JP2010060809A - Method of calculating identification score posterior probability classified for each number of errors, identification learning device and method with the number of errors weight using the method, voice recognition device using the device, program and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To calculate an identification score and a posterior probability for each number of errors. <P>SOLUTION: A method of calculating the identification score for each number of errors includes: a local score local error calculation process; identification score calculation process for each number of errors; and posterior probability calculation process for each number of errors. The identification score calculation process for each number of errors calculates a local score, the number of local errors, the identification score for each number of errors of a lattice by inputting the lattice, a forward accumulative score for each number of errors, and a backward accumulative score for each number of errors. The posterior probability calculation process for each number of errors calculates the identification score of the lattice for each number of errors, the forward accumulative score for each number of errors, the backward accumulative score for each number of errors, and the posterior probability that a correct symbol sequence includes each directed arc included in the lattice by inputting the local score, for each number of errors. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  This invention uses a feature amount information sequence of a signal expressing conceptual information on a time axis or a space axis such as a voice, a still image, and a moving image to recognize a pattern as a symbol sequence expressed by a predetermined discrete value. Error number-specific identification score / posterior probability calculation method used for model parameter learning, error number weighted identification learning device using the method, method thereof, speech recognition device using the device, program, and recording It relates to the medium.

  In order to reduce pattern recognition errors, it is effective to estimate the number of errors as a model parameter function (loss function) and to learn model parameters so that the loss function value becomes small. As a typical conventional technique for performing such learning, an MPE / MWE (Minimum Phone Error / Minimum Word Error) learning method is known (Non-Patent Document 1). A case where this learning method is applied to continuous speech recognition, which is a typical example of pattern recognition, will be described as an example.

  FIG. 7 shows a functional configuration example of the speech recognition apparatus 800. The speech recognition device 800 includes an acoustic model learning device 700 using the MPE / MWE learning method, a speech feature amount extraction unit 80, a word sequence search unit 81, and an acoustic model parameter recording unit 82. The acoustic model learning apparatus 700 includes a local score / local error calculation unit 70, an error number average value calculation unit 71, a forward / backward score calculation unit 72, a posterior probability calculation unit 73, a loss function value calculation unit 74, and a partial differential coefficient value calculation. Unit 75 and model parameter update unit 76.

The voice feature quantity extraction unit 80 calculates a voice feature quantity information series _Xr of the voice information string by using a voice information string formed of discrete values as an input. Audio feature information series _{X r} is extracted by, for example, Mel Frequency Cepstral Coefficients (MFCC) analysis. Word sequence search unit 81, in accordance with the audio feature information sequence X _r, outputs word lattice represented by searching the recorded feature amount information sequence in the acoustic model parameter recording section 82 by a plurality of symbol sequences To do. FIG. 8 shows an example of a word lattice corresponding to “Today's Hokkaido”. The word lattice is also called a word graph, and has a structure in which a plurality of nodes are connected by a directed arc q corresponding to a word or a partial word sequence.

The directed arc q also corresponds to a partial feature quantity sequence included in the speech feature quantity information series _Xr . The correspondence between each directed arc q included in the lattice and the word or partial word sequence, and the correspondence between each directed arc q included in the lattice and the partial feature amount sequence are obtained as a result of word string search. It is.

The local score / local error calculation unit 70 receives the correct answer sequence word string S _r , the speech feature quantity information sequence X _r , the model parameter λ _(t), and the word lattice, and inputs the local of the directed arc q. A local score p _q which is a score and a local error number e _q which is a local error number are calculated. The local score p _q can be calculated by equation (1). Since the local score p _q and the local error number e _q are calculated for each effective arc q included in the lattice, they form a set for the effective arc q. In FIG. 7, this is expressed as {p _q }, {e _q }. The same applies to other signals.

