JP6549064B2 - Speech recognition device, speech recognition method, program - Google Patents

Description

  The present invention relates to a speech recognition apparatus, a speech recognition method, and a program.

  Patent Document 1 discloses a document summarizing apparatus capable of flexibly coping with the limitation of the number of characters. The document summarizing device of Patent Document 1 includes a sentence shortening device, a sentence score determining device, and a sentence selecting device. The sentence shortening apparatus shortens the sentences in the input document at a plurality of designated shortening rates and outputs a text and a short sentence. The sentence score determination device refers to the word score database and the sentence shortening rate, the appearance position information of the sentence, and the position information score obtained by the input parameter with respect to the original sentence and the short sentence output by the sentence shortening device. The sentence score is determined on the basis of the word score obtained as a measure indicating the weight of the words constituting the sentence. The sentence selection device selects a combination of sentences having the largest sum of sentence scores determined by the sentence score determination device as a summary under the input character number restriction.

JP, 2010-55236, A

  Speech recognition is a technology for converting speech into text, but under all circumstances it is currently difficult to perform speech recognition with a conversion rate of 100%, and in many cases erroneous words will be included in the recognition result. In addition, the recognition result may include redundant phrases that do not need to be converted into text. For example, the recognition result "This is difficult so it may be" may be shortened to the expression "this is difficult" by deletion of a redundant phrase.

  In this way, correcting the error of the speech recognition result and deleting the unnecessary part is an essential process to obtain high-quality recognition result. A process of improving the above two points with respect to the recognition result and improving the readability of the recognition result and the ease of application of the language processing in the latter stage will be referred to as “recognition result shaping technology”.

  Normal speech recognition is performed by tuning an acoustic model and a language model to a form optimum for the target speech and then decoding (real-time text conversion using both models). The above two models are models that use only information on the probability of sound, the number of languages, and the arrangement of words. The above two models can not use more information.

  On the other hand, in discriminative reranking, correction of recognition results is performed using long-distance information (such as certainty in sentence units) that can not be considered in a normal language model, using a pair of speech recognition results and correct data Research exists. However, in the discriminative reranking method, correct data for speech recognition must be manually prepared, and the problem is that this cost is high.

  Therefore, it is an object of the present invention to provide a speech recognition apparatus capable of correcting speech recognition results without using correct answer data.

  The speech recognition apparatus of the present invention includes a speech recognition unit, a 3 gram calculation unit, a tf-idf calculation unit, an importance calculation unit, and an unnecessary word deletion unit. Note that N is an integer of 2 or more.

  The speech recognition unit outputs the first to N-th speech recognition results based on the input speech data. The 3 gram calculating unit calculates the 3 gram probability of the text data prepared in advance. The tf-idf calculation unit is configured to calculate the tf of each word included in the first to N-th speech recognition results and the idf of each word included in the first speech recognition result among idf prepared based on text data. And calculate the tf-idf of each word included in the first speech recognition result. The degree-of-importance calculation unit calculates the NRD of each word included in the first-ranked speech recognition result based on tf-idf, and outputs a value based on the calculated NRD as the degree of importance of each word. The unnecessary word deletion unit is included in the reliability of each word included in the first-ranked speech recognition result, the 3-gram probability of three consecutive words included in the first-ranked speech recognition result, and the first-ranked speech recognition result The unnecessary words included in the first speech recognition result are deleted based on the solution of the integer programming problem formulated using the importance of each word.

  According to the speech recognition apparatus of the present invention, the speech recognition result can be corrected without using the correct answer data.

FIG. 1 is a block diagram showing the configuration of a speech recognition device according to a first embodiment. 3 is a flowchart showing the operation of the speech recognition device of the first embodiment. FIG. 7 is a block diagram showing the configuration of a speech recognition device according to a second embodiment. 6 is a flowchart showing the operation of the speech recognition device of the second embodiment. FIG. 7 is a block diagram showing the configuration of a speech recognition device according to a third embodiment. 10 is a flowchart showing the operation of the speech recognition device of the third embodiment. FIG. 7 is a block diagram showing the configuration of a speech recognition device according to a fourth embodiment. 10 is a flowchart showing the operation of the speech recognition device of the fourth embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. Note that components having the same function will be assigned the same reference numerals and redundant description will be omitted.

