JP2014081516A - Symbol string conversion method, voice recognition method using the same, device of symbol string conversion and voice recognition and program, and recording medium for the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a symbol string conversion method that makes a hypothesis generation rule incapable of a cascade connection dependable.SOLUTION: The symbol string conversion method is configured to include: an auxiliary hypothesis generation step in which an input signal selection unit selects any one of an input symbol string to be input to a hypothesis generation unit of a previous stage with reference to an input/output symbol string arrangement storage unit and an output signal string to be output by each hypothesis generation unit of the previous stage and a stage before the previous stage as an input symbol string of a hypothesis generation unit of an m-th stage (m is an integer of 1 to M), and the hypothesis generation unit of the m-th stage repeats an auxiliary hypothesis generation step where the hypothesis generation unit of the m-th stage generates an auxiliary hypothesis fto an M-th stage while incrementing m; a hypothesis correction step in which a hypothesis correction unit corrects a weight of an auxiliary hypothesis fof the previous stage by an accumulation weight of the auxiliary hypothesis f, and repeats the correction until an accumulation weight of a hypothesis fis corrected by an accumulation weight of the auxiliary hypothesis while decrementing m; and a hypothesis narrowing-down step in which a hypothesis narrowing-down unit extracts a hypothesis which a corrected accumulation weight points to a specific value or an auxiliary hypothesis thereof, and narrows down the extracted hypothesis or auxiliary hypothesis as an input symbol string of the hypothesis narrowing-down unit of a next stage.

Description

  The present invention is applied from many output symbol strings that can be generated for an input symbol string by a symbol string conversion rule expressed by a weighted finite state transformer (WFST). A symbol string conversion method capable of efficiently finding an output symbol string having a singular value having a minimum or maximum cumulative weight of conversion rules, a speech recognition method using the same, and a device and a program thereof The present invention relates to a recording medium.

  In two different weighted finite state converters, if the output symbol string of one hypothesis generation rule can be the input symbol string of the other hypothesis generation rule, these two hypothesis generation rules are combined to create one hypothesis generation rule. Can be created. For example, speech recognition makes use of these features, hypothesis generation rules based on acoustic models, hypothesis generation rules that convert triphones (phoneme triplets) to phonemes, hypothesis generation rules based on pronunciation dictionaries, hypothesis generation based on language models By synthesizing the rules, a mechanism for converting to a word string that is a speech recognition result is realized.

  On the other hand, when the hypothesis generation rules are synthesized, there is a problem that the memory at the time of execution increases because the number of states and the number of state transitions of the hypothesis generation rules become very large. In order to solve this problem, for example, a symbol string conversion device 900 disclosed in Patent Document 1 is considered. FIG. 14 shows a functional configuration of the symbol string conversion apparatus 900 and its operation will be described.

  The symbol string converter 900 includes hypothesis generators 201-1 to 201-M cascaded in M stages, and each of the M hypothesis generators 201-2 to the M-th hypothesis generator 201-M. M hypothesis generation rule storage units 101-1 to 101-M, hypothesis correction units 301-1 to 301-M, and hypothesis correction units 301-1 to 301-M, respectively. It is constituted by M hypothesis narrowing-down units 401-1 to 401 -M provided therewith.

  Each hypothesis correction unit 301-1 to 301-M includes weight correction means 301A and repetition control means 301B. The weight correction unit 301A corrects the cumulative weight of the hypothesis generated by the preceding hypothesis generation unit with the value of the singular value with which the cumulative weight in the hypothesis generated by the subsequent hypothesis generation unit is maximum or minimum. The iterative control unit 301B repeatedly executes the operation of correcting the accumulated weight from the (M−1) th stage to the first stage hypothesis correcting unit.

  The hypothesis narrowing-down units 401-1 to 401-M include a hypothesis extraction unit 401A and an input processing unit 401B. The hypothesis extraction unit 401A extracts a hypothesis having the smallest cumulative weight, for example, from the corrected hypotheses generated by the respective hypothesis generation units 201-1 to 201-M when all the symbol strings are read. Then, the process of inputting the hypothesis to the hypothesis narrowing down section in the next stage is executed. The input processing unit 401B uses the hypothesis extracted by the hypothesis extraction unit 401A as the next input.

  The first-stage hypothesis generation unit 201-1 generates a hypothesis based on the hypothesis generation rule stored in the first-stage hypothesis generation rule storage unit 101-1 for one symbol input of the input symbol string. Generate. A hypothesis is a possibility that a certain symbol string is input in order and the state transition is repeated from the initial state according to the input symbol string in the hypothesis generation rule for the input symbol string read up to now. Represents one state transition process.

  Hypotheses (candidate symbol strings) generated by the preceding hypothesis generation unit are input to the second and subsequent hypothesis generation units 201-2 to 201-M. Thereafter, for the purpose of distinguishing the hypothesis output from the first-stage hypothesis generation unit 201-1 from the hypotheses output from the second-stage and subsequent hypothesis generation units 201-2 to 201-M, the second-stage and subsequent hypotheses. The hypothesis generated by the generation units 201-2 to 201-M is referred to as an m-th stage auxiliary hypothesis or simply an auxiliary hypothesis.

  The hypotheses generated by the respective hypothesis generation units 201-1 to 201-M are generated through weighted finite state transition processes stored in the respective hypothesis generation rule storage units 101-1 to 101-M, and Multiple hypotheses are generated based on the possibilities. Each hypothesis is weighted according to the conditions attached to the weighted finite state transition process. The weight is accumulated according to the state transition from the first-stage hypothesis generation unit 201-1 to the final-stage hypothesis generation unit 201-M.

  When a hypothesis is generated in the final hypothesis generation unit 201-M, the hypothesis correction units 301-1 to 301-M start operation. The hypothesis correction unit 301-M extracts, for example, a hypothesis having a minimum cumulative weight from the hypotheses generated by the hypothesis generation unit 201-M, and uses the cumulative weight of the hypothesis to obtain the previous M-1 stage. The hypothesis generation unit 201- (M-1) corrects the accumulated weight of the hypothesis. This correction is performed by adding the minimum cumulative weight extracted in the final stage to all the cumulative weights of the hypotheses generated by the previous hypothesis generation unit 201- (M-1). This correction is repeated until the hypothesis output by the first-stage hypothesis generation unit 201-1 is corrected.

  When all the symbol strings are read and the process of correcting the hypotheses is completed, the hypothesis narrowing-down units 401-1 to 401-M accumulate among all the hypotheses generated by the respective hypothesis generation units 201-1 to 201-M. Select the (probable) hypothesis that shows the singular value of the weight, and narrow down the hypothesis. The narrowing-down operation can be performed as follows.

  The first-stage hypothesis narrowing-down unit 401-1 extracts the hypothesis having the smallest cumulative weight, for example, from the hypotheses generated by the first-stage hypothesis generation unit 201-1, and sends the hypothesis to the second-stage hypothesis generation unit 201-2. Given as an input symbol string, a second hypothesis generation unit 201-2 generates an auxiliary hypothesis according to a finite state transition process possible with this input symbol string.

  The second-stage hypothesis narrowing-down section 301-2 extracts a hypothesis having the smallest cumulative weight, for example, from the auxiliary hypotheses generated by the second-stage hypothesis generation section 201-2, and generates the auxiliary hypothesis as the third-stage hypothesis. The procedure of setting the input symbol string of the unit 201-3 is repeated until the auxiliary hypothesis output by the final-stage hypothesis generation unit 201-M. The auxiliary hypotheses extracted from the auxiliary hypotheses generated by the final-stage hypothesis narrowing-down section 401-M from the auxiliary hypothesis generation section 201-M are output to the symbol string output section 106, and the conversion of the symbol strings is completed.

