JP4528076B2 - Speech recognition apparatus and speech recognition program - Google Patents

Description

  The present invention relates to a speech recognition apparatus and a speech recognition program, and more particularly to a speech recognition apparatus and a speech recognition program for improving speech recognition accuracy.

  Conventionally, as a method for improving speech recognition accuracy in speech recognition technology, for example, a method of improving speech recognition rate by using one feature amount as a measure of word reliability, such as a posteriori probability of a word hypothesis. Has been proposed (see, for example, Non-Patent Document 1).

In addition, as feature quantities other than posterior probabilities, there are acoustic stability, word hypothesis density, and the like (see, for example, Non-Patent Documents 2 and 3).
F. Wessel et al. , “Confidence measurement for large vocabulary continuous speech recognition,” IEEE Trans. Speech and Audio Processing, Vol. 9, PP. 288-298, Marth 2001. T. T. Zeppenfeld, M.M. Finke, and K.K. Ries, “Recognition of conventional telephone use the the Janus speech engine,” IEEE int. Conf. Acoustics, Speech and Signal Pressing, PP. 1815-1818, 1997. T. T. Kemp and T.W. Shaaf, “Estimating Confidence Using Word lattices,” Eurospech, Rhodes, Greece, PP. 827-830, 1997.

  By the way, as shown in the above-mentioned non-patent document, the conventional method obtains the reliability of the word using only the posterior probability of the word hypothesis or other single feature amount, and improves the speech recognition rate. .

  However, the reliability obtained from a single feature amount is low in accuracy, and the improvement in recognition accuracy is small. Therefore, in order to improve the accuracy of correct / incorrect classification of whether the word hypothesis is correct or incorrect, it is preferable to integrate a plurality of measures. That is, it is possible to expect a reduction in word error rate in speech recognition by increasing classification accuracy.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to provide a speech recognition device and a speech recognition program for accurately determining whether a word output is correct and improving speech recognition accuracy. .

  In order to solve the above problems, the present invention employs means for solving the problems having the following characteristics.

The invention described in claim 1 is a speech recognition apparatus for recognizing speech, speech recognition means for generating a word network from input speech, a word network obtained by the speech recognition means, an acoustic model, and a language Based on a model, feature quantity calculation means for calculating a plurality of preset feature quantities for a plurality of word hypotheses included in the word network, and an acoustic score, a language score, and a word posterior obtained by the feature quantity calculation means probability, acoustic stability, word hypothesis density, number of active HMM, a phoneme number average frame of the back-off case, and a plurality of feature amounts set in advance from among the feature amount by the correctness label history of the word hypotheses with a respective confidence measure x each consisting of time series of the plurality of feature quantities and the feature quantity, it indicates the correctness of the word hypotheses By changing either the value of the correctness for each of the word hypothesis, consisting of the following formula (1) reliability model P _ME | setting the correctness labels y for maximizing the value of _(y x) If each reliability measure x is larger than a threshold value set in advance for each reliability measure x and the correct / incorrect label y is correct, “1” is set. Otherwise, “1” is set. 0 "consisting of the binary function that feature function f (x, y), the feature function f (x, y) log confidence score and a predetermined weight λ for the logarithm of the probability of correctness of the word hypotheses And the calculated log confidence score and the logarithmic acoustic score and logarithmic language score obtained by the speech recognition means are added for each sentence hypothesis included in the word network to obtain the highest score. Sentence hypothesis, from highest to lowest score And a score calculating means for outputting because the set number sentence hypothesis, or the correct word sequence of preset score or more sentence hypotheses as the speech recognition result, the log confidence score, the feature function f (x , Y) and the weight λ, the probability of being a correct answer given the word hypothesis obtained from the reliability model P _ME (y | x) consisting of the following equation (1): It is calculated by taking the logarithm of .

  According to the first aspect of the invention, by using the word network, it is possible to accurately determine whether the word output is correct. Thereby, the accuracy of the speech recognition result can be improved.

The invention described in claim 2 is characterized in that the score calculation means obtains the plurality of feature quantities and a time series of the feature quantities from a word hypothesis included in the word network.

  According to the second aspect of the present invention, the accuracy of the speech recognition result can be improved by calculating the reliability score based on a plurality of feature quantities and the time series of the feature quantities from the word hypothesis.