Here, P _ΛG (s _q ) is a local language score assigned to the directed arc q. p _ΛA (x _q | s _q ) is a local acoustic score assigned to the directed arc q. The product of the local language score and the local acoustic score is the local score. η and ψ are control coefficients, and the larger the η, the larger the contribution ratio of the local language score P _ΛG (s _q ) in the local score. ψ is a coefficient that suppresses (ψ → small) or emphasizes (ψ → large) variations for each partial word sequence s _q when calculating the local score p _q . Also, the local score local error calculating unit 70 counts the local number of errors e _q. Counting Examples of topical error count e _q shown in FIG. FIG. 9A shows an example of the number of errors in units of words. FIG. 9B shows an example of the number of errors per phoneme, and FIG. 9C shows an example of the number of errors per time frame. All are examples of the number of errors per unit when the correct sentence is "Today's weather in Hokkaido is sunny". For example, the number of errors in word units is 2, the number of errors in phoneme units is 5, and the number of errors is 35 in frame units. The number of local errors e _q is counted by comparing the correct word sequence S _r , the speech feature amount information sequence X _r, and the word lattice Q. Here, Q is a set including the connection relation of the directed arc q.

  The error number average value calculation unit 71 calculates an average value c (q) of error numbers shown in Expression (2) for all word sequences passing through the directed arc q.

ここで、Ｓは単語系列の全体を意味し、Ｅは単語系列全体の誤り数を意味する。

Here, S means the whole word series, and E means the number of errors in the whole word series.

The average value c (q) of the number of errors can be efficiently calculated by a combination of a forward / backward algorithm and an average error number propagation algorithm. That is, as shown in Expression (3), the local error number e _q can be obtained by adding the forward average error number α ′ _q and the backward average error number β ′ _q .

A conceptual diagram of how to obtain the forward average error number α ′ _q is shown in FIG. Before obtaining the forward average error number α ′ _q , the forward cumulative probability α _q of the directed arc q is obtained by calculation of Equation (4). The forward cumulative probability α _q of the directed arc _q is the forward cumulative of each of the preceding directed arcs q ⁻ (i) (i = 1,..., N _q ) whose end is connected to the start end of the directed arc q. This is the accumulation of values obtained by multiplying the probability α _{q− (i)} by the respective local probabilities p _{q (i)} .

Backward cumulative probability beta _q to be described later with forward cumulative probability alpha _q of directed arcs q is the forward-backward score calculation 72 is calculated as an input local score p _q. The forward cumulative probability α _q and the backward cumulative probability β _q described later are output to the error number average value calculating unit 71 and the posterior probability calculating unit 73.

The error number average value calculation unit 71 receives the local score p _q , the local error number e _q, and the word lattice q as input, and calculates the number of errors for all word sequences passing through the directed arc q by the calculation of Equation (3). An average value c (q) is calculated. In calculating Equation (3), the required forward average error number α ′ _q is calculated by Equation (5).

ｅ_{ｑ−（ｉ）}は先行有向弧ｑ⁻（ｉ）の局所誤り数である。

e _{q- (i)} prior directed arcs ^q - is the local number of errors (i).

Similarly, the backward cumulative probability β _q and the backward average error number β ′ _q can be calculated by the equations (6) and (7).

FIG. 11 shows a conceptual diagram of how to obtain the average value c (q) of the number of errors. The partial word sequence s _q assigned to the directed arc q is, for example, “Tokaido”, and the average value c (q) of the number of errors in the directed arc is given by Equation (3).

Equations (4) and (5) are applied recursively in order from the directional arc starting from the start node of the word lattice, and further recursively from the directional arc starting from the end node of the word lattice. If Expressions (6) and (7) are applied, the average value c (q) of the number of errors can be obtained for all directed arcs q by calculation of Expression (3). The directed arc q is ((q, {q ⁻ (i)}, {q ⁺ (i)}) ∈Q). Hereinafter, qεQ is simply written.

The posterior probability calculation unit 73 receives the forward cumulative probability α _q and the backward cumulative probability β _q as inputs, and calculates the posterior probability P _Λ (q | X _r ) that the recognition candidate word sequence passes through the directed arc q using the formula (8 ) To calculate.

The loss function value calculation unit 74 receives the average value c (q) of the number of errors and the posterior probability P _Λ (q | X _r ) as inputs, calculates the equation (9), and performs a loss function F _MPE ( _Xr ) is obtained.