  In this way, the hypothesis generator is cascaded in multiple stages, and the hypothesis whose cumulative weight shows a singular value from the hypotheses generated by the final hypothesis generator is output as a conversion result (voice recognition result), thereby increasing memory. Is avoiding.

Japanese Patent No. 4478088

  The symbol string that can be converted by the conventional symbol string conversion apparatus 900 is limited to the case where the output symbol string of the hypothesis generation rule positioned immediately before can be the input symbol string of the hypothesis generation rule positioned immediately after. For example, in speech recognition using a class language model, a hypothesis generation rule that converts a word string into a class string in addition to a hypothesis generation rule based on an acoustic model, a hypothesis generation rule that converts triphones into phonemes, and a hypothesis generation rule based on a word pronunciation dictionary Speech recognition is performed by synthesizing hypothesis generation rules based on a class language model or by cascading the hypothesis generation units. For example, when adding a paraphrase process after the speech recognition process, since the output of the hypothesis generation rule by the above class language model is a class string, the hypothesis generation rule for paraphrasing using the word string as an input is cascaded. Can not do it.

  As described above, the conventional symbol string conversion apparatus 900 has a problem that only the input symbol string of the hypothesis generation rule can be connected, and there is a problem that lacks versatility.

  The present invention has been made in view of this problem, and by making it possible to change the cascade connection relationship of the hypothesis generation unit to an arbitrary combination, it is possible to subordinate hypothesis generation rules that cannot be cascaded in the past. It is an object of the present invention to provide a symbol string conversion method, a speech recognition method using the same, an apparatus and a program thereof, and a recording medium thereof.

The symbol string conversion method of the present invention is associated with each of M (M ≧ 2) cascaded hypothesis generation units, M hypothesis correction units associated with each of the hypothesis generation units, and hypothesis correction units. M hypothesis narrowing-down sections, input symbol strings of the first stage hypothesis generation section provided between cascaded hypothesis generation sections, and output symbol strings output from the preceding and previous hypothesis generation sections Whether the input symbol string selection unit that selects the input symbol string of the hypothesis generation unit of the next stage as input, the input symbol string selection information selected by the input symbol string selection unit, and the hypotheses extracted by the hypothesis narrowing unit are output A symbol string conversion device including an input / output symbol string arrangement storage unit storing the output selection information. The hypothesis generation unit of the first stage, the hypothesis generation process of generating in accordance with the hypothesis generation rule hypothesis f ₁ is a state transition process capable state transition for the input symbol string is input, the input signal selecting unit Is one of the input symbol string input to the first hypothesis generation unit with reference to the input / output symbol string arrangement storage unit and the output symbol string output from the preceding and previous hypothesis generation units. The m-th hypothesis generation unit (m is an integer from 1 to M) is used as an input symbol string, and the m-th hypothesis generation unit generates an auxiliary hypothesis f _m. The auxiliary hypothesis generation process that repeats up to the M-th stage while incrementing, and the hypothesis correction unit corrects the weight of the auxiliary hypothesis f _{m−1 in} the previous stage with the cumulative weight of the auxiliary hypothesis f _m , and uses the cumulative weight of the hypothesis f ₁ as the auxiliary hypothesis Repeatedly decrementing m until corrected with cumulative weight A hypothesis correction process, the hypothesis narrowing portion down is accumulated weights are corrected extracts the hypothesis or auxiliary hypothesis exhibits a singular value, the extracted hypotheses or auxiliary hypothesis as input symbol string of the next hypothesis refining unit.

  The speech recognition method of the present invention includes a speech feature symbol string extraction process in addition to the above-described symbol string conversion method. In the speech feature symbol string extraction process, an acoustic feature symbol string is extracted from a time series of short-time acoustic patterns of speech by using a speech signal as an input. The first-stage hypothesis generation unit receives the symbol string representing the acoustic feature extracted by the speech feature symbol string extraction unit, and inputs the hypothesis generation rule for the acoustic model having the standard feature of the fixed speech unit and various words. With reference to the word pronunciation dictionary hypothesis generation rule that associates pronunciation with the above-mentioned sequence of fixed speech units, a hypothesis is generated with a word-based symbol string as an output symbol string.

  According to the symbol string conversion method of the present invention, it is possible to arbitrarily combine hypothesis generation units that match the input symbol strings based on information stored in the input / output symbol string arrangement storage unit. After the generation rule, it is possible to cascade the hypothesis generation rule having the word string as an input, and the versatility of the symbol string conversion method can be improved. Further, according to the speech recognition method using the symbol string conversion method of the present invention, it is possible to easily add processing such as paraphrase processing even when a class language model is used, for example.

The figure which shows the function structural example of the symbol string converter 100 of this invention. The figure which shows the rough operation | movement flow of the symbol string converter 100. FIG. The figure which shows the function structural example of the symbol string conversion apparatus 200 of this invention. Alphabet - shows the hypothesis generation rule T1 stored in the Hiragana conversion rule storing unit 210 _1. Hiragana - shows the hypothesis generation rule T2 stored in kanji conversion rule storing unit 210 _2. Hiragana - shows the hypothesis generation rule T3 stored in the katakana conversion rule storing unit 210 _3. The figure which shows the example of the information memorize | stored in the input / output symbol sequence arrangement | positioning memory | storage part 15. FIG. The figure which shows the operation | movement flow of the symbol string converter 200 of this invention. The figure which shows the operation | movement flow of the symbol string converter 200 of this invention. The figure which shows the operation | movement flow of the symbol string converter 200 of this invention. The figure which shows the operation | movement flow of the symbol string converter 200 of this invention. The figure which shows the function structural example of the speech recognition apparatus 300 of this invention. The figure which shows the example of the information memorize | stored in the input-output symbol sequence arrangement | positioning memory | storage part 15 'of the speech recognition apparatus 300. FIG. The figure which shows the function structure of the conventional symbol string converter 900. FIG.

  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 a functional configuration example of the symbol string conversion apparatus 100 of the present invention. The symbol string conversion apparatus 100 includes M (M ≧ 2) cascaded hypothesis generation units 10 ₁ to 10 _M and M hypothesis correction units 12 associated with the respective hypothesis generation units 10 ₁ to 10 _{M. 1} to 12 _M and, each hypothesis correction unit ₁₂ 1 and the M hypothesis refining unit ₁₃ 1 to 13 _M associated with to 12 _M, 1 stage is provided between the cascade-connected are hypothesis generating unit ₁₀ 1 to 10 _M An input symbol string selector 14 _{1 to} 14 _M- that selects an input symbol string of the next-stage hypothesis generation unit with an input symbol string of the hypothesis generation unit of the eye and a symbol string output from the hypothesis generation unit before the previous stage as inputs. ₁ and input symbol string selection information selected by the input symbol string selection units 14 _{1 to} 14 _M−1 and output selection information indicating whether or not the hypotheses extracted by the hypothesis narrowing-down units 13 _{1 to} 13 _M are output. The input symbol string arrangement storage unit 15 and the time-series operation of each unit are controlled. That a control unit 16 comprises a. The symbol string conversion apparatus 100 is realized by reading a predetermined program into a computer composed of, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

FIG. 2 shows a schematic operation flow of the symbol string conversion apparatus 100 and its operation will be described. 1 stage hypothesis generation unit 11 ₁ generates in accordance with the input symbol hypothesis generation rule hypothesis f ₁ is a state transition process capable state transitions for a column input (step S11a, 11b). (·) In each step represents a corresponding operation step in a detailed operation flowchart described later.