The invention described in claim 3, wherein the score calculating means, based on the feature amount of the threshold value set in advance, said word classified into correct or incorrect answer each word hypothesis, the results were classified as the errata label It is characterized by giving to a hypothesis.

  According to invention of Claim 3, a score calculation can be efficiently performed by providing a correct / incorrect label.

According to a fourth aspect of the present invention, the score calculation means includes: a binary classifier expressed by the feature function f (x, y) that is classified into a correct answer or an incorrect answer according to a preset feature amount threshold value. A plurality of binary classifiers associated with a time series of feature quantities, the log confidence score obtained from the combined binary classifier , the logarithmic acoustic score and the logarithmic language score Are added for each sentence hypothesis included in the word network, and the correct word string is output.

According to the fourth aspect of the present invention, it is possible to determine the correctness of the word output with high accuracy in correspondence with the time series, and to improve the accuracy of the speech recognition result.

According to a fifth aspect of the present invention, there is provided a speech recognition program for causing a computer to execute speech recognition processing for recognizing speech, wherein the computer recognizes speech recognition means for generating a word network from input speech, and the speech recognition Feature amount calculating means for calculating a plurality of preset feature amounts for a plurality of word hypotheses included in the word network based on a word network obtained by the means, an acoustic model, and a language model; and the feature acoustic score obtained by the amount computing means, the language score, the word posterior probabilities, acoustic stability, word hypothesis density, number of active HMM, a phoneme number average frames, due to the back-off case, and correctness label history of the word hypotheses Using the plurality of feature amounts set in advance from among the feature amounts, the plurality of feature amounts and the feature amounts are used. Each confidence measure x each consisting of time series of amounts, shows the correctness of said word hypotheses, by changing to one of the values of the correctness for each of the word hypothesis consists of the following equation (1) reliability model P _ME | values of _(y x) sets the correctness labels y for maximizing, greater than a predetermined threshold the respective confidence measure x is relative to the respective confidence measure x, In addition, when the correct / incorrect label y is correct, it is set to “1”, otherwise it is set to “0”, and a feature function f (x, y) composed of a binary function and the feature function f (x, calculates log confidence score is the logarithm of the probability of correctness of the word hypotheses from the predetermined weight λ for y), and the calculated the log confidence scores, logarithmic acoustic score and obtained by the speech recognition means Log language score and included in the word network Add as each sentence hypothesis, output the sentence hypothesis that has the highest score, the sentence hypothesis of a preset number from the highest score, or the correct word string of sentence hypotheses greater than the preset score as the speech recognition result The logarithmic reliability score is the reliability model P _ME (y | x) comprising the following function (1) including the feature function f (x, y) and the weight λ. It is calculated by taking the logarithm of the probability with respect to the probability of the correct answer when the word hypothesis obtained from the above is given .

According to the fifth aspect of the present invention, it is possible to determine the correctness of word output with high accuracy. Thereby, the accuracy of the speech recognition result can be improved. In addition, voice recognition can be realized at a low cost without requiring a special device configuration. Moreover, voice recognition can be easily realized by installing the program.

  According to the present invention, speech recognition accuracy can be improved.

<Outline of the present invention>
The present invention generates a word network from input speech such as generated speech, obtains various feature amounts of each word hypothesis on the word network, and corrects / incorrects speech recognition based on the feature amount and its time series. The correct answer is expressed by binary classification, the score (reliability score) is calculated from a reliability model that integrates the binary classification result (evaluation of correct answer rate), and further from the language score obtained from the language model and the acoustic model A correct word string is output as a speech recognition result together with the obtained acoustic score.

  Note that the binary classification results can be integrated using, for example, a maximum entropy (ME) model that has a feature that a model in which different knowledge sources are integrated can be easily created and is often used in classification problems.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments in which a speech recognition apparatus and a speech recognition program according to the present invention are suitably implemented will be described in detail with reference to the drawings.

<Functional configuration>
FIG. 1 is a diagram illustrating an example of a functional configuration of a speech recognition apparatus according to the present invention. The speech recognition apparatus 10 in FIG. 1 includes a speech recognition unit 11, a feature amount calculation unit 12, a score calculation unit 13, an acoustic model 14, a language model 15, and a reliability model 16. .