Recognition errors are reduced by minimizing the loss function F _MPE (X _r ) with respect to the model parameter λ using an optimization method. As a specific optimization method, a probabilistic descent (PD) method, a Quickprop method, an Rprop method, an extended Baum-Welch (EBW) method, or the like can be used. In any optimization method, it is necessary to calculate the partial differential coefficient value ∂F _MPE (X _r ) / ∂λ of F _MPE (X _r ) for the model parameter λ (∈Λ). When the partial differential coefficient value ∂F _MPE (X _r ) / ∂λ is set as logp _Λ (X _r , q) _≈log (P _ΛG (q) ^ηψ p _ΛA (X _r , | q) ^ψ ), It can be disassembled as shown in (10).

Among the formula (10), the identification score logp representing the likelihood that X _r is characteristic amount information sequence of speech uttered with the intention of word string including partial word sequence corresponding to the directed arcs q The partial derivative value for _Λ (X _r , q) is important.

The partial differential coefficient value calculation unit 75 receives the average value c (q) of the number of errors, the posterior probability P _Λ (q | X _r ), and the loss function F _MPE (X _r ) as an input in the directed arc q. Is obtained by the calculation of equation (11).

That is, in the MPE / MWE learning method applied to the word lattice, the difference between the average value c (q) of the number of errors of the directed arc q and the average value F _MPE (X _r ) of the number of recognition errors, that is, the recognition error. Learning is performed on the basis of the average value F _MPE (X _r ) of the numbers. If the average value c (q) of the number of errors is smaller than the function F _MPE (X _r ), the learning is advanced so as to increase the identification score logp _Λ (X _r , q). Conversely, when the directed arc q has an average value c (q) of the number of errors larger than F _MPE (X _r ), the learning is advanced so as to lower the identification score logp _Λ (X _r , q). By repeating this, the loss function F _MPE (X _r ) can be minimized. The model parameter updating unit 76 updates the model parameter λ _(t) recorded in the acoustic model parameter recording unit 82 to the model parameter λ _{(t + 1)} so that the loss function becomes low.
D. Povey and P. Woodland, “Minimum Phone Error and I−smoothing for improved discriminative training,” in Proc.ICASSP02, pp.105-108, 2002.

  In the conventional MPE / MWE learning method, the difference between the average value of the number of recognition errors in the target symbol sequence set and the average value of the number of recognition errors as the entire lattice, and the posterior probability of the target word sequence set The model parameters are learned based on the size of. Therefore, the following problems arise.

  If the amount of learning data is not enough, or if the learning data and the feature quantity information sequence to be recognized have different statistical properties, the number of errors in the feature quantity information sequence to be recognized is sufficient. The learning effect to reduce cannot be obtained. In particular, if the distribution of the number of symbol sequences for each number of errors differs greatly between the result of recognizing a feature quantity information sequence to be recognized by the initial model and the result of recognizing learning data, the learning effect cannot be obtained.

  Further, depending on the recognition result output characteristics of the search unit, the lattice output from the search unit may include a large number of symbol sequences that differ only in the time of symbol boundaries even with the same symbol arrangement. In such a case, the bias of the distribution of the number of symbol sequences for each number of errors becomes large, and the learning effect is limited.

  The present invention has been made in view of such problems, and has a small adverse effect due to a bias in learning data and a bias in the appearance tendency of symbol sequences, and is classified by the number of errors used for learning parameters that can provide higher recognition accuracy. It is an object of the present invention to provide an identification score / posterior probability calculation method, an error number weighted identification learning device using the method, a speech recognition device using the device, a program, and a recording medium.

  The number-of-errors identification score / posterior probability calculation method of the present invention includes a local score / local error calculation process, an error number-specific identification score calculation process, and an error number-specific posterior probability calculation process. In the local score / local error calculation process, a feature quantity information series, a correct symbol series corresponding to the feature quantity information series, a lattice representing the feature quantity information series as a plurality of recognition symbol sequences, and a model parameter are input. As described above, the local score for each directed arc included in the lattice and the number of local errors included in the directed arc are calculated. The error score identification score calculation process is performed by using the local score, the local error count, and the lattice as input, the discrimination score by error count of the lattice, a forward cumulative score by error count, and a backward cumulative score by error count. And calculate. The posterior probability calculation process according to the number of errors includes the identification symbol for each number of errors in the lattice, the forward cumulative score by the number of errors, the backward cumulative score by the number of errors, and the local score, and the correct symbol series. Calculates the a posteriori probability of including each directed arc included in the lattice according to the number of errors.