The input symbol selection unit 14 _m refers to the input / output symbol string arrangement storage unit 15 and inputs it to the first-stage hypothesis generation unit 11 ₁ , the preceding stage, and the hypothesis generation units 11 _{1 to} 11 before that. m-th stage selects one of the output symbol string output from the _m-1 (m is an integer between 1 to m) as input symbol string hypotheses generator 11 _m of, m-th hypothesis generation unit 11 _m There repeated process of generating an auxiliary hypothesis _{f m,} to M-th stage while incrementing the m (step S12a, 12b, 12c). First stage of the input symbol selecting section 14 _1, the input symbol string and hypothesis generation unit 11 ₁ of one of the output to the output symbol string, the hypothesis generator reference selected by the input and output symbol string arranged storage section 15 112 output to ₂ . The input symbol selection unit 14 _{M-1 at the} final stage includes an input symbol string to be input to the first-stage hypothesis generation unit 11 ₁ and a plurality of outputs from the preceding and previous hypothesis generation units 11 _{1 to} 11 _M−1. one symbol string from the symbol string, and outputs the hypothesis generator 11 _M reference selected by the input and output symbol string arranged storage unit 15.

Each hypothesis corrector 12 _m is the weight of the auxiliary hypothesis f _m-1 of the previous stage in the cumulative weight of the auxiliary hypothesis f _m corrected, decrements m until correcting the cumulative weight of the hypothesis f ₁ the cumulative weight of the auxiliary hypothesis (Steps S12d and 12c). Here, the correction is selected by referring to the input / output symbol string arrangement storage unit 15 that uses the hypothesis as an auxiliary hypothesis for the hypothesis representing one state transition process of the hypothesis generation unit 11 _m-1 in the previous stage. possible in any hypothesis or subsequent hypothesis generation unit 11 _m when the output symbol string of the auxiliary hypothesis as input symbol string of the subsequent hypothesis generation unit 11 _m of hypothesis generation unit _{11 1} ~11 m-1 to be In the state transition process, the state transition having the minimum cumulative weight is obtained, and the cumulative weight is added to the hypothetical cumulative weight.

The hypothesis narrowing-down unit 13 _m extracts a hypothesis or auxiliary hypothesis in which the corrected cumulative weight exhibits a singular value, uses the extracted hypothesis or auxiliary hypothesis as an input to the next-stage hypothesis narrowing-down unit 13 _{m + 1} , and the final-stage temporary narrowing-down unit It outputs a 13 _M extracted ancillary hypothesis outside (step S13a). Hypothesis refining unit 13 _m is hypothesized or auxiliary hypothesis was extracted with reference to output symbol sequence arranged storage unit 15 may be output directly to the outside (step S13b).

According to the symbol string converter 100 as described above, select an input symbol string of input symbol selecting section 14 _m is next hypothesis generation unit 11 _m on the basis of the information stored in the output symbol string arranged storage section 15 Therefore, for example, after generating a hypothesis using a hypothesis generation rule for hiragana-kanji conversion, it is possible to provide a hypothesis generation unit that uses a hypothesis generation rule for hiragana-katakana conversion. Therefore, the versatility of the symbol string conversion method can be improved.

FIG. 3 shows an example of the functional configuration of the symbol string converter 200, and the present invention will be described more specifically. Symbol string transformer 200 first stage hypothesis generator 211 ₁ of hypothesis generation rules alphabet - Hiragana conversion rule (T1: 4), 2-stage hypothesis generator 211 ₂ of hypothesis generation rules Hiragana - Kanji conversion rule (T2: 5), the third stage of the final stage of hypothesis generation unit 211 ₃ of hypothesis generation rules Hiragana - Katakana conversion rule (T3: 6), in which a. Each conversion rule T1~T3 is alphabetical - Hiragana conversion rule storing unit 210 ₁ and the Hiragana - kanji conversion rule storing unit 210 ₂ and hiragana - stored katakana conversion rule storing unit 210 _3. FIG. 7 shows information stored in the input / output symbol string arrangement storage unit 215. The first column of the information is the stage number of the hypothesis generation units 211 _{1 to} 211 ₃ , the second column is the stage number of the output symbol string of the hypothesis generation unit that becomes the input symbol string of each hypothesis generation unit, and the third column is the hypothesis narrowing down A value indicating whether the part outputs an auxiliary hypothesis.

The operation will be described by taking as an example a case where “a, m, e, d, a, m, a” is input to the symbol string converter 200 as an input symbol string. The operation flow is shown in FIGS. The current state number of the hypothesis generation rule T1 is expressed as (s ₁ , O, W, L) where s _{1 is} an output symbol string, O is a cumulative weight, and L is an auxiliary hypothesis list. The auxiliary hypothesis in the m-th stage is similarly expressed as (s _m , O, W, L _m ). The m-th stage auxiliary hypothesis list is represented as {(s _m , O, W, L _m ), (s _m ′, O ′, W ′, L _m ′),.

  When the symbol string conversion device 200 starts operating (step S401), the control unit 216 creates empty hypothesis lists H and H ′ in, for example, RAM, which is a storage means of a computer constituting the symbol string conversion device 200. (Step S402). Then, a hypothesis h = (0, φ, 0, {(0, φ, 0, φ)}) is inserted into the hypothesis list H (step S403).

[Processing of “a”]
Hypothesis generation unit 211 ₁ reads the symbol "a" of the input symbol string (step S404). Then, a hypothesis (0, φ, 0, {(0, φ, 0, φ)}) is extracted from the hypothesis list H, and the state transition <0 where the input symbol is “a” from the current state 0 of the alphabet-hiragana conversion rule. A list E ₁ including 0, a: a / 0> is created (step S405). At this time, m = 1 is set.

Since the hypothesis generation unit 211 ₁ is not the list E ₁ = φ (no content), the state transition <0 → 0, a: a / 0> is extracted from the list E ₁ (step S407), and a new hypothesis f ₁ = (0, Oh, 0, {(0, φ, 0, {(0, φ, 0, φ)})} is generated (step S408).

If hypothesis _{f 1} is generated, the second stage of the hypothesis generator 211 ₂ starts operating (step S409). At this time, m = 2 is set. Input symbol selection unit 214 _1, output symbol since the input symbol column number n of reference column disposed storage unit 215 m = 2 is n = 1 (FIG. 7), the symbols of the output of the hypothesis generator 211 ₁ The column o [e ₁ ] is selected and input to the hypothesis generation unit 211 ₂ . Hypothesis correcting unit 212 ₁ generates an empty auxiliary hypothesis list _{L 2} (step S502). Since the input symbol string o _{[e 1]} is not epsilon (empty), the flow proceeds to step S503 is a process of hypothesis generation unit 211 _2.

Hypothesis generation unit 211 ₂ performs the calculation of equation (1) determine the correction term AW _{[f 1]} weights.

In this case, AW [f ₁ ] = 0. Hypothesis generation unit 211 _2, an auxiliary hypothesis list _{L [f 1] = {(} 0, φ, 0, {0, φ, 0, φ})} auxiliary hypothesis from _{g 2 = (0, φ,} 0, {0 , Φ, 0, φ}), and from the hypothesis generation rule T2 (Hiragana-Kanji conversion rule storage unit 2102), the transitionable state transition <0 → 1, a with the input symbol o [e ₁ ] = “a” : rain / 0>, <0 → 2 , Oh: to create a list _{E 2,} including a candy / 0> (step S504).