  The voice recognition means 11 inputs the spoken voice or the like as the input voice, and stores the acoustic model 14 in which information indicating the likelihood obtained from the waveform of the voice and the pronunciation of the word is stored, and the ease of connection between words. A word network is generated with reference to the language model 15 in which the information shown is stored. Further, the voice recognition unit 11 outputs the generated word network to the feature amount calculation unit 12.

  The feature quantity calculation means 12 calculates a plurality of preset feature quantities for each word hypothesis of the word network with reference to the acoustic model 14 and the language model 15. Also, the feature quantity calculation means 12 calculates the feature quantity from the inputted word network itself. The feature quantity calculation unit 12 outputs the calculated feature quantity to the score calculation unit 13.

  The score calculation means 13 generates a correct word string based on the reliability model 16 and outputs it as a speech recognition result. Next, the specific content of each component described above will be described.

<Voice recognition means 11>
The speech recognition means 11 inputs the spoken speech or the like as input speech, and is connected by word hypotheses obtained from the acoustic score of the acoustic model 14 and the language score of the language model 15 with the end time of the word as a vertex. Generate a word network with parts as edges. The voice recognition unit 11 outputs the generated word network to the feature amount calculation unit 12.

<Feature amount calculation means 12>
For example, (a) an acoustic score, (b) a language score, (c) a word posterior probability, and (d) an acoustic stability (acoustic) for the word hypothesis on each side of the input word network. (e) word hypothesis density, (f) number of active HMMs (Hidden Markov Models), (g) average number of phoneme frames, (h) backoff (Back-off) case, ( i) A feature amount based on a history of correct / wrong labels of the word hypothesis is obtained.

  Here, the above-mentioned (a) acoustic score can be obtained by referring to the acoustic model 14, and (b) the language score can be obtained by referring to the language model 15. Further, (c) the word posterior probability indicates how much each side (score) of the word network is used, and can be obtained by the method described in Non-Patent Document 1, for example.

  Moreover, (d) acoustic stability shows the frequency of the route used in a word network, for example, can be calculated | required by the method shown in the nonpatent literature 2. FIG. Moreover, (e) word hypothesis density shows the frequency of the word used at a certain time on a word network, for example, can be calculated | required by the method of a nonpatent literature 3.

  In addition, (f) the number of active HMMs indicates the number of HMMs that are simultaneously searched when searching for words in speech recognition. For example, the larger the number of HMMs in a certain word search, It can be determined that the word is unlikely to be correct.

  Also, the back-off case is, for example, when the ease of connection between a certain word hypothesis is not required, and the acoustic score and language score when the word network is generated are written in descending order of score. The feature amount is set using a predetermined number of sentence hypotheses.

  The feature amount calculation unit 12 outputs the word network to which the respective feature amounts obtained as described above are added to the score calculation unit 13.

<Score calculation means 13>
The score calculation means 13 generates an optimal correct word string for the word network to which the feature amount is assigned, and outputs the generated correct word string as a speech recognition result. Here, a specific configuration example of the score calculation unit 13 will be described with reference to the drawings.

  FIG. 2 is a diagram showing a configuration example of the score calculation means in the present invention. The score calculation unit 13 illustrated in FIG. 2 includes a sentence hypothesis generation unit 21, a reliability calculation unit 22, and a rescoring unit 23.

  The sentence hypothesis generation means 21 receives the feature-added word network obtained from the feature quantity calculation means 12 and generates N sentence hypotheses (n-best sentence hypotheses) set in advance by performing rescoring or the like. To do. Moreover, the sentence hypothesis generation means 21 can obtain a feature amount for the generated N sentence hypotheses by, for example, a back-off case. In this case, in the back-off case, a feature value is set to “1” when n-gram exists in the language model 15 and “0” otherwise.

  In addition, the sentence hypothesis generation unit 21 outputs the obtained feature amount to the reliability calculation unit 22. The reliability calculation means 22 calculates the reliability based on the input feature quantity and the reliability model 16.

  Here, the reliability model 16 can be obtained by, for example, the maximum entropy method. (For the maximum entropy method, see, for example, “A. Berger, SD Pietra, and V.D. 1996. ”, etc.) By the maximum entropy method, the reliability model is given by the following equation (1).

