, S16). The determination unit 114 refers to the columns 39 and 40 in which the secondary HMM numbers A-1 and A-2 are recorded, and acquires the parameters of the corresponding two secondary HMMs (9109 and 9110). The parameter whose secondary HMM number is “9109” is an HMM created for the partial speech signal in the difference section (state numbers 2 to 6) of the teacher speech of “kyu” at the time of learning. The parameter whose secondary HMM number is “9110” is an HMM created for the partial speech signal in the difference section (state numbers 2 to 6) of the teacher speech of “10” at the time of learning. Note that the feature amount calculated for each frame having the second time length is also used during learning.

  The determination unit 114 calculates the likelihood of generating a feature amount series in the difference interval for the two secondary HMMs. When the likelihood is calculated, the keyword number that is the basis of the secondary HMM with the higher likelihood is output as a recognition result using the secondary HMM, and this is used as the final result. For example, when the secondary HMM with the secondary HMM number “9110” has a higher likelihood, the recognition result is determined as “10” (S18).

  In the case where the determination is based on “kyu” and “chu”, “3” and “6” are stored in the similar word database 203 when the columns 42 and 43 in the row of “kyu” are viewed. This indicates that, in order to determine “kyu” and “chu”, the difference section is a section divided by state numbers 3 to 6.

  As described above, according to the present embodiment, even if there is a possibility of erroneous recognition in the primary recognition processing using the primary HMM, keywords that are acoustically similar are re-recognized. Therefore, misrecognition between keywords can be reduced. As a result, the overall recognition rate can be improved.

  In this embodiment, the secondary recognition is performed if there is a keyword similar to the estimated word regardless of the likelihood of the estimated word in the primary recognition. For example, the estimation in the primary recognition is performed. Only when the likelihood of a word is less than or equal to a predetermined value, the processing subsequent to the determination of whether a similar keyword exists (S10) may be performed. Or after performing secondary recognition, you may consider the likelihood of the presumed word in primary recognition.

<Modification>
In the above embodiment, when there are two similar words, the secondary recognition is performed for the similar word having the higher likelihood in the primary recognition, but the secondary recognition is also performed including the similar word having the lower likelihood. May be performed.

  For example, when the primary recognition result is “kyu”, after extracting the partial speech signal of the difference section based on the HMM of “kyu”, from the feature amount of the partial speech signal, “kyu”, “ju” and “ The likelihood of each secondary HMM may be compared, and the word that is the source of the secondary HMM with the highest likelihood may be determined as a recognition result. In this case, it is assumed that three secondary HMM numbers to be used in secondary recognition are recorded in the similar word database 203.

  Alternatively, when the primary recognition result is “kyu”, first, it is determined which one of “ju” and “chu” between similar words has the highest likelihood, and then the one with the higher likelihood It may be determined which of the similar words and “kyu” has the highest likelihood. In this case, the similar word database 203 shown in FIG. 3 can be used as it is.

  Specifically, first, the determination unit 106 accesses a line in which “10”, which is an HMM number of “10”, is recorded in the column 32 of the similar word database 203 illustrated in FIG. When the extraction unit 108 searches for a column that matches the HMM number of “Chu”, the column 41 of “primary HMM number B of similar words” is found to be that column. Therefore, “3” and “5” are read as the start state number B (column 42) and the end state number B (column 43), respectively. The extraction unit 108 performs Viterbi alignment of the audio signal with the primary HMM of “10”, and extracts the audio signal in the difference section divided by the states 3 to 5 (S12).

  When the partial speech signal in the difference section is extracted, the analysis unit 112 analyzes the partial speech signal and obtains the feature amount in the difference section (S14, S16). The determination unit 114 refers to the columns 44 and 45 in which the secondary HMM numbers B-1 and B-2 are recorded, and acquires the parameters of the corresponding two secondary HMMs (10210 and 10212). The parameter whose secondary HMM number is “10210” is an HMM created for the partial speech signal in the difference section (state numbers 3 to 5) of the teacher speech of “10” during learning. The parameter whose secondary HMM number is “10212” is an HMM created for the partial speech signal in the difference section (state numbers 3 to 5) of the teacher speech of “Chu” during learning.

  The determination unit 114 calculates the likelihood of generating a feature amount series in the difference interval for the two secondary HMMs. When the likelihood is calculated, the processing after S12 is performed again for the similar word (“ju” or “chu”) that is the basis of the higher-order secondary HMM and the estimated word “kyu”. Done. Since the secondary recognition process of the similar word with the higher likelihood and the estimated word is the same as described above, detailed description will not be repeated.

  In the present embodiment, it has been described that the number of keywords similar to the estimated word is two or less.

  The speech recognition method executed by the speech recognition apparatus 1 according to the present embodiment can also be provided as a program. Such a program can be provided by being recorded on an optical medium such as a CD-ROM (Compact Disc-ROM) or a computer-readable non-transitory recording medium such as a memory card. A program can also be provided by downloading via a network.