  The configuration and operation of the speech recognition apparatus 1 according to the first embodiment will be described below with reference to FIGS. 1 and 2. As shown in FIG. 1, the speech recognition apparatus 1 of this embodiment includes a corpus storage unit 10, a speech recognition unit 11, a 3 gram calculation unit 12, a tf-idf calculation unit 13, and an importance calculation unit 14. The unnecessary word deletion unit 15 is included. It is assumed that the corpus storage unit 10 stores text data prepared in advance.

  The speech recognition unit 11 outputs the first to N-th speech recognition results based on the input speech data (S11). The 3 gram calculating unit 12 calculates the 3 gram probability of the text data prepared in advance (S12). The tf-idf calculation unit 13 calculates tf of each word included in the first to N-th speech recognition results and each word included in the first speech recognition result among idf prepared in advance based on text data. Based on idf, tf-idf of each word included in the first speech recognition result is calculated (S13). The importance degree calculation unit 14 calculates NRD (Normalized Relevance Distance) of each word included in the first-ranked speech recognition result based on tf-idf, and outputs the calculated NRD value as the importance degree of each word To do (S14). The unnecessary word deletion unit 15 includes the reliability of each word included in the first-ranked speech recognition result, the 3-gram probability of three consecutive words included in the first-ranked speech recognition result, and the first-ranked speech recognition result The unnecessary words included in the first speech recognition result are deleted on the basis of the solution of the integer programming problem formulated using the importance of each word to be processed (S15).

The operation of each component will be described in detail below.
<Voice recognition unit 11>
Input: Audio data (time-series data of sound pressure. Format is pcm, wav etc.), upper limit N of the order number
Output: 1st to Nth sentences of speech recognition result, reliability of each word in each sentence Here, it is assumed that the inputted speech data is processed as a sentence for each one utterance.

[Example of sentences for each input utterance]
Example of the first utterance: "I ate curry curry today"
The speech recognition unit 11 acquires speech data as input, executes speech recognition by a general speech recognition method, and outputs speech recognition results from the first place to the Nth place (S11). However, as described above, N is an integer of 2 or more.

  In step S11, recognition results of multiple (= N) hypotheses having degrees of reliability and reliability for each word are output as follows for each sentence (sentences are separated based on time information) of speech data . Note that N may be specified manually. For example, N may be five.

  The multiple hypotheses of recognition results consist of multiple recognition results that include, among the speech recognition results, sentences evaluated as likely to be the first by the speech recognition system, and sentences cited as candidates other than those mentioned above during calculation of speech recognition. Show a hypothesis.

  The degree of reliability is a probability value indicating how much likelihood this word is correct for each word of the recognition result, and has a value of 0 or more and 1 or less.

[Example of reliability]
1st place: Today I ate curry curd Reliability: Today → 0.7 → → 0.5 → → 0.4 Curry → 0.5 → 0.7 Eat → 0.9
2nd place: Today I ate light curry Confidence: Today → 0.7 ate → 0.5 light → 0.35 Curry → 0.5 ate → 0.7 → 0.9
...
N: I ate Kan curry today
Confidence: Today → 0.7 → 0.5 Can → 0.2 Curry → 0.5 → 0.7 Eat → 0.9

<3 gram calculation unit 12>
Input: large amount of text data output: 3gram probability for text data The 3gram calculating unit 12 uses a large amount of text data (not a speech recognition result) prepared in advance. In the present embodiment, a large amount of text data is stored in advance in the corpus storage unit 10. For example, a newspaper article can be used as a large amount of text data. As a large amount of text data, there can be considered, for example, a corpus in which there are about 150,000 articles of about 30 sentences each.