Because the E ₂ = φ (there is no content) not (N in step S505), the state transition from the list _{E 2} <0 → 1, Oh: rain / 0> is taken out as a list _{e 2} (step S506), auxiliary hypothesis f ₂ = (1, rain, 0 {(0, φ, 0, φ)} is generated (step S507).

As m = 3 and m is incremented since at this point it is m = 2, M = 3 Since m <M, 3-stage hypothesis generator 211 ₃ starts operating. Input symbol selector 214 _2, output symbol since the input symbol column number n of reference column disposed storage unit 215 m = 3 is n = 1 (FIG. 7), the symbols of the output of the hypothesis generator 211 ₁ input to the hypothesis generator 211 _3. select a column o _{[e 1].} The hypothesis correction unit 212 ₃ generates an empty auxiliary hypothesis list L ₃ (step S502). Since the input symbol string o _{[e 1]} is not epsilon (empty), the flow proceeds to step S503 is a process of hypothesis generation unit 211 _3.

Hypothesis generation unit 211 ₃ performs the calculation of equation (1) determine the correction term AW _{[f 2]} in weight. In this case, AW [f ₂ ] = 0. The hypothesis generation unit 211 ₃ takes out the auxiliary hypothesis g ₃ = (0, φ, 0, φ) from the auxiliary hypothesis list L [f ₂ ] = {(0, φ, 0, φ)}, and generates a hypothesis generation rule T3 ( A list E ₃ including a transitionable state transition <0 → 0, a: a / 0> with the input symbol o [e ₁ ] = “a” is created from the hiragana-katakana conversion rule storage unit 210 ₃ ) (step S3) S504).

E ₃ = φ since (content is not) not (step S505 N), the state transition from the list _{E 3} <0 → 0, Oh: A / 0> is taken out as a list _{e 3,} auxiliary hypothesis _f 3 = ( 0, a, 0, φ) are generated (step S507).

Since not m = 3, M = 3 Since m <M at step S508, the next stage of hypothesis generation unit does not exist, the hypothesis generator 211 ₃ inserts an auxiliary hypothesis _{f 3} to the auxiliary hypothesis list _{L 3} (step S509). As a result list _{E 3} is empty (φ).

When the list E ₃ = φ, the hypothesis correction unit 212 ₃ starts to operate. Since the auxiliary hypothesis list L [f ₂ ] = φ, the element {0, a, 0, φ)} of the auxiliary hypothesis list L ₃ is added to the auxiliary hypothesis list L [f ₂ ] of the hypothesis correction unit 212 _{2 in} the previous stage. (Step S511).

Since the cumulative weight of the auxiliary hypothesis in the auxiliary hypothesis list L [f ₂ ] is 0, after setting W [f ₂ ] = 0, the auxiliary hypothesis f ₂ = (1, rain, 0 {(0, φ, 0, φ)} Is completed, m is decreased by 1 to m = 2 (step S513), and the process returns to step S508 (step S514).

After the correction procedure _{f 2} at step S508, the process proceeds to step S509. At step S509, the inserting the _{f 2} to _{L 2,} the flow returns to step S505. Since E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from the list _{E 2} <0 → 2, Ah: candy / 0> and List _{e 2} removed, auxiliary hypothesis _f 2 = (2 in step S507, the candy, 0, {(0, φ , 0, φ)} ) Is generated. In step S508, the flow advances since it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is not φ, the process proceeds to step S503. . In step S503, AW [f ₂ ] = 0.

In step S504, the auxiliary hypothesis g ₃ = (0, φ, 0, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, φ, 0, φ)} and input from state 0 of the hypothesis generation rule T3. A list E ₃ including a transitionable state transition <0 → 0, a: a / 0> with the symbol o [e ₁ ] = “a” is created.

Since E ₃ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from E ₃ <0 → 0, Ah: A / 0> and _{e 3} removed, auxiliary hypothesis _f 3 = at step S507 (0, A, 0, phi) to produce a. Not a m = 3 and M = 3 Since m <M at step S508, the process proceeds to step S509, the inserting auxiliary hypothesis _{f 3} Listing _{L 3.} Returning to step S505, since E ₃ = φ, the process proceeds to step S510.

In step S510, since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511. In step S511, in step S512 transferred to the auxiliary hypothesis list _{L 3} elements all auxiliary hypotheses list L _{[f 2],} the cumulative weight of L _{[f 2]} Minimum cumulative weight during the auxiliary hypothesis auxiliary hypothesis _{g 3} 'is Since 0, W [f ₂ ] = 0. Step S513 The m by reducing 1 in the m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S508.

After the correction procedure _{f 2} at step S508, the process proceeds to step S509. At step S509, the inserting the _{f 2} to _{L 2,} the flow returns to step S505. Since E ₂ = φ in step S505, the process proceeds to step S510. In step S510, since L [f ₁ ] = φ, the process proceeds to step S511. In step S511, all the elements of the auxiliary hypothesis list _{L 2} transferred to L _{[f 1].} In step S512, L [f ₁ ] = {(1, rain, 0, {(0, a, 0, φ)}), (2, 飴, 0, {(0, a, 0, φ)}) }, The cumulative weights of the auxiliary hypotheses are both 0, so W [f ₁ ] = 0. Exit correction auxiliary hypothesis _{f 1} at step S514, the flow returns to step S409.

After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410. In step S410, insert the hypothesis _{f 1} to the hypothesis list H ', the flow returns to step S406. Since the state transition list E ₁ = φ in step S406, the process proceeds to step S411. Since the hypothesis list H = φ in step S411, the process proceeds to step S412 and the elements of H ′ {(0, a, 0, {(1, rain, 0, {(0, a, 0, φ)}), (2 , 飴, 0, {(0, a, 0, φ)})})} are moved to the hypothesis list H. Proceeding to step S413, since the next input symbol exists, the process returns to step S404.

[Processing of “m”]
In step S404, the symbol “m” is read. In step S405, a hypothesis h = (0, a, 0, {(1, rain, 0, {(0, a, 0, φ)}), (2 , 飴, 0, {(0, a, 0, φ)})}). A list E ₁ including a state transition <0 → 1, m: ε / 0> with an input symbol “m” is created from the current state 0 of the hypothesis generation rule T1 of this hypothesis.

Since _{E 1} = not at φ at step S406 proceeds to step S407, the state transition list state transition from _{E 1 <0 → 0, m} : ε / 0> and _{e 1} removed, new hypotheses in step S408 _f 1 = ( 1, 2, 0, {(1, rain, 0, {(0, a, 0, φ)}), (2, 飴, 0, {(0, a, 0, φ)})}) To do. The process proceeds to step S501 to correct the hypothesis _{f 1} in step S409. At this time, m = 2 is set.

In step S502 (m = 2) produces an empty auxiliary hypothesis list _{L 2.} In step S521 (m = 2), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 2 with reference to the input / output symbol string arrangement storage unit 215 is ε, step S522 (m = Go to 2). In step S522 (m = 2), the auxiliary hypothesis list L [f ₁ ] = {(1, rain, 0, {(0, a, 0, φ)}), (2, 飴, 0, {(0, a , 0, φ)})} extracts the auxiliary hypothesis f ₂ = (1, rain, 0, {(0, a, 0, φ)}).