ここで、ｘは仮説の事後確率等の信頼度尺度（特徴量及びその時系列）を示し、ｙ∈｛−１，１｝は単語仮説の正解又は不正解の正誤ラベルを示している。また、ｆ_ｉ（ｘ，ｙ）は、観測値ペア（ｘ，ｙ）に関して特徴の条件で“０”又は“１”を返す素性関数と呼ばれる二値関数を示し、λ_ｉは素性関数に対する重みを示している。

Here, x represents a reliability measure (feature and its time series) such as a posteriori probability of a hypothesis, and y∈ {−1, 1} represents a correct answer or incorrect label of a word hypothesis. Further, f _i (x, y) represents a binary function called a feature function that returns “0” or “1” under the feature condition for the observed value pair (x, y), and λ _i is a weight for the feature function. Is shown.

P _ME (y | x) is a reliability model and indicates a probability that the focused word hypothesis is correct (or incorrect). Note that λ _i described above can be obtained by a GIS (Generalized Interactive Scaling) algorithm or the like.

  The reliability calculation means 22 uses a feature quantity obtained from the feature-added word network and functions as a binary classifier that determines the correct / incorrect answer of the word hypothesis according to the feature value threshold (configuration). Have The binary classifier is expressed by the feature function of the above-described equation (1).

  Here, since the feature function is a binary function, it can be regarded as a simple “binary classifier”. What is important in expressing reliability (features) using such a function is “how to express reliability with feature functions” and “how the reliability changes over time (time series)” "How do you represent it?" Therefore, in the present invention, in order to describe the reliability using a feature function, a threshold is set for a certain reliability and a feature function that is activated before and after the threshold is defined.

In other words, it has a plurality of binary classifiers that classify word hypotheses into correct or incorrect answers based on feature amount threshold values, and these multiple binary classifiers are combined as a binary classifier connected with respect to the time series of feature amounts, A reliability score and a correct / incorrect label for the word hypothesis are obtained by a reliability model integrating all binary classifiers, and a correct word string is output. Accordingly, it is possible to determine whether the word output is correct or not with high accuracy in correspondence with the time series, and to improve the accuracy of the speech recognition result.

Here, the _{c t} and reliability, as correctness (correct or incorrect) label word hypothesis to predict _{y t,} shown against a threshold _{c thresh1} for reliability, for example, the feature functions of formula (1) below It is defined as equation (2).

また、式（２）に示したｆ_ｉは二値関数であり、特徴量を表現するには不十分である。そこで信頼度の詳細な表現を行うため、同じ信頼度尺度に対して複数の閾値ｃ_{ｔｈｒｅｓｈ２}，ｃ_{ｔｈｒｅｓｈ３}，・・・を定め、それぞれの閾値に対して素性関数を定義する。例えば、閾値ｃ_{ｔｈｒｅｓｈ２}及びｃ_{ｔｈｒｅｓｈ３}を用いた場合には、以下に示す式（３）、式（４）のようになる。

Further, f _i shown in equation (2) is a binary function, is insufficient to represent the feature quantity. Therefore, in order to express the reliability in detail, a plurality of threshold _values c _thresh2 , c _thresh3 ,... _Are defined for the same reliability measure, and a feature function is defined for each threshold value. For example, when threshold _values c _thresh2 and c _thresh3 are used, the following equations (3) and (4) are obtained.

また、信頼度の時間的変化は、単語仮説に対して得られた信頼度の系列を素性関数に取り込むことで実現する。特徴量の時間的変化を表現するためには、例えば信頼度ｃ_ｔ−１，ｃ_ｔに対して、以下に示す式（５）により素性を決定する。

Further, the temporal change in the reliability is realized by incorporating the reliability series obtained for the word hypothesis into the feature function. In order to express the temporal change of the feature amount, for example, the feature is determined by the following equation (5) with respect to the reliability c _t−1 and c _t .

なお、上述した素性関数の定義では、信頼度に対する閾値を決定することが重要である。そこで、信頼度に対する閾値は、次の手順で決定する。まず、任意の二値分類器を用いて閾値を１つ決定する。次に、閾値の上下に一定の間隔で、新たな閾値を設定し、素性を定義して最大エントロピーモデルを学習する。閾値の設定は、モデルによる分類誤り率が下がり始めてから上がらなくなるまで繰り返し行う。

In the above-described feature function definition, it is important to determine a threshold for reliability. Therefore, the threshold value for the reliability is determined by the following procedure. First, one threshold is determined using an arbitrary binary classifier. Next, a new threshold value is set at regular intervals above and below the threshold value, the feature is defined, and the maximum entropy model is learned. The threshold setting is repeated until the classification error rate by the model starts to decrease and does not increase.