The 3 gram calculating unit 12 calculates 3 gram probabilities for all sentences of a large amount of text data. The 3 gram probability represents the probability that three words appear in succession in a sentence. Step S12 will be described using a specific example. For example, suppose that we want to calculate a 3gram probability p (hot | today, today) in which a sequence of 3gram (today, hot) appears. When this is calculated, a sequence of three words that continue as "today, xxx" is searched for all text data, and the number is calculated. As a result, for example, it is assumed that only the following three patterns are found, and the number of each line is as follows.
(1) It is hot 100 today
(2) It is fine today 95
(3) Rarely today 5
In this case, the 3gram probability p (hot | today) is

Is calculated. 3gram calculation unit 12, for any of the words that appear in the text data, any three-row _w i _{considered, w j,} w _k conditional probability occurs _{p (w k | w i,} w j) to calculate the ( i, j, k are symbols representing arbitrary indexes, and also appear in the following mathematical expressions).

Here, there are also cases where arbitrary three lines w _i , w _j and w _k do not appear at least once in text data, and p (w _k | w _i , w _j ) can not be calculated directly. In such a case, it may be dealt with by a method called backoff. Backoff is a method of using 2gram or 1gram instead of 3gram in the above case. The backoff is disclosed, for example, in Reference Non-Patent Document 1.
(Reference Non-Patent Document 1: Kenji Kita, Junichi Sakurai, “Language and Computation (4) Probabilistic Language Model”, The University of Tokyo Press, November 1999, p. 67-69)

  Also, p (o | start) and p (end | o, o) are also calculated. p (o | start) is the probability that the word ○ appears immediately after the beginning of a sentence. p (end | ,,)) is the probability of becoming a tail end of the sequence of ,, 。. The beginning of the sentence and the end of the sentence are judged based on the line feed symbol in the text data.

<Tf-idf calculator 13>
Input: First to N speech recognition results, large amount of text data output: tf-idf for all words appearing in the first speech recognition result
tf-idf is an index calculated based on two indices of tf (term frequency, word occurrence frequency) and idf (inverse document frequency, inverse document frequency), and the importance of a word in a sentence is It is an indicator to represent.

  First, the same data as the large amount of text data used in step S12 is prepared. In the present embodiment, text data stored in advance in the corpus storage unit 10 may be diverted. As described above, the text data to be prepared in advance may be, for example, a newspaper article or a corpus having about 150,000 articles each having about 30 sentences.

  The method of calculating idf is described below. The number of documents of text data (a set of document segments, which is assumed to be attached to the text data in advance, for example, one article in a newspaper), is D, of which the number of documents in which the target word a appears Assuming that d, idf is calculated as log (D / d). The base of the logarithm is a positive real number greater than one. In the following example, the bottom is assumed to be 10.

For example, the total number of documents in the corpus storage unit 10 is 150000,
Number of documents where "today" appears ... 400
Number of documents where "ha" appears ... 300
Number of documents where "curry" appears-3000
The number of documents in which "o" appears ... 50000
Number of documents that "eat" appears ... 40000
Number of documents in which "ta" appears ... 50000
It is assumed that

In this case, the tf-idf calculator 13 calculates idf of each word as
Idf of "today" = log ₁₀ (150000/400) = 2.24
Idf = log ₁₀ (150,000 / 300) = 2.70 of "ha"
Idf of "curry" = log ₁₀ (150000/3000) = 1.70
Idf = log ₁₀ (150,000 / 50000) = 0.478 of "to"
"Eat" idf = log ₁₀ (150000/40000) = 0.57
Idf of the "ta" = log ₁₀ (150,000 / 50000) = 0.478
Calculate

  Next, the method of calculating tf will be described. Generally, tf is calculated as tf = A / M using the appearance frequency A of the word of interest a in the document when the total number of words in a document is M.