In step S523 (m = 2), proceeds from it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is ε, step S522 (m = Go to 3). In step S522 (m = 3), the auxiliary hypothesis f ₃ = (0, a, 0, φ)} is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, a, 0, φ)}.

In step S523 (m = 3), since at m = 3 not M = 3 Since m <M, the process proceeds to step S524 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.} In step S525 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511. In step S511 (m = 3), the elements of the auxiliary hypothesis list _{L 3} {(0, A, 0, φ)} all transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 0, so W [f ₂ ] = 0. Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S523 (m = 2).

After the correction procedure _{f 2} at step S523 (m = 2), the process proceeds to step S524 (m = 2). In step S524 (m = 2) inserting the _{f 2} to _{L 2.} Since L [f ₁ ] is not φ in step S525, the process returns to step S522 (m = 2). In step S522 (m = 2), from the auxiliary hypothesis list L [f ₁ ] = {(2, 飴, 0, {(0, a, 0, φ)})}, the auxiliary hypothesis f ₂ = (2, 飴, 0 , {(0, a, 0, φ)}). In step S523 (m = 2), proceeds from it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is ε, step S522 (m = Go to 3). In step S522 (m = 3), the auxiliary hypothesis f ₃ = (0, a, 0, φ)} is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, a, 0, φ)}.

In step S523 (m = 3), since at m = 3 not M = 3 Since m <M, the process proceeds to step S524 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.} In step S525 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511. In step S511 (m = 3), the elements of the auxiliary hypothesis list _{L 3} {(0, A, 0, φ)} all transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 0, so W [f ₂ ] = 0. Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S523 (m = 2). After the correction procedure _{f 2} at step S523 (m = 2), the process proceeds to step S524 (m = 2). In step S524 (m = 2), insert the _{f 2} to _{L 2,} the flow returns to step S522 (m = 2).

In step S522 (m = 2), from the auxiliary hypothesis list L [f ₁ ] = {(2, 飴, 0, {(0, a, 0, φ)})}, the auxiliary hypothesis f ₂ = (2, 飴, 0 , {(0, a, 0, φ)}). Since L [f ₁ ] = φ in step S525 (m = 2), the process proceeds to step S511 (m = 2). In step S511 (m = 2), all the elements of the auxiliary hypothesis list _{L 2} transferred to L _{[f 1].} In step S512 (m = 2), L [f ₁ ] = {(1, rain, 0, {(0, a, 0, φ)}), (2, 飴, 0, {(0, a, 0 , Φ)})}, the cumulative weights of the auxiliary hypotheses are both 0, so W [f ₁ ] = 0. In step S514 (m = 2) terminates the correction of the auxiliary hypothesis _{f 1,} the flow returns to step S409.

After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410. In step S410 inserts the hypothesis _{f 1} to the hypothesis list H ', processing proceeds to step S406. In step S406, since state transition list E ₁ = φ, the process proceeds to step S411, and since hypothesis list H = φ, the process proceeds to step S412. In step S412, the elements {(1, a, 0, {(1, rain, 0, {(0, a, 0, φ)}), (2, 飴, 0, {(0, a) , 0, φ)})})} are all moved to the hypothesis list H. Proceeding to step S413, since the next input symbol exists, the process returns to step S404.

[Processing of “e”]
In step S404, the symbol “e” is read, and in step S405, a hypothesis h = (1, a, 0, {(1, rain, 0, {(0, a, 0, φ)}), (2 , 飴, 0, {(0, a, 0, φ)})}). A state transition list E ₁ including a state transition <1 → 0, e: me / 0> with an input symbol “e” is created from the current state 0 of the hypothesis generation rule T1 of this hypothesis.

Since _{E 1} = not a φ in step S406 proceeds to the step S407, the state transition from _{E 1:} the _{e 1} Remove the <1 → 0, e because / 0>, a new hypothesis in step S408 _f 1 = (0, rain , 0, {(1, rain, 0, {(0, a, 0, φ)}), (2, 飴, 0, {(0, a, 0, φ)})}). The process proceeds to step S501 to correct the hypothesis _{f 1} in step S409. At this time, m = 2 is set.

To generate an empty auxiliary hypothesis list _{L 2} in step S502. In step S521, since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 2 with reference to the input / output symbol string arrangement storage unit 215 is not ε, the process proceeds to step S503. f ₁ ] = 0.
In step S504, auxiliary hypothesis list L [f ₁ ] = {(1, rain, 0, {(0, a, 0, φ)}), (2, 飴, 0, {(0, a, 0, φ) })} Is extracted from the auxiliary hypothesis g ₂ = (1, rain, 0, {(0, a, 0, φ)}), and the state transition that can be transitioned from the state 1 of the hypothesis generation rule T2 with the input symbol “M” <1 → 5, because: ε / 1> make a state transition list _{E 2,} including.

Since the state transition E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition list _{E 2} state transition from <1 → 5, because: ε / 1> and _{e 2} taken _{out, W [f 2] = 0} + 0 + 1 = 1 is an auxiliary hypothesis _f 2 = (5 in step S507, the rain, 1 , {(0, a, 0, φ)}).

In step S508, the flow advances since it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is not φ, the process proceeds to step S503. . In step S503 (m = 3), AW [f ₂ ] = 1-0 = 1.

In step S504 (m = 3), the auxiliary hypothesis g ₃ = (0, a, 0, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, a, 0, φ)}, and the hypothesis generation rule T3 A list E ₃ including a transitionable state transition <0 → 0, second: me / 0> with an input symbol of o [e ₁ ] = “m” is created from the state 0 of FIG. Since E ₃ = φ is not satisfied in step S505 (m = 3), the process proceeds to step S506 (m = 3). State transition from E ₃ <0 → 0, because: main / 0> and _{e 3} removed, step S507 (m = 3) in _{W [f 3] = 0 +} 1 + 0 = 1 and comprising an auxiliary hypothesis _f 3 = (0, candy , 1, φ).

Not a m = 3 and M = 3 Since m <M at step S508 (m = 3), the process proceeds to step S509 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.} Returning to step S505 (m = 3), since E ₃ = φ, the process proceeds to step S510 (m = 3). In step S510 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511 (m = 3). In step S511 (m = 3), all the elements of the auxiliary hypothesis list _{L 3} is transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 1, so W [f ₂ ] = 1. Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S508.

After the correction procedure _{f 2} at step S508, the process proceeds to step S509. At step S509, the inserting the _{f 2} to _{L 2,} the flow returns to step S505. Since E ₂ = φ in step S505, the process proceeds to step S510. In step S510, since L [f ₁ ] = φ is not satisfied, the process returns to step S504.

In step S504, from the auxiliary hypothesis list L [f ₁ ] = {(2, 飴, 0, {(0, a, 0, φ)})}, the auxiliary hypothesis g ₂ = (2, 飴, 0, {(0, A, 0, φ)}) is taken out, and a state transition list E ₂ including a state transition <2 → 5, fifth: ε / 2> that can be transitioned from the state 2 of the hypothesis generation rule T2 with the input symbol “m” is created. .

Since the state transition E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition <2 → 5, fifth: ε / 2> is extracted from state transition list E ₂ and set as e _2, and auxiliary hypothesis f ₂ = (5, 飴, 2) where W [f ₂ ] = 0 + 0 + 2 = 2 in step S507. , {(0, a, 0, φ)}).