  In addition, the error number weighted identification learning method of the present invention uses the error number-specific identification score / posterior probability calculation method of the present invention. The speech recognition apparatus of the present invention uses the error number weighted identification learning method of the present invention.

  The method of calculating the identification score / posterior probability according to the number of errors of the present invention calculates the identification score by the number of errors of the lattice, the forward cumulative score by the number of errors, and the backward cumulative score by the number of errors in the process of calculating the identification score by the number of errors. Then, the posterior probability that the correct symbol sequence includes each directed arc included in the lattice is calculated for each error number. In other words, since the identification score and the posterior probability are calculated for each error number, the identification score for each error number and the posterior can be used in a model parameter learning device that is not easily affected by the bias of the learning data and the bias of the appearance tendency of the recognition symbol series. It can be a probability calculation method.

  Further, the error number weighted identification learning method using the error number-specific identification score / posterior probability calculation method of the present invention can be a learning method for improving recognition performance. Furthermore, the speech recognition apparatus using the error number weighted discrimination learning method can improve the recognition rate.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows an example of the functional configuration of an error number weighted identification learning apparatus 100 according to the present invention. The operation flow is shown in FIG. The error number weighted identification learning apparatus 100 includes a local score / local error calculation unit 12, an error number-specific identification score calculation unit 13, an error number-specific posterior probability calculation unit 14, a loss function value calculation unit 15, and a partial differentiation. A coefficient value calculation unit 16 and a model parameter update unit 17 are provided. The symbol sequence search unit 10 and the model parameter recording unit 11 provided outside the error number weighted identification learning device 100 include a word sequence search unit 81 provided outside the acoustic model learning device 700 described in the related art. This is basically the same as the acoustic model parameter recording unit 82. The local score / local error calculation unit 12 and the model parameter update unit 17 in the error number weighted identification learning device 100 are the same as the local score / local error calculation unit 70 and the model parameter update in the acoustic model learning device 700 of the prior art. This is basically the same as the section 76. That is, for these portions, there is only a difference whether or not the model parameter is for speech recognition.

  In the following description, only different parts will be described. The operations up to step S10 and step S12 in FIG. 2 are the same as those of the conventional acoustic model learning apparatus 700. The error number weighted identification learning device 100 and the acoustic model learning device 700 are realized by a predetermined program being read into a computer including, for example, a ROM, a RAM, and a CPU, and the CPU executing the program. Is done.

The number-of-errors identification score calculation unit 13 includes a local score p _q calculated by the local score / local error calculation unit 12, a local error number e _q, and a lattice that represents a feature amount information sequence by a plurality of recognition symbol sequences, As an input, an identification score G _j (X _r ) (hereinafter abbreviated as an identification score G _j (X _r )) for each number of lattice errors shown in Expression (12) is obtained (step S13).

The identification score G _j (X _r ) can be efficiently calculated by a forward / backward algorithm for each number of errors. Now, local score p _q local scores in directed arcs _q, and local error count e _q local errors Number of directed arcs q. Assuming that the partial feature quantity information sequence assigned to the directed arc q as a result of continuous pattern recognition is x _q , and the partial word sequence is s _q , the local score p _q of the directed arc q is the same as the formula ( It can be calculated in 1). The sum of forward probabilities α _q with the local score of the directed arc q and the forward cumulative score α _{q- (i)} of all the preceding directed arcs q ⁻ (i) (i = 1,..., N _q− ). _{, J} is calculated by equation (13).

That is, in all combinations of the directed arc q (i) and the cumulative error number k before the preceding directional arc, the cumulative error number k before the preceding directional arc and the local error of the preceding directional arc q ⁻ (i). For each j that is the sum of the number e _q− (i), the sum of forward probabilities is obtained.