Therefore, the tf-idf calculation unit 13 regards a set of speech recognition results from the first place to the Nth place as one document, and calculates tf for each word included in at least the first speech recognition result. For example, assuming that the total number M of words in the set of speech recognition results from the first to the Nth is M = 1000 and the appearance frequency A of the word "curry" included in the first speech recognition result is 200,
"Curry" tf = 200/1000 = 0.20
Is calculated.

The tf-idf calculator 13 calculates tf-idf as follows using the values of idf and tf.
tf-idf = tf x idf
For example, tf-idf of "curry" in the above example is
tf-idf = 0.20 × 1.70 = 0.34
It is.

  The tf-idf calculation unit 13 may calculate tf-idf of each word included in the at least one speech recognition result. The important thing here is that although the tf calculation requires the first to N speech recognition results, as tf-idf, all the words appearing in the first to N speech recognition results are not necessarily required. It is not essential.

<Importance calculation unit 14>
Input: tf-idf of speech recognition result of first place, speech recognition result of first place
Output: NRD (Normalized Relevance Distance) of each word included in the first speech recognition result
The importance degree calculation unit 14 calculates NRD, which is a measure representing the similarity between two words in the first speech recognition result, based on tf-idf of the first speech recognition result (S14). The purpose of calculating NRD that represents the similarity between two words is, for example, if a word is highly similar to another word, that word is a correct word that is not misrecognized, and unnecessary words that should be deleted It is because it can be judged that there is not.

Here, in order to calculate NRD, first f _NRD (w) for each word and f _NRD (w ₁ , w ₂ ) between two words are calculated. They are,

Defined as TFIDF (w, d) represents tf-idf of the word w in the document d. S is the total number of documents.

The importance calculator 14 calculates, for example, the NRD of each word included in the first speech recognition result using the above-mentioned definition. For example, each word included in the first speech recognition result is w ₁ ,. . . , W _M and the word consistency score of any word w _{i in} this

That is calculated as the sum of the inverse of the NRD of all words and word w _i excluding i of up index j = 1 to M. As this score is higher, it is determined that the word is likely to be a correct word and is a word that is likely to be an unnecessary (necessary) word.

<Unnecessary word deletion unit 15>
Input: Speech recognition result of first place, reliability for each word of speech recognition result of first place, 3gram probability for each word of speech recognition result of first place, importance degree output for each word of speech recognition result of first place: unnecessary Speech recognition result in which a word is deleted The unnecessary word deletion unit 15 executes a process of deleting an unnecessary word from the speech recognition result of the first place. Now, the first speech recognition result of certain speech data is the words w ₁ , w ₂ ,. . . , W _T are assumed to be obtained. Unnecessary words from this sentence, NRD-based values (word consistency score, the importance of the word, the degree that should not be deleted) and the probability of connection of a three-word sequence (3 gram probability), Formulate into an integer programming problem using the degree of confidence (the degree to which the word is considered correct as a speech recognition result).

In the description of step S15, variables are defined. Each of δ _i , α _i , β _ij , and γ _ijk takes an integer value of 1 or 0, δ _i is defined for i = 1 to T, and 1 leaves the word w _i (does not delete it), If it is 0, it is a variable that represents deleting. α _i is a variable that is defined for i = 1 to T. If it is 1, it is a variable representing that the word w _i is the head word of the sentence and 0 if it is not the head word. β _ij is defined for all (i, j) combinations that satisfy 0 ≦ i <j ≦ T, 1 if the sequence immediately after w _i , w _j is the end of the sentence, 0 otherwise It is a variable to represent. γ _ijk is defined for all combinations of (i, j, k) that satisfy 0 ≦ i <j ≦ k ≦ T, and three consecutive words w _i , w _j and w _k are sentences after deletion If it exists in, it is defined as 0, if it does not exist.

This variable is used to calculate δ _i , α _i , β _ij , γ _ijk which become solutions maximizing the next function (integer programming problem).