In step S508, the flow advances since it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set. In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is not φ, the process proceeds to step S503. . In step S503 (m = 3), AW [f ₂ ] = 2-0 = 2.

In step S504 (m = 3), the auxiliary hypothesis g ₃ = (0, a, 0, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, a, 0, φ)}, and the hypothesis generation rule T3 A list E ₃ including a transitionable state transition <0 → 0, second: me / 0> with an input symbol of o [e ₁ ] = “m” is created from the state 0 of FIG.

Since E ₃ = φ is not satisfied in step S505 (m = 3), the process proceeds to step S506 (m = 3). State transition from E ₃ <0 → 0, because: main / 0> and _{e 3} removed, step S507 (m = 3) in _{W [f 3] = 0 +} 2 + 0 = 2 to become the auxiliary hypothesis _f 3 = (0, candy , 2, φ).

Not a m = 3 and M = 3 Since m <M at step S508 (m = 3), the process proceeds to step S509 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.} Returning to step S505 (m = 3), since E ₃ = φ, the process proceeds to step S510 (m = 3). In step S510 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511 (m = 3). In step S511 (m = 3), all the elements of the auxiliary hypothesis list _{L 3} is transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 1, so W [f ₂ ] = 1. Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S508. After the correction procedure _{f 2} at step S508, the process proceeds to step S509. At step S509, the inserting the _{f 2} to _{L 2,} the flow returns to step S505. However, since an auxiliary hypothesis (5, rain, 1, {(0, candy, 1, φ)}) already exists in L ₂ and the arrival state in the hypothesis generation rule T2 is 5, auxiliary with a small cumulative weight The hypothesis (5, rain, 1, {(0, candy, 1, φ)}) is left, and (5, 飴, 2, {(0, candy, 2, φ)}) is deleted. At this time, the auxiliary hypothesis list of each auxiliary hypothesis is connected by correcting the cumulative weight (5, rain, 1, {(0, candy, 1, φ), (0, candy, 2, φ)}). To do. Further, since the arrival state in the hypothesis generation rule T3 is 0 in the auxiliary hypothesis list of this auxiliary hypothesis, the auxiliary hypothesis (0, candy, 1, φ) having a small cumulative weight is left, and (5, rain, 1, {( 0, candy, 1, φ)}).

Further, since the auxiliary hypothesis that has reached state 5 reaches state 0 without any input symbol via state transition <5 → 0, ε: ε / 0>, the auxiliary hypothesis list L ₂ includes auxiliary hypotheses ( 0, rain, 1, {(0, candy, 1, φ)}) remain. Since E ₂ = φ in step S505, the process proceeds to step S510. In step S510, since L [f ₁ ] = φ, the process proceeds to step S511. In step S511, transferred all the elements of the _{L 2} to L _{[f 1].} In step S512, since the cumulative weight of the auxiliary hypothesis of L [f ₁ ] is 1, W [f ₁ ] = 1. Exit correction auxiliary hypothesis _{f 1} at step S514, the flow returns to step S409. After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410. In step S410, insert the hypothesis _{f 1} to the hypothesis list H ', the flow returns to step S406. Since the state transition list E ₁ = φ in step S406, the process proceeds to step S411. Since the hypothesis list H = φ in step S411, the process proceeds to step S412 and the element {(0, rain, 1, {(0, rain, 1, {(0, candy, 1, φ)})} of H ′ is hypothesized. Move to list H. Proceed to step S413, and the next input symbol exists, so return to step S404.

[Processing of “d”]
In step S404, the symbol “d” is read, and in step S405, a hypothesis h = (0, rain, 1, {(0, rain, 1, {(0, candy, 1, φ)}) is extracted from the hypothesis list H. A list E ₁ including a state transition <0 → 3, d: ε / 0> whose input symbol is “m” is created from the current state 0 of the hypothesis generation rule T1 of this hypothesis, because E ₁ = φ is not satisfied in step S406. the process proceeds to step S407, the state transition list state transition from _{E 1 <0 → 3, d} : ε / 0> and _{e 1} removed, new hypotheses in step S408 _f 1 = (3, rain, 1, {(0, rain, 1, {(0, candy, 1, phi)}) for generating a. proceeds to step S501 to correct the hypothesis _{f 1} at step S409. setting this time m = 2.

In step S502 (m = 2) produces an empty auxiliary hypothesis list _{L 2.} In step S521 (m = 2), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 2 with reference to the input / output symbol string arrangement storage unit 215 is ε, step S522 (m = Go to 2). In step S522 (m = 2), auxiliary hypothesis list L [f ₁ ] = {(0, rain, 1, {(0, candy, 1, φ)} to auxiliary hypothesis f ₂ = (0, rain, 1, { (0, candy, 1, phi) taken out. step S523 in (m = 2), since at m = 2 is M = 3 since m <M, step S501 (m = 3 to correct the auxiliary hypothesis _{f 2} At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is ε, step S522 (m = Go to 3). In step S522 (m = 3), the auxiliary hypothesis f ₃ = (0, candy, 1, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, candy, 1, φ)}. In step S523 (m = 3), since at m = 3 not M = 3 Since m <M, the process proceeds to step S524 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.} In step S525 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511.

In step S511 (m = 3), the elements of the auxiliary hypothesis list _{L 3} {(0, candy, 1, φ)} all transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 1, so W [f ₂ ] = 1. Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S523 (m = 2).

After the correction procedure _{f 2} at step S523 (m = 2), the process proceeds to step S524 (m = 2). In step S524 (m = 2) inserting the _{f 2} to _{L 2.} Since L [f ₁ ] = φ in step S525 (m = 2), the process proceeds to step S511 (m = 2). In step S511 (m = 2), all the elements of the auxiliary hypothesis list _{L 2} transferred to L _{[f 1].} In step S512 (m = 2), since the cumulative weight of the auxiliary hypothesis of L [f ₁ ] = {(0, rain, 1, {(0, candy, 1, φ)})} is 1, W [f ₁ ] = 1. In step S514 (m = 2) terminates the correction of the auxiliary hypothesis _{f 1,} the flow returns to step S409.

After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410. In step S410 inserts the hypothesis _{f 1} to the hypothesis list H ', processing proceeds to step S406. In step S406, since state transition list E ₁ = φ, the process proceeds to step S411, and since hypothesis list H = φ, the process proceeds to step S412. In step S412, all the elements {(3, candy, 1, {(0, rain, 1, {(0, candy, 1, φ)})})} of the hypothesis list H ′ are moved to the hypothesis list H. Proceeding to step S413, since the next input symbol exists, the process returns to step S404.

[Processing of “a”]
In step S404, the symbol “a” is read, and in step S405, a hypothesis h = (3, candy, 1, {(0, rain, 1, {(0, candy, 1, φ)})}) is obtained from the hypothesis list H. Take out. A state transition list E ₁ including a state transition <3 → 0, a: da / 0> whose input symbol is “e” is created from the current state 0 of the hypothesis generation rule T1 of this hypothesis. Because it is not a E1 = φ in step S406 proceeds to the step S407, the state transition from _{E 1:} the _{e 1} Remove the <1 → 0, e because / 0>, a new hypothesis in step S408 _f 1 = (0, but rain , 1, {(0, rain, 1, {(0, candy, 1, φ)})}). The process proceeds to step S501 to correct the hypothesis _{f 1} in step S409. At this time, m = 2 is set.