If the formula (13) is applied recursively using the preceding directed arc without omission in order from the directed arc starting from the starting node of the word lattice, the number of forward accumulated errors for all the directed arcs q∈Q Another forward probability α _{q, j} (hereinafter referred to as a forward cumulative score by number of errors) can be obtained. As for the headed directional arc, the virtual nodes of p _{q− (1)} = 1, α _{q− (1), j} = 1 (for all j) and e _q− = 0 with the start node as the end Consider the leading directed arc q ⁻ (1) = q ^start .

Similarly, if Equation (14) is applied recursively using the subsequent directed arc without omission in order from the directed arc that ends at the end of the word lattice, backward accumulation is performed for all directed arcs q∈Q. A total sum β _{q, j} of backward probabilities by number of errors (hereinafter referred to as backward cumulative score by number of errors) can be obtained.

If the forward cumulative score α _{q, j} by number of errors or the backward cumulative score β _{q, j} by number of errors is calculated for all directional arcs q∈Q, the identification score G _j (X _r ) can be _expressed by the equation (15). It is obtained by.

That is, the identification score G _j (X _r ) is the forward cumulative score α ^final _j = G _j (X _r ) by the number of errors in the terminal directed arc or the backward cumulative score β ^start _j by the number of errors in the ^start directed arc. = G _j (X _r ).

The posterior probability calculation unit 14 according to the number of errors includes the identification score G _j (X _r ), the forward cumulative score α _{q, j} by the number of errors, the backward cumulative score β _{q, j} by the number of errors _, and the locality of the directed arc q. Using the score p _q as an input, the posterior probability γ _{q, j} that the recognition symbol sequence having the error number j passes through the directed arc q is calculated by the equation (16) (step S14).

FIG. 3 shows a conceptual diagram of the operations of the equations (13) to (16). The forward / backward cumulative score integration ζ _{q, j} (hereinafter abbreviated as cumulative score integration ζ _{q, j} ) of the directed arc q with the local error e _q = 2 can be calculated by the equation (17). That is, the numerator of formula (16).

The cumulative score integration ζ _{q, j} is a cumulative score α _{q, k} before the preceding directed arc for each cumulative error number, a cumulative score β _{q, u} until the subsequent directed arc, and a directed arc for which the posterior probability is to be obtained. It is the cumulative product of _q and the local score p _q . A value obtained by dividing the cumulative score integration ζ _{q, j} by the identification score G _j (X _r ) is the posterior probability γ _{q, j} that the recognition symbol sequence having the error number j passes through the directed arc q. α _{q, k} , β _{q, j} , and γ _{q, j} are the respective values for the directed arc q and the number of errors j.

The loss function value calculation unit 15 receives the identification score G _j (X _r ) for each number of errors and calculates, for example, a loss function value shown in Expression (18) (step S15).

Here, φ is a control coefficient for suppressing (φ → small) or emphasizing (φ → large) the variation of the identification score G _j (X _r ) for each error number for each error number j.

  The partial differential coefficient value calculation unit 16 calculates a partial differential coefficient value (formula (19)) that minimizes the loss function of the formula (18) for the model parameter (step S16).

The value of the first term on the right side of Equation (19) is important, and the partial differential coefficient value calculation unit 16 uses the first term on the right side as the identification score G _j (X _r ) for each error number and the loss function value. F _MGE1 (X _r ) and the posterior probability γ _{q, j} are used as inputs to calculate according to equation (20).

By applying the optimization method using Equation (18) and Equation (19), the loss function value can be minimized. The total loss Γ _MGE1 (Z) (formula 21) and the partial differential coefficient value (formula (22)) of the feature amount information series are used to determine the convergence of the optimization. Z represents various partial feature amount information series X _r1 ,
.., X _rm is an entire feature amount information series (Zε {X _{r (m)} | m = 1,..., M}). _{Xr (m)}
Is correct in the expression.