Here, Sig (w _i ) is the importance (value based on NRD) of the word w _i , and p (w _k | w _i , w _j ) is a 3 gram in which the words w _i , w _j and w _k are three consecutive probability, q _{(w i)} is the reliability of the word _{w i.} The unnecessary word deletion unit 15 calculates the solution of this problem, and w ₁ , w ₂ ,. . . For each word w _i of, w _T, a word whose δ _i is 1 is output as it is, and a word (corresponding to an unnecessary word) whose δ _i is 0 is not output. Delete unnecessary words in sentences.

When the above evaluation function is maximized, the less important words are deleted. If any word is deleted and the arrangement of other words is a natural arrangement, the arbitrary word is deleted.
For example, it is assumed that the unnecessary words are deleted by this method for the sentence "I ate today's curry", which is the speech recognition result, and the sentence is compressed. In this case, “Karu” is assumed to be low in importance (low in value based on NRD). Then, if it is natural that “Curray” is natural (the 3 gram probability is high) in the sequence assuming that “Karu” is deleted, the corresponding “Karu” may be deleted. The above expression is a mathematical expression of this.

Therefore, the unnecessary word deletion unit 15 outputs the speech recognition result from which the unnecessary word has been removed using δ _i of the solution of the maximum value problem described above.

  The configuration and operation of the speech recognition apparatus according to the second embodiment will be described below with reference to FIGS. 3 and 4. As shown in FIG. 3, the speech recognition apparatus 2 of the present embodiment includes a corpus storage unit 10, a speech recognition unit 11, a 3 gram calculation unit 12, an importance calculation unit 24, and an unnecessary word deletion unit 15. Except that the tf-idf calculation unit 13 existing in the speech recognition device 1 of the first embodiment is omitted, and the importance calculation unit 14 of the first embodiment is replaced by the importance calculation unit 24 in the present embodiment. The same as in the first embodiment will not be described as appropriate.

In this embodiment, word2vec is used instead of NRD when calculating the importance of a word. word2vec is a method of converting each word of a large amount of text data into a U-dimensional (U is an integer of 2 or more) real value vector using DNN (Deep Neural Network). The word 2vec is disclosed, for example, in Reference Non-Patent Document 2.
(Reference Non-Patent Document 2: Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, Jeffrey Dean, “Distributed Representations of Words and Words and Their Compositionality”, [online], Oct 2013, [May 30, 2016 Search ], Internet <URL: https://papers.nips.cc/paper/5021-distributed-representations-of-words-and-phrases-and-their-compositionality.pdf>)

  The dimension U is specified manually. For example, U = 100 is used. Hereinafter, the real-valued vector of word2vec of word w will be written as x (w) (this vector is assumed to be a vertical vector). Using this,

の代わりに、

Instead of,

Use That is, the importance calculator 24 converts each word w of the text data into a U-dimensional real value vector x (w) using DNN, and based on the U-dimensional real value vector x (w), The importance of each word included in the first speech recognition result is calculated (S24).

  The configuration and operation of the speech recognition apparatus according to the third embodiment will be described below with reference to FIGS. 5 and 6. The speech recognition apparatus 3 of this embodiment is a modification of the speech recognition apparatus 2 of the second embodiment. As shown in FIG. 5, the speech recognition apparatus 3 of this embodiment includes a corpus storage unit 10, a speech recognition unit 11, a 3 gram calculation unit 12, an importance calculation unit 34, and an unnecessary word deletion unit 15. The second embodiment is the same as the second embodiment except that the importance calculation unit 24 of the second embodiment is replaced by the importance calculation unit 34 in the present embodiment, and thus the description will be omitted as appropriate.