  The program according to the present invention is a program module that is provided as a part of a computer operating system (OS) and calls necessary modules in a predetermined arrangement at a predetermined timing to execute processing. Also good. In that case, the program itself does not include the module, and the process is executed in cooperation with the OS. A program that does not include such a module can also be included in the program according to the present invention.

  The program according to the present invention may be provided by being incorporated in a part of another program. Even in this case, the program itself does not include the module included in the other program, and the process is executed in cooperation with the other program. Such a program incorporated in another program can also be included in the program according to the present invention.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

  DESCRIPTION OF SYMBOLS 1 Voice recognition apparatus, 11 CPU, 12 ROM, 13 RAM, 14 Hard disk, 15 Operation part, 16 Display part, 17 Drive apparatus, 18 Communication I / F, 19 Input part, 20 Microphone, 50, 60 Difference area, 100 1 Next recognition processing unit, 102 analysis unit, 104 estimation unit, 106 determination unit, 108 extraction unit, 110 secondary recognition processing unit, 112 analysis unit, 114 determination unit, 116 output unit, 201 primary HMM data, 202 secondary HMM Data, 203 similar words database.

Claims (9)

First recognition processing means for estimating a first word from the plurality of keywords by executing recognition processing based on the feature amount of the audio signal and the first model parameter for each of the plurality of keywords. When,
A storage unit for preliminarily storing similar word information about keywords that are acoustically similar to each other;
Judgment means for judging whether or not there is a second word that is similar to the first word estimated by the first recognition processing means by referring to the similar word information; ,
An extracting means for extracting a partial speech signal in a predetermined specific section from the speech signal when the judging means judges that the second word is present;
Second recognition processing means for executing recognition processing using the feature amount of the partial speech signal extracted by the extraction means;
The second recognition processing means, based on the feature amount of the partial speech signal and the second model parameter corresponding to the specific section for each of the first word and the second word, A speech recognition apparatus including a determination unit that executes a determination process using a keyword having a higher likelihood of the first word and the second word as a recognition result. The similar word information includes identification information about the similar keyword for each specific keyword that may be erroneously recognized as an acoustically similar keyword,
The storage means further stores section information defining the specific section for each specific keyword,
The said extraction means divides | segments the said audio | voice signal into a some area with a predetermined | prescribed algorithm, The said partial audio | voice signal is extracted based on these divided | segmented some area and the said area information. Voice recognition device. The first recognition processing means includes first analysis means for cutting out the voice signal in units of a frame having a first time length and calculating a feature value of the voice signal by performing analysis for each frame,
The second recognition processing means calculates the feature amount of the partial speech signal by cutting out the partial speech signal in units of frames having a second time length shorter than the first time and performing analysis for each frame. The speech recognition apparatus according to claim 1, further comprising: a second analysis unit that performs the analysis. An output unit for outputting the keyword determined as the recognition result by the determination unit;
The output means outputs the first word estimated by the first recognition processing means as the recognition result when the determination means determines that the second word does not exist. Item 4. The speech recognition device according to any one of Items 1 to 3.   When the determination unit determines that there are a plurality of keywords similar to the first word, the determination unit selects a keyword having a higher likelihood in the recognition process in the first recognition processing unit. The speech recognition apparatus according to claim 1, wherein the determination process is executed as the second word.   When the determination unit determines that there are a plurality of keywords similar to the first word, the determination unit determines the second model parameter for each of the first word and a plurality of similar keywords. The speech recognition apparatus according to claim 1, wherein a keyword having the highest likelihood among the first word and the plurality of similar keywords is determined as the recognition result.   When the determination unit determines that there are a plurality of similar keywords to the first word, the determination unit uses the second model parameter for each of the plurality of similar keywords to determine the plurality of similarities. The speech recognition device according to claim 1, wherein a keyword having a higher likelihood among keywords to be determined is determined, and the determination process is performed using the keyword having a higher likelihood as the second word. .   The speech recognition apparatus according to any one of claims 1 to 7, wherein the specific section is determined as a section having a highest recognition rate between keywords that are acoustically similar to each other by a simulation during learning. Estimating a first word from the plurality of keywords by performing a recognition process based on a feature amount of the audio signal and a first model parameter for each of the plurality of keywords;
It is determined whether or not there is a second word that is a keyword similar to the estimated first word by referring to the similar word information about the keywords that are acoustically similar to each other stored in advance. And steps to
Extracting a partial speech signal in a predetermined specific section from the speech signal when it is determined that the second word is present;
Based on the extracted feature amount of the partial speech signal and the second model parameter corresponding to the specific section for each of the first word and the second word, the first word and the The speech recognition program which makes a computer perform the step which performs the determination process which uses as a recognition result the keyword with a higher likelihood among 2nd words.

Images