  Next, the feature is re-determined by changing the window width (the number of series of reliability scales). The above processing is performed for all reliability measures, and integration by the maximum entropy method is performed again.

  Thus, a reliability model can be obtained by defining a threshold and a feature and integrating them by a maximum entropy method.

  When the reliability model is obtained, the reliability calculation means 22 obtains the correct correct / incorrect correct / incorrect correct label sequence according to the following equation (6).

ここで、上述の式（６）のａｒｇｍａｘ操作は、各単語仮説に対する最適な正解・不正解の正誤ラベルの系列ｙ^＊を、ｖｉｔｅｒｂｉアルゴリズム（ｖｉｔｅｒｂｉアルゴリズムについては、例えば「Ｇ．Ｄ．Ｆｏｒｎｅｙ，“Ｔｈｅ Ｖｉｔｅｒｂｉ Ａｌｇｏｒｉｔｈｍ，” Ｐｒｏｃ．ＩＥＥＥ，Ｖｏｌ．６１，ｐｐ．２６８−２７８，１９７３」等に示されている。）により求めることを示している。

Here, the argmax operation of the above equation (6) is performed by calculating the optimum correct / incorrect correct / incorrect correct label sequence y ^* for each word hypothesis, and the viterbi algorithm (for the Viterbi algorithm, for example, “GD Forney,” The Viterbi Algorithm, “Proc. IEEE, Vol. 61, pp. 268-278, 1973”, etc.).

At this time, a reliability score P _ME (y _t | x _t ) is given to each word hypothesis w _t ⁽ⁿ⁾ in the N sentence hypotheses. The reliability calculation unit 22 outputs the reliability score of the above-described reliability model to the rescoring unit 23.

Rescoring means 23, for example with respect to the n-th sentence hypothesis w of N sentence hypotheses generated by sentence hypothesis generation means 21 ^(n), by adding to weight the output scores of the binary classifier The rescoring score S (w ⁽ⁿ⁾ ) is corrected by the calculation of the following equation (7).

ただし、ａｃ（ｗ_ｔ ^（ｎ））は文仮説ｗ_ｔ ^（ｎ）の音響モデルの対数スコアを示し、ｌｍ（ｗ_ｔ ^（ｎ））はｗ_ｔ ^（ｎ）の言語モデルの対数スコアを示し、ｃｆ（ｗ_ｔ ^（ｎ））は信頼度モデルの対数スコアを示している。また、ｇｗは言語スコアに対する重みを示し、ｃｗは二値分類器スコア（信頼度スコア）に対する重みを示している。

Where ac (w _t ⁽ⁿ⁾ ) represents the logarithmic score of the acoustic model of the sentence hypothesis w _t ⁽ⁿ⁾ , lm (w _t ⁽ⁿ⁾ ) represents the logarithmic score of the language model of w _t ⁽ⁿ⁾ , cf (w _t ⁽ⁿ⁾ ) indicates a logarithmic score of the reliability model. Gw represents a weight for the language score, and cw represents a weight for the binary classifier score (reliability score).

Here, as a method of correcting the rescoring score S (w ⁽ⁿ⁾ ), there are a method of simply correcting the score by adjusting the weight cw and a method of correcting the classifier score according to the recognition rate of the sentence hypothesis. is there.

The method will be described later, and label the correctness label by the maximum entropy model specifically for each n-best sentence hypotheses, determined the correctness label _{y t,} the score ^{_{cf (w t (n))}} , the above-mentioned Rescore the n-best sentence hypothesis using equation (7).

  At this time, the classifier score cw is corrected according to the word error rate of the sentence hypothesis. The correction formula is given by the following formula (8).

ここで、Ｃａ，Ｃｂはそれぞれ定数を示している。また、ｅ（ｗ_ｔ ^（ｎ））（０≦ｅ（ｗ_ｔ ^（ｎ））≦１）は単語誤り率を示しているが、これは事前には得ることができない。そこで、上述した式（６）を用いて信頼度計算手段２２で得られた正誤ラベル系列を用いて、以下に示す式（９）を計算し、単語誤り率を代替する。

Here, Ca and Cb are constants, respectively. Also, e (w _t ⁽ⁿ⁾ ) (0 ≦ e (w _t ⁽ⁿ⁾ ) ≦ 1) indicates the word error rate, but this cannot be obtained in advance. Therefore, using the correct / wrong label sequence obtained by the reliability calculation means 22 using the above-described equation (6), the following equation (9) is calculated to replace the word error rate.