When calculating the degree of importance of a word, the degree of importance calculation unit 34 of this embodiment calculates the degree of importance (naturalness in a sentence) on the basis of the degree of variation of a real-valued vector by word2vec. Now, each word w ₁ , w ₂ ,. . . , W _V , consider a real-valued vector x (w _i ).
Is the mean vector of this real-valued vector

を計算する。これを用いて、

Calculate Using this,

をその単語の重要度（文内の自然さ）を表す指標とし、

As a measure of the importance of the word (naturalness in the sentence),

の代わりに用いる。

In place of

  The configuration and operation of the speech recognition apparatus according to the fourth embodiment will be described below with reference to FIGS. 7 and 8. The speech recognition device 4 of this embodiment is a modification of the speech recognition device 2 of the second embodiment. As shown in FIG. 7, the speech recognition apparatus 4 of this embodiment includes a corpus storage unit 10, a speech recognition unit 11, a 3 gram calculation unit 12, an importance calculation unit 44, and an unnecessary word deletion unit 15. The second embodiment is the same as the second embodiment except that the importance calculation unit 24 of the second embodiment is replaced by the importance calculation unit 44 in the present embodiment, and thus the description will be omitted as appropriate.

  Similar to the third embodiment, the importance calculation unit 44 of the present embodiment calculates the importance (naturalness in a sentence) based on the degree of variation of the real value vector due to word2vec when calculating the importance of the word. calculate.

Now, each word w ₁ , w ₂ ,. . . , W _Y , consider real value vectors w _i . Assuming that this real-valued vector group follows one mixed normal distribution,

を用いて、対数ガウス確率

Log Gaussian probability using

を、その単語の重要度とする。これを、

And the importance of the word. this,

の代わりに用いる。

In place of

<Supplementary Note>
The apparatus according to the present invention is, for example, an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected as a single hardware entity, or a communication device (for example, communication cable) capable of communicating outside the hardware entity. Communication unit that can be connected, CPU (central processing unit, cache memory, registers, etc. may be provided), RAM or ROM that is memory, external storage device that is hard disk, input unit for these, output unit, communication unit , CPU, RAM, ROM, and a bus connected so as to enable exchange of data between external storage devices. If necessary, the hardware entity may be provided with a device (drive) capable of reading and writing a recording medium such as a CD-ROM. Examples of physical entities provided with such hardware resources include general purpose computers.

  The external storage device of the hardware entity stores a program necessary for realizing the above-mentioned function, data required for processing the program, and the like (not limited to the external storage device, for example, the program is read) It may be stored in the ROM which is a dedicated storage device). In addition, data and the like obtained by the processing of these programs are appropriately stored in a RAM, an external storage device, and the like.

  In the hardware entity, each program stored in the external storage device (or ROM etc.) and data necessary for processing of each program are read into the memory as necessary, and interpreted and processed appropriately by the CPU . As a result, the CPU realizes predetermined functions (each component requirement expressed as the above-mentioned,...

  The present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present invention. Further, the processing described in the above embodiment may be performed not only in chronological order according to the order of description but also may be performed in parallel or individually depending on the processing capability of the device that executes the processing or the necessity. .

  As described above, when the processing function in the hardware entity (the apparatus of the present invention) described in the above embodiment is implemented by a computer, the processing content of the function that the hardware entity should have is described by a program. Then, by executing this program on a computer, the processing function of the hardware entity is realized on the computer.

  The program describing the processing content can be recorded in a computer readable recording medium. As the computer readable recording medium, any medium such as a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, etc. may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only) Memory), CD-R (Recordable) / RW (Rewritable), etc. as magneto-optical recording medium, MO (Magneto-Optical disc) etc., as semiconductor memory EEP-ROM (Electronically Erasable and Programmable Only Read Memory) etc. Can be used.

  Further, this program is distributed, for example, by selling, transferring, lending, etc. a portable recording medium such as a DVD, a CD-ROM or the like in which the program is recorded. Furthermore, this program may be stored in a storage device of a server computer, and the program may be distributed by transferring the program from the server computer to another computer via a network.