The model parameter updating unit 17 receives the partial differential coefficient value of the equation (18), the loss function value F _MGE1 (X _r ), and the model parameter λ _(t) recorded in the model parameter recording unit 11 as an optimum. The parameter is updated to the model parameter λ _{(t + 1)} that minimizes the loss function using the optimization method (step S17).

  As described above, since the model parameters are updated using the identification score for each error number, it is possible to realize a model parameter learning apparatus that is hardly affected by the bias of the learning data and the bias of the appearance tendency of the recognition symbol series. The reason why it is difficult to be influenced by the bias of the learning data and the bias of the appearance tendency of the recognition symbol series is shown in FIGS. 4 and 5 show the relationship between the number of errors per word symbol sequence, the posterior probability, and the partial differential coefficient value. The horizontal axis represents the number of errors per word symbol sequence, the vertical axes in FIGS. 4 (a), 4 (c), 5 (a), and 5 (c) represent posterior probabilities, and FIGS. 4 (b) and 4 (c). d), the vertical axis of FIGS. 5B and 5D is the partial differential coefficient value.

  All the parameters ψ for adjusting the variation for each partial word sequence are fixed to 0.1. Parameters φ for adjusting the variation for each error number j are 1.0 in FIGS. 4A, 4B, 5A, and 5B, and FIGS. 4C, 4D, and 5C. ), (D) are set to φ = 0.25. When φ = 1 is set here, the posterior probability and the partial differential coefficient value are average values of the word symbol series. This is clear from equation (18). That is, FIGS. 4A and 4B and FIGS. 5A and 5B show the characteristics of the prior art. Therefore, the characteristics shown in FIGS. 4C and 4D and FIGS. 5C and 5D are the posterior probabilities and partial differential coefficient values obtained in the first embodiment.

  Comparing FIG. 4 (a) with FIG. 4 (c), FIG. 4 (b) and FIG. 4 (d), the characteristics of the prior art do not change until the number of errors per word symbol sequence exceeds 30. On the other hand, there is a change in the characteristics of Example 1. The comparison between FIG. 5 (a) and FIG. 5 (c) and FIG. 5 (b) and FIG. 5 (d) is the same. The characteristics shown in FIGS. 5C and 5D are obtained by using the loss function and the partial differential coefficient values shown in the equations (22) and (23), and the parameters ν = 0.25 and ε = 10. It is. This change means that it contributes to learning.

  Expressions other than those described above can be applied to the expressions for obtaining the loss function value and the partial differential coefficient value. For example, as shown in Equation (23) and Equation (24), a threshold value ε may be provided for the number of errors, and the loss function value may be weighted using the threshold value ε as a boundary.

  Here, ν is a parameter for controlling variation with the threshold ε as a boundary. Moreover, you may use a loss function value and a partial differential coefficient value as shown to Formula (25) and Formula (26).

Here, σ is an attenuation coefficient that exponentially attenuates the identification score as the number of errors increases. [Application example]
An example in which the error number weighted identification learning apparatus 100 of the present invention is applied to speech recognition will be described. FIG. 6 shows a functional configuration example in which a speech recognition apparatus 600 is configured using the error number weighted identification learning apparatus 100 of the present invention. The speech recognition device 600 is obtained by replacing the speech recognition learning device 700 of the speech recognition device 800 described in the prior art with the error number weighted identification learning device 100 of the present invention.

  Since the speech recognition apparatus 600 according to the present invention learns the acoustic model according to the number of errors, the accuracy of the recognition score can be improved. Therefore, it is possible to realize a voice recognition device with few erroneous recognitions.

〔Experimental result〕
An experiment was conducted for the purpose of confirming the effect of the error number weighted discriminative learning method of the present invention. According to the learning method of the present invention, the loss function value is obtained by the equation (22), the partial differential coefficient value is obtained by the equation (23), and the conditions of ψ = 0.04, φ = 0.25, ν = 0.65, ε = 20. So, I learned about 230 hours of speech in a Japanese conference. After that, the word error rate as a result of speech recognition of the evaluation speech having a length of about 130 minutes different from the learning data by the speech recognition apparatus 700 of the present invention was 18.8%. The word error rate as a result of speech recognition of the evaluation speech by the speech recognition apparatus 800 of the prior art was 19.3%. The word error rate as a result of speech recognition using the initial model parameters was 21.6%. Therefore, the relative error reduction rate with an initial error rate of 100 is 13.0% for the method of the present invention and 10.6% for the conventional method, and the recognition performance of the learning method of the present invention is superior. It was confirmed that