このように、単語認識率に応じてペナルティを設けてリスコアリングスコアに与えることにより、単語誤り率の高い文仮説をｎ−ｂｅｓｔの下位に押し下げることができる。

Thus, by providing a penalty according to the word recognition rate and giving it to the rescoring score, it is possible to push down a sentence hypothesis having a high word error rate to the lower part of n-best.

  The re-scoring means 23 corrects sentence hypotheses that have the highest reliability score obtained by rescoring, a predetermined number of sentence hypotheses in descending order of score, or a sentence hypothesis greater than or equal to a preset score. A word string is output as a speech recognition result.

  As described above, the speech recognition apparatus according to the present invention can determine whether a word output is correct or incorrect with high accuracy. Thereby, the accuracy of voice recognition can be improved.

  Here, the voice recognition device described above can perform voice recognition in the present invention using the dedicated device configuration described above, but generates an execution program that can cause a computer to execute the processing in each configuration. For example, voice recognition in the present invention can be realized by installing a program in a general-purpose personal computer, workstation, or the like.

<Hardware configuration>
Here, a hardware configuration example of a computer capable of performing speech recognition according to the present invention will be described with reference to the drawings. FIG. 3 is a diagram illustrating an example of a hardware configuration capable of realizing speech recognition according to the present invention.

  3 includes an input device 31, an output device 32, a drive device 33, an auxiliary storage device 34, a memory device 35, a CPU (Central Processing Unit) 36 that performs various controls, and a network connection device. 37, and these are connected to each other by a system bus B.

  The input device 31 has a pointing device such as a keyboard and a mouse operated by a user, and inputs various operation signals such as execution of a program from the user. The output device 32 has a display for displaying various windows and data necessary for operating the computer main body for performing the processing in the present invention, and for recognizing the sound in the present invention by the control program of the CPU 36. Program execution progress and results can be displayed.

  Here, in the present invention, the execution program installed in the computer main body is provided by, for example, the recording medium 38 such as a CD-ROM. The recording medium 38 on which the program is recorded can be set in the drive device 33, and the execution program included in the recording medium 38 is installed in the auxiliary storage device 34 from the recording medium 38 via the drive device 33.

  The auxiliary storage device 34 is a storage means such as a hard disk, and can store an execution program in the present invention, a control program provided in a computer, and the like, and can perform input / output as necessary.

  Based on a control program such as an OS (Operating System) and an execution program read and stored by the memory device 35, the CPU 36 performs various operations and inputs / outputs data to / from each hardware component. Each process in the speech recognition can be realized by controlling the process. Various information necessary during the execution of the program can be acquired from the auxiliary storage device 34 and can also be stored.

  The network connection device 37 acquires an execution program from another terminal connected to the communication network by connecting to a communication network or the like, or an execution result obtained by executing the program or an execution in the present invention The program itself can be provided to other terminals.

  With the hardware configuration described above, voice recognition can be realized at low cost without requiring a special device configuration. Moreover, voice recognition can be easily realized by installing the program.

  Next, a processing procedure in the execution program will be described using a flowchart.

<Voice recognition processing>
FIG. 4 is a flowchart showing an example of a voice recognition processing procedure using the voice recognition program according to the present invention. The spoken voice or the like is input as input voice (S01), the input voice data is recognized, and a word network is generated (S02).

  Next, based on the word network obtained in S02, a plurality of feature amounts set in advance are calculated with reference to the acoustic model and the language model (S03). Further, a reliability score is calculated based on the word network to which the feature amount obtained in S03 and the reliability model are given (S04), and the sentence hypothesis and score having the highest reliability are obtained as a result of the calculation. A correct number of sentence hypotheses from a descending order or a correct word string of sentence hypotheses equal to or higher than a preset score is output as a speech recognition result (S05).