  For example, a computer that executes such a program first temporarily stores a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. Then, at the time of execution of the process, the computer reads the program stored in its own recording medium and executes the process according to the read program. Further, as another execution form of this program, the computer may read the program directly from the portable recording medium and execute processing according to the program, and further, the program is transferred from the server computer to this computer Each time, processing according to the received program may be executed sequentially. In addition, a configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes processing functions only by executing instructions and acquiring results from the server computer without transferring the program to the computer It may be Note that the program in the present embodiment includes information provided for processing by a computer that conforms to the program (such as data that is not a direct command to the computer but has a property that defines the processing of the computer).

  Further, in this embodiment, the hardware entity is configured by executing a predetermined program on a computer, but at least a part of the processing content may be realized as hardware.

Claims (7)

Let N be an integer of 2 or more,
A voice recognition unit that outputs a first to N-th voice recognition result based on input voice data;
3gram calculation unit which calculates 3gram probability of text data prepared in advance;
The tf of each word included in the first to N speech recognition results and the idf of each word included in the first speech recognition result among idf prepared in advance based on the text data, A tf-idf calculator for calculating tf-idf of each word contained in the first-ranked speech recognition result;
An importance calculator configured to calculate an NRD of each word included in the first-ranked speech recognition result based on the tf-idf, and outputting the calculated value based on the NRD as the importance of each word;
The reliability of each word included in the first speech recognition result, the 3-gram probability of three consecutive words included in the first speech recognition result, and each word included in the first speech recognition result A speech recognition apparatus including an unnecessary word deletion unit which deletes an unnecessary word included in the first speech recognition result on the basis of the solution of the integer programming problem formulated using the importance and the importance. A speech recognition unit that outputs a first speech recognition result based on input speech data;
3gram calculation unit which calculates 3gram probability of text data prepared in advance;
Each word of the text data is converted into a multi-dimensional real-valued vector using DNN, and the importance of each word included in the first-ranked speech recognition result is calculated based on the multi-dimensional real-valued vector Importance calculation unit,
The reliability of each word included in the first speech recognition result, the 3-gram probability of three consecutive words included in the first speech recognition result, and each word included in the first speech recognition result A speech recognition apparatus including an unnecessary word deletion unit which deletes an unnecessary word included in the first speech recognition result on the basis of the solution of the integer programming problem formulated using the importance and the importance. The speech recognition apparatus according to claim 2, wherein
The importance calculation unit
A speech recognition apparatus which calculates the importance based on the degree of variation of the real value vector. A speech recognition method performed by a speech recognition device, comprising
Let N be an integer of 2 or more,
Outputting the first to N-th speech recognition results based on the input speech data;
Calculating a 3-gram probability of text data prepared in advance;
The tf of each word included in the first to N speech recognition results and the idf of each word included in the first speech recognition result among idf prepared in advance based on the text data, Calculating tf-idf of each word contained in the first-place speech recognition result;
Calculating an NRD of each word included in the first speech recognition result based on the tf-idf, and outputting a value based on the calculated NRD as an importance of each word;
The reliability of each word included in the first speech recognition result, the 3-gram probability of three consecutive words included in the first speech recognition result, and each word included in the first speech recognition result A speech recognition method including the step of deleting unnecessary words included in the first-ranked speech recognition result based on the solution of the integer programming problem formulated using the importance and the importance. A speech recognition method performed by a speech recognition device, comprising
Outputting a first speech recognition result based on the input speech data;
Calculating a 3-gram probability of text data prepared in advance;
Each word of the text data is converted into a multi-dimensional real-valued vector using DNN, and the importance of each word included in the first-ranked speech recognition result is calculated based on the multi-dimensional real-valued vector Step and
The reliability of each word included in the first speech recognition result, the 3-gram probability of three consecutive words included in the first speech recognition result, and each word included in the first speech recognition result A voice recognition method comprising the step of deleting unnecessary words included in the first place voice recognition result based on the solution of the integer programming problem formulated using the degree of importance. The speech recognition method according to claim 5, wherein
A speech recognition method for calculating the degree of importance based on the degree of variation of the real value vector.   A program that causes a computer to function as the speech recognition device according to any one of claims 1 to 3.

Images