  The number-of-errors identification score / a posteriori probability calculation method based on the technical idea of the present invention, and the number-of-errors weighted identification learning apparatus and method using the method are not limited to the above-described embodiment. Modifications can be made as appropriate without departing from the spirit of the invention. The processes described in the above-described apparatus and method are not only executed in time series according to the order described, but may be executed in parallel or individually as required by the processing capability of the apparatus that executes the process. .

  Further, when the processing means in the above apparatus is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape, or the like, and 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. can be used as magneto-optical recording media, MO (Magneto Optical disc) can be used, and flash memory can be used as semiconductor memory.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

The figure which shows the function structural example of the identification device 100 with an error number weighting of this invention. The figure which shows the operation | movement flow of the error number weighted identification device. The figure which shows notionally the operation | movement which calculates | requires the forward accumulation score classified by error number-posterior probability (gamma) _{q, j} . It is a figure which shows the reason which is hard to be influenced by the bias of learning data and the appearance tendency of a recognition symbol series, (a) is a figure which shows the posterior probability according to the number of errors of a prior art, (b) is a partial differential of a prior art The figure which shows a coefficient value, (c) is a figure which shows the posterior probability according to error number of Example 1, (d) is the figure which shows the partial differential coefficient value according to error number of Example 1. It is a figure which shows the reason which is hard to be influenced by the bias of learning data and the appearance tendency of a recognition symbol series, (a) is a figure which shows the posterior probability according to the number of errors of a prior art, (b) is a partial differential of a prior art The figure which shows a coefficient value, (c) is a figure which shows the posterior probability according to error number of Example 1, (d) is the figure which shows the partial differential coefficient value according to error number of Example 1. The figure which shows the function structure of the speech recognition apparatus 600 of this invention. The figure which shows the function structural example of the conventional acoustic model learning apparatus. The figure which shows an example of the word lattice corresponding to "Today's Hokkaido." Is a diagram showing a counting Examples of local error count e _q, (a) is a diagram showing an example of a number of errors word unit, (b) is a diagram showing the number of errors phoneme, (c), the time frame It is a figure which shows the example of the number of errors. The figure which shows the conceptual diagram of how to obtain _| require forward average error number (alpha) ' _q . The figure which shows the conceptual diagram of how to obtain | require the average value c (q) of the number of errors.

Claims (10)