  FIG. 5 is a flowchart illustrating an example of the reliability score calculation processing procedure in S04. In the reliability score calculation process in S04, first, a word network with a feature amount to which a plurality of feature amounts are added is input (S11), and N sentence hypotheses (n-best) preset by re-scoring are input. Sentence hypothesis) is generated (S12). In addition, the reliability (reliability score) is calculated based on the feature amount of the generated sentence hypothesis (S13). In addition, the re-scoring of the reliability is performed by adjusting the weighting of the reliability in the above-described formula (7) for calculating the reliability score (S14).

  With these processes, it is possible to determine the correctness of word output with high accuracy. In addition, the accuracy of voice recognition can be improved. Furthermore, voice recognition can be easily realized by installing an execution program for performing the above-described processing in a general-purpose computer or the like.

<Comparison between the conventional method and the present invention>
Next, a comparison result with the conventional method using the above-described speech recognition method will be described with reference to the drawings. FIG. 6 is a diagram illustrating an example of comparison data and comparison results. Here, FIG. 6 (a) shows an example of comparison data, FIG. 6 (b) shows a classification error rate when a correct / incorrect label is given, and FIG. 6 (c) shows a comparison result of rescoring. Yes.

  FIG. 6A shows the number of sentences in the maximum entropy learning data and the evaluation data of the method in the present invention, the number of words, the perplexity that is a parameter indicating the difficulty of the sentence, the unknown word rate, Indicates the word error rate.

  Here, the learning data is used for learning the maximum entropy model of the reliability model after outputting the recognition result by the speech recognition device and obtaining each feature amount. Of the learning data, 18,106 words are used to determine the feature function threshold of the maximum entropy model, the classifier threshold based on each feature, and parameters (language weight, insertion penalty) of the speech recognition system. It was. For the reliability model, a feature function of 8,210 words was selected from the learning data.

  FIG. 6B shows the classification error rate for each feature when the reliability calculation unit 22 assigns correct / incorrect correct / incorrect labels. In FIG. 6B, the word posterior probability, the acoustic stability, and the word hypothesis density indicate the classification error rate (conventional method) when a binary classifier is configured using each feature amount, and is a reliability model. Indicates the classification error rate according to the invention. According to FIG. 6B, it can be seen that the reliability model has a lower error rate and higher performance than the classification based on the conventional feature amount.

  FIG. 6C shows the word error rate in the speech recognition result. In addition, the conventional method in FIG.6 (c) performs speech recognition by the method which does not add the reliability score cf of a reliability model by (7) Formula shown above. The word posterior probability is obtained by performing speech recognition using the logarithm of the word posterior probability as the reliability score cf. As can be seen from FIG. 6C, the reliability model obtained by the method of the present invention improves the word error rate, and the present invention is effective.

  As described above, according to the present invention, it is possible to determine the correctness of word output with high accuracy. Thereby, the accuracy of the speech recognition result can be improved. Specifically, a plurality of feature quantities are integrated to obtain reliability regarding correct and incorrect answers of the speech recognition word hypothesis. Further, by integrating the reliability measure with the maximum entropy model and labeling the correctness / incorrectness of the word, the correctness / incorrectness determination of the word output can be performed with high accuracy, and the speech recognition accuracy can be improved.

  The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to such specific embodiments, and various modifications, within the scope of the gist of the present invention described in the claims, It can be changed.

It is a figure which shows an example of a function structure of the speech recognition apparatus in this invention. It is a figure which shows the example of 1 structure of the score calculation means in this invention. It is a figure which shows an example of the hardware constitutions which can implement | achieve the speech recognition in this invention. It is a flowchart of an example which shows the speech recognition process sequence using the speech recognition program in this invention. It is a flowchart of an example which shows the reliability score calculation process procedure. It is a figure which shows an example of comparison data and a comparison result.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speech recognition apparatus 11 Voice recognition means 12 Feature-value calculation means 13 Score calculation means 14 Acoustic model 15 Language model 16 Reliability model 21 Sentence hypothesis generation means 22 Reliability calculation means 23 Rescoring means 31 Input device 32 Output device 33 Drive Device 34 Auxiliary storage device 35 Memory device 36 CPU
37 Network connection device 38 Recording medium

Claims (5)