To generate an empty auxiliary hypothesis list _{L 2} in step S502. In step S521, since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 2 with reference to the input / output symbol string arrangement storage unit 215 is not ε, the process proceeds to step S503. f ₁ ] = 0.
In step S504, the auxiliary hypothesis list L [f ₁ ] = {(0, rain, 1, {(0, rain, 1, φ)})} is supplemented by the auxiliary hypothesis g ₂ = (0, rain, 1, {(0, Candy, 1, φ)}), and includes state transition <0 → 4: ball / 1> that can transition from state 1 of hypothesis generation rule T2 with input symbol o [e ₁ ] = “da” make the transition list _{E 2.}

Since the state transition E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from the state transition list _{E 2} <0 → 4, it is: Ball / 1> and _{e 2} taken out, is a _{W [f 2] = 1 +} 0 + 1 = 2 auxiliary hypothesis _f 2 = in step S507 (4, hard candy, 2, {(0, candy, 1, φ)}). In step S508, the flow advances since it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is not φ, the process proceeds to step S503. . In step S503 (m = 3), AW [f ₂ ] = 2−1 = 1.

In step S504 (m = 3), the auxiliary hypothesis g ₃ = (0, candy, 1, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, candy, 1, φ)}, and the hypothesis generation rule T3 A list E ₃ including a transitionable state transition <0 → 0, where: da / 0> with the input symbol o [e ₁ ] = “da” is created from the state 0 of FIG.

Since E ₃ = φ is not satisfied in step S505 (m = 3), the process proceeds to step S506 (m = 3). State transition from E ₃ <0 → 0, it: Da / 0> and _{e 3} removed, step S507 (m = 3) in _{W [f 3] = 1 +} 1 + 0 = 2 to become the auxiliary hypothesis _f 3 = (0, Ameda , 2, φ).

Not a m = 3 and M = 3 Since m <M at step S508 (m = 3), the process proceeds to step S509 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.} Returning to step S505 (m = 3), since E ₃ = φ, the process proceeds to step S510 (m = 3). In step S510 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511 (m = 3). In step S511 (m = 3), all the elements of the auxiliary hypothesis list _{L 3} is transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 2, so W [f ₂ ] = 2.

Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S508. After the correction procedure _{f 2} at step S508, the process proceeds to step S509. At step S509, the inserting the _{f 2} to _{L 2,} the flow returns to step S505. Since E ₂ = φ in step S505, the process proceeds to step S510. In step S510, since L [f ₁ ] = φ, the process proceeds to step S511. In step S511, transferred all the elements of the _{L 2} to L _{[f 1].} In step S512, since the cumulative weight of the auxiliary hypothesis of L [f ₁ ] is 2, W [f ₁ ] = 2 is set. Exit correction auxiliary hypothesis _{f 1} at step S514, the flow returns to step S409. After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410. In step S410, insert the hypothesis _{f 1} to the hypothesis list H ', the flow returns to step S406. Since the state transition list E ₁ = φ in step S406, the process proceeds to step S411. Since the hypothesis list H = φ in step S411, the process proceeds to step S412 and the element of H ′ {(0, Ameda, 2, {(4, rainball, 2, {(0, Ameda, 2, φ)})} Is moved to the hypothesis list H. The process proceeds to step S413, and since the next input symbol exists, the process returns to step S404.

[Processing of “m”]
In step S404, the symbol “d” is read, and in step S405, the hypothesis h = (0, Ameda, 2, {(4, rainball, 2, {(0, AMeDA, 2, φ)}) is read from the hypothesis list H. A list E ₁ including a state transition <0 → 1, m: ε / 0> having an input symbol “m” is created from the current state 0 of the hypothesis generation rule T1 of this hypothesis.

Since _{E 1} = not at φ at step S406 proceeds to step S407, the state transition list state transition from _{E 1 <0 → 1, m} : ε / 0> and _{e 1} removed, new hypotheses in step S408 _f 1 = ( 1, but rain, 2, proceed to {(4, candy, 2, {(0, Ameda, 2, phi)}) for generating a. step S501 to correct the hypothesis _{f 1} at step S409. in this case m = Set to 2.

In step S502 (m = 2) produces an empty auxiliary hypothesis list _{L 2.} In step S521 (m = 2), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 2 with reference to the input / output symbol string arrangement storage unit 215 is ε, step S522 (m = Go to 2). In step S522 (m = 2), from the auxiliary hypothesis list L [f ₁ ] = {(4, rainball, 2, {(0, AMeDA, 2, φ)})}, the auxiliary hypothesis f ₂ = (4, rainball) , 2, {(0, AMeDA, 2, φ)}). In step S523 (m = 2), proceeds from it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is ε, step S522 (m = Go to 3). In step S522 (m = 3), the auxiliary hypothesis f ₃ = (0, AMeDA, 2, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, AMeDA, 2, φ)}. In step S523 (m = 3), since at m = 3 not M = 3 Since m <M, the process proceeds to step S524 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.}

In step S525 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511.

In step S511 (m = 3), the elements of the auxiliary hypothesis list _{L 3} {(0, Ameda, 2, φ)} all transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 1, so W [f ₂ ] = 1. Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S523 (m = 2). After the correction procedure _{f 2} at step S523 (m = 2), the process proceeds to step S524 (m = 2). In step S524 (m = 2) inserting the _{f 2} to _{L 2.}

Since L [f ₁ ] = φ in step S525 (m = 2), the process proceeds to step S511 (m = 2). In step S511 (m = 2), all the elements of the auxiliary hypothesis list _{L 2} transferred to L _{[f 1].} In step S512 (m = 2), the cumulative weight of the auxiliary hypothesis of L [f ₁ ] = {(4, rainball, 2, {(0, AMeDA, 2, φ)})} is 2, so W [f ₁ ] = 2. In step S514 (m = 2) terminates the correction of the auxiliary hypothesis _{f 1,} the flow returns to step S409. After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410. In step S410 inserts the hypothesis _{f 1} to the hypothesis list H ', processing proceeds to step S406.

In step S406, since state transition list E ₁ = φ, the process proceeds to step S411, and since hypothesis list H = φ, the process proceeds to step S412. In step S412, all the elements {(1, candy, 2, {(4 rainball, 2, {(0, AMeDA, 2, φ)})} of the hypothesis list H ′ are moved to the hypothesis list H. Step S413 Since the next input symbol exists, the process returns to step S404.

[Processing of “a”]
In step S404, the symbol “a” is read, and in step S405, the hypothesis h = (1, candy, 2, {(4, rainball, 2, {(0, AMeDA, 2, φ)}) is read from the hypothesis list H. A state transition list E ₁ including a state transition <1 → 0, a: ma / 0> with an input symbol “e” is created from the current state 0 of the hypothesis generation rule T1 of this hypothesis.

Since _{E 1} = not a φ in step S406 proceeds to the step S407, the state transition from _{E 1 <1 → 0, a} : or / 0> and _{e 1} taken out, a new hypothesis in step S408 _f 1 = (0, rain setting the lumps, 2, {(4, candy, 2, {(0, Ameda, 2, phi)}) for generating a. proceeds to step S501 to correct the hypothesis _{f 1} at step S409. in this case m = 2 To do.

To generate an empty auxiliary hypothesis list _{L 2} in step S502. In step S521, since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 2 with reference to the input / output symbol string arrangement storage unit 215 is not ε, the process proceeds to step S503. f ₁ ] = 0.