The local score / local error calculation unit inputs a feature quantity information series, a correct symbol series corresponding to the feature quantity information series, a lattice representing the feature quantity information series as a plurality of recognition symbol series, and model parameters. A local score for each directed arc included in the lattice and a local score / local error calculation process for calculating the number of local errors included in the directed arc;
The number-of-errors identification score calculation unit receives the local score, the number of local errors, and the lattice as input, and the identification score for each number of errors of the lattice, the forward cumulative score by the number of errors, and the backward cumulative score by the number of errors. And an identification score calculation process according to the number of errors for calculating
The posterior probability calculation unit by number of errors receives the identification score by number of errors of the lattice, the forward cumulative score by number of errors, the backward cumulative score by number of errors, and the local score, and the correct symbol series A posterior probability calculation process according to the number of errors for calculating the posterior probability of each including the directed arc included in the lattice, according to the number of errors,
A method of calculating the identification score and posterior probability according to the number of errors including. In the identification score / posterior probability calculation method according to the number of errors described in claim 1,
The identification score calculation process according to the number of errors is as follows:
Forward finding the forward cumulative score by the number of errors as the sum of forward probabilities accumulated for each sum of the local score of the preceding directed arc of the directed arc and the number of errors of the forward cumulative score by the number of errors up to the preceding directed arc Total probability calculation step;
Backward probabilities for obtaining the backward cumulative score by the number of errors as the sum of backward probabilities accumulated for each sum of the local score of the subsequent directed arc of the directed arc and the number of errors of the backward cumulative score by the number of errors up to the subsequent effective arc A sum calculation step;
A posteriori probability calculation step for obtaining an identification score for each number of errors of the lattice as a posterior probability for each sum of the number of errors of the local score of the directed arc, the forward probability sum, and the backward probability sum;
An error-specific identification score / posterior probability calculation method characterized by including: The identification score / posterior probability calculation method according to the number of errors according to claim 1 or 2,
A pattern recognition process in which a pattern recognition unit searches and outputs a lattice representing a plurality of symbol sequences from the model parameter recording unit by inputting a feature amount information sequence;
A loss function value calculation process in which the loss function value calculation unit calculates a loss function value using the discrimination score for each number of errors of the lattice as an input,
The partial differential coefficient value calculation unit calculates a partial differential coefficient value in a directed arc with the identification score for each number of errors of the lattice, the loss function value, and the posterior probability for each error number as inputs. Numerical calculation process,
A model parameter update unit updates the model parameter with the partial differential coefficient value, the loss function, and the model parameter as inputs, and
An error-weighted discriminative learning method including: In the error-weighted discriminative learning method according to claim 3,
The loss function value calculation process divides a value obtained by raising the identification score for each number of errors of the lattice by a control coefficient, and dividing by an accumulated value of values obtained by raising the identification scores of all directed arcs by the power of the control coefficient. An error number weighted discriminative learning method, characterized in that a value obtained by multiplying the value by the error number is accumulated in the error number to obtain the loss function. A model parameter recording unit for recording model parameters;
A pattern recognizing unit that searches the model parameter recording unit for a lattice that represents a plurality of symbol sequences with a feature amount information sequence as an input;
The feature quantity information series, a correct symbol series corresponding to the feature quantity information series, a lattice that represents the feature quantity information series as a plurality of recognition symbol sequences, and model parameters as inputs, are included in the lattice. A local score for the directed arc of, and a local score / local error calculation unit for calculating the number of local errors included in the directed arc;
Discriminating by error number, using the local score, the local error number, and the lattice as inputs, and calculating an identification score by error number, a forward cumulative score by error number, and a backward cumulative score by error number A score calculator,
Each of the correct symbol sequences included in the lattice by inputting the identification score for each number of errors of the lattice, the forward cumulative score by the number of errors, the backward cumulative score by the number of errors, and the local score. An a posteriori probability calculation unit for each error number that calculates the a posteriori probability of including an arc for each error number;
A loss function value calculation unit for calculating a loss function value by using the discrimination score according to the number of errors of the lattice as an input;
A partial differential coefficient value calculation unit that calculates a partial differential coefficient value in a directed arc by using the discrimination score for each error number of the lattice, the loss function value, and the posterior probability for each error number as input,
A model parameter updating unit that updates the model parameter with the partial differential coefficient value, the loss function, and the model parameter as inputs;
An error-weighted identification learning device comprising: In the error number weighted identification learning device according to claim 5,
The loss function value calculation unit divides a value obtained by raising the identification score for each number of errors of the lattice by a control coefficient, and a value obtained by dividing a cumulative value of the identification score for each number of errors of the lattice by a power of the control coefficient, Further, an error number weighted identification learning apparatus characterized in that a value obtained by multiplying the number of errors is accumulated by the number of errors to form the loss function. An error number weighted identification learning device according to claim 5 or 6,
A voice feature quantity extraction unit that calculates a voice feature quantity information sequence of the voice information string using the voice information string as an input;
An acoustic model recording unit for recording an acoustic model learned by the error number weighted identification learning device;
A word string search unit for searching a feature amount information sequence recorded in the acoustic model parameter recording unit and outputting a word lattice represented by a plurality of symbol sequences in accordance with the speech feature amount information sequence;
A speech recognition apparatus comprising:   An apparatus program for causing a computer to function as the error number weighted identification learning apparatus according to claim 5.   An apparatus program for causing a computer to function as the voice recognition apparatus according to claim 7.   A computer-readable recording medium on which the apparatus program according to claim 8 is recorded.

Images