In a speech recognition device that recognizes speech,
Speech recognition means for generating a word network from input speech;
Feature quantity calculation means for calculating a plurality of preset feature quantities for a plurality of word hypotheses included in the word network based on the word network obtained by the speech recognition means, an acoustic model, and a language model;
Acoustic score obtained by the feature amount calculating means, the language score, the word posterior probabilities, acoustic stability, word hypothesis density, number of active HMM, a phoneme number average frame of the back-off case, and the correctness label of said word hypotheses history by using a plurality of feature amounts set in advance from among the feature amount of a respective confidence measure x each consisting of time series of the plurality of feature quantities and the feature amount, the correctness of the word hypothesis shows , by changing to one of the values of the correctness for each of the word hypothesis, reliability model P _ME consisting of the following formulas (1) | and correctness labels y for maximizing the value of _(y x) And set to “1” when each reliability measure x is larger than a threshold value set in advance for each reliability measure x and the correct / incorrect label y is correct, and otherwise. Two with “0” Feature function f (x, y) composed of functions and to calculate the feature function f (x, y) log confidence score is logarithmic and a predetermined weight λ probability of correctness of the word hypothesis for, it is calculated The log reliability score and the logarithmic acoustic score and logarithmic language score obtained by the speech recognition means are added for each sentence hypothesis included in the word network, and the sentence hypothesis having the highest score, A score calculation means for outputting as a speech recognition result a sentence hypothesis of a predetermined number of sentence hypotheses from a descending order, or a correct word string of sentence hypotheses equal to or higher than a preset score ;
The logarithmic reliability score is given by the word hypothesis obtained from the reliability model P _ME (y | x) including the following function (1) including the feature function f (x, y) and the weight λ. A speech recognition apparatus , wherein the probability is calculated by taking the logarithm of the probability with respect to the probability of being a correct answer .

The score calculation means includes
The speech recognition apparatus according to claim 1, wherein the plurality of feature quantities and a time series of the feature quantities are obtained from a word hypothesis included in the word network. The score calculation means includes
3. The word hypothesis is classified into a correct answer or an incorrect answer based on a preset feature amount threshold, and the classified result is assigned to the word hypothesis as the correct / incorrect label. Voice recognition device. The score calculation means includes
A plurality of binary classifiers expressed by the feature function f (x, y) for classifying the correct answer or the incorrect answer according to a preset feature amount threshold;
The plurality of binary classifiers are combined in correspondence with a time series of feature quantities, and the log reliability score obtained from the combined binary classifier, the logarithmic acoustic score, and the logarithmic language score are combined into the word network. 4. The speech recognition apparatus according to claim 1, wherein the correct word string is output by adding each sentence hypothesis included in the sentence hypothesis. 5. In a speech recognition program for causing a computer to execute speech recognition processing for recognizing speech,
The computer,
Speech recognition means for generating a word network from input speech;
Feature quantity calculation means for calculating a plurality of preset feature quantities for a plurality of word hypotheses included in the word network based on a word network obtained by the speech recognition means, an acoustic model, and a language model; and ,
Acoustic score obtained by the feature amount calculating means, the language score, the word posterior probabilities, acoustic stability, word hypothesis density, number of active HMM, a phoneme number average frame of the back-off case, and the correctness label of said word hypotheses history by using a plurality of feature amounts set in advance from among the feature amount of a respective confidence measure x each consisting of time series of the plurality of feature quantities and the feature amount, the correctness of the word hypothesis shows By changing the value of each of the word hypotheses to one of the correct and incorrect values, a correct / incorrect label y for maximizing the value of the reliability model P _ME (y | x) consisting of the following equation (1) is obtained: And set to “1” when each reliability measure x is larger than a threshold value set in advance for each reliability measure x and the correct / incorrect label y is correct, and otherwise. Two with “0” Feature function f (x, y) composed of functions and to calculate the feature function f (x, y) log confidence score is logarithmic and a predetermined weight λ probability of correctness of the word hypothesis for, it is calculated The log reliability score and the logarithmic acoustic score and logarithmic language score obtained by the speech recognition means are added for each sentence hypothesis included in the word network, and the sentence hypothesis having the highest score, Function as score calculation means for outputting as a voice recognition result a sentence hypothesis of a sentence hypothesis of a preset number of sentences from a large order or a sentence hypothesis greater than a preset score ,
The logarithmic reliability score is given by the word hypothesis obtained from the reliability model P _ME (y | x) including the following function (1) including the feature function f (x, y) and the weight λ. A speech recognition program which is calculated by taking the logarithm of the probability with respect to the probability of being a correct answer when the answer is made .

Images