In step S504, the auxiliary hypothesis list L [f ₁ ] = {(4, rainball, 2, {(0, AMeDA, 2, φ)} is used, and the auxiliary hypothesis g ₂ = (4, rainball, 2, {(0, State transition list including state transitions <4 → 5 or ε / 1> that can be transitioned from state 1 of hypothesis generation rule T2 with input symbol o [e ₁ ] = “ma”. make the E _2.

Since the state transition E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition list _{E 2} state transition from <4 → 5, or: epsilon / 1> and _{e 2 removed, W [f 2] = 2} + 0 + 1 = 3 a is an auxiliary hypothesis _f 2 = (5 in step S507, the rain ball, 3, {(0, AMeDA, 2, φ)}). In step S508, the flow advances since it is M = 3 Since m <M with m = 2, the step S501 of correcting the auxiliary hypothesis _{f 2} (m = 3). At this time, m is increased by 1, and m = 3 is set.

In step S502 (m = 3) for generating an empty auxiliary hypothesis list _{L 3.} In step S521 (m = 3), since the output symbol o [e ₁ ] of the stage number n = 1 that is input to m = 3 with reference to the input / output symbol string arrangement storage unit 215 is not φ, the process proceeds to step S503. In step S503 (m = 3), AW [f ₂ ] = 3−2 = 1. In step S504 (m = 3), the auxiliary hypothesis g ₃ = (0, AMeDA, 2, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, AMeDA, 2, φ)}, and the hypothesis generation rule T3 A list E ₃ including a transitionable state transition <0 → 0, m: ma / 0> of the input symbol o [e ₁ ] = “ma” is created from the state 0 of FIG.

Since E ₃ = φ is not satisfied in step S505 (m = 3), the process proceeds to step S506 (m = 3). State transition from E ₃ <0 → 0, or: Ma / 0> and _{e 3} removed, step S507 (m = 3) in _{W [f 3] = 2 +} 1 + 0 = 3 and comprising the auxiliary hypothesis _f 3 = (0, Amedama , 3, φ). Not a m = 3 and M = 3 Since m <M at step S508 (m = 3), the process proceeds to step S509 (m = 3), inserting an auxiliary hypothesis _{f 3} Listing _{L 3.}
Returning to step S505 (m = 3), since E ₃ = φ, the process proceeds to step S510 (m = 3).

In step S510 (m = 3), since the auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511 (m = 3). In step S511 (m = 3), all the elements of the auxiliary hypothesis list _{L 3} is transferred to the auxiliary hypothesis list L _{[f 2].} In step S512 (m = 3), the cumulative weight of the auxiliary hypothesis of L [f ₂ ] is 3, so W [f ₂ ] = 3. Step S513 (m = 3) in to the m Reduce 1 and m = 2, and terminates the correction of the auxiliary hypothesis _{f 2} at step S514, the flow returns to step S508.

After the correction procedure _{f 2} at step S508, the process proceeds to step S509. At step S509, the inserting the _{f 2} to _{L 2,} the flow returns to step S505. Since E ₂ = φ in step S505, the process proceeds to step S510. In step S510, since L [f ₁ ] = φ, the process proceeds to step S511. In step S511, transferred all the elements of the _{L 2} to L _{[f 1].} In step S512, since the cumulative weight of the auxiliary hypothesis of L [f ₁ ] is 3, W [f ₁ ] = 3. Exit correction auxiliary hypothesis _{f 1} at step S514, the flow returns to step S409.

After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410. In step S410, insert the hypothesis _{f 1} to the hypothesis list H ', the flow returns to step S406. Since the state transition list E ₁ = φ in step S406, the process proceeds to step S411. Since the hypothesis list H = φ in step S411, the process proceeds to step S412 and the element of H ′ {(0, candy, 3, {(5, rainball, 3, {(0, Amedama, 3, φ)})} the process proceeds to. step S413 to move the hypothesis list H, since there is no next input symbol, the F _m of the flow proceeds to step S414. step S414, a set representing the end state of the m-th stage hypothesis generation rule.

In step S414, the hypothesis h _{1 in} H is equal to (0, Amedama, 3, {(5, Rainball, 3, {(0, Amedama, 3, φ)})}) Assuming that the auxiliary hypothesis h ₂ = (5, rainball, 3, {(0, AMeDama, 3, φ)}) and the auxiliary hypothesis h ₃ = (0, AMeDama, 3, φ), Cumulative weight W [h ₂ ] = 3 + (3-3) + 0 = 3, W [h ₃ ] = 3 + (3-3) + 0 = 3. In step S421, to output the resulting converted output symbol string "Amedama" of the input and output symbols with reference to the column placement storage unit 215 of the auxiliary hypothesis h ₂ output symbol string "candy" auxiliary hypothesis h _3, at step S415 The symbol string conversion process ends.

  In this way, when correcting the hypothesis score, the input symbol of the auxiliary hypothesis is input not only immediately before but also the output symbol of the previous hypothesis generator, so that the hypothesis that cannot be synthesized or cascaded originally. The generation rule can be made dependent, and the versatility of the symbol string converter can be improved. In addition, since it is possible to select whether or not to output a plurality of auxiliary hypotheses to the outside based on the information stored in the input / output symbol string arrangement storage unit 15, there is also an effect that a plurality of conversion results can be obtained simultaneously.

  A speech recognition device 300 in which the above-described symbol string conversion device of the present invention is applied to speech recognition using a class language model will be described. The speech recognition apparatus 300 is an embodiment in the case where a paraphrase process is added after the speech recognition apparatus using the class language model.

  A class language model is a word-based language model that gives ease of word connection (for example, probability value), while assigning similar words to groups called classes to give ease of class connection. Is. This is used when a text corpus for learning a language model is not sufficiently given or when a word not in the text corpus is added. The paraphrase process is to convert the utterance content into another sentence style while performing voice recognition. For example, a process of rewriting a spoken sentence into a written style.

FIG. 12 shows a functional configuration example of the speech recognition apparatus 300. The speech recognition apparatus 300 is configured by combining a symbol string conversion device 100 ′, which is a modification of the above-described symbol string conversion device 100, and a speech feature symbol string extraction unit 31. The symbol string conversion device 100 ′ has a five-stage configuration in which the number m of the hypothesis generation unit 11 is set to m = 5 with respect to the symbol string conversion device 100, and information stored in the input / output symbol string arrangement storage unit 15 ′ includes Only the part changed to the content corresponding to the number of stages is different. Therefore, in FIG. 12, only the hypothesis generation units 11 ₁ ′ to 11 ₅ ′ and the hypothesis generation rule storage units 10 ₁ ′ to 10 ₆ ′ accompanying the hypothesis generation units 11 ₁ ′ to 11 ₅ ′ are described, and the hypothesis correction unit 12 ′ and other hypothesis narrowing down units 13 'and the input / output symbol string arrangement storage unit 15' are omitted.

The voice feature symbol string extraction unit 31 receives a voice signal and extracts an acoustic feature symbol string from a time series of short-time acoustic patterns of voice. The hypothesis generation unit 11 ₁ ′ at the first stage outputs the speech signal of the standard feature of the speech fixed unit (for example, element) stored in the acoustic model hypothesis generation rule storage unit 10 ₁ ′ for a short time (for example, 10 milliseconds). An acoustic pattern obtained by analysis for each word, and a word utterance dictionary that associates pronunciations of various words stored in the word utterance dictionary hypothesis generation rule storage unit 10 ₂ ′ with a sequence of fixed speech units, A hypothesis f ₁ is generated in which the symbol string is a word-based symbol string.

The input symbol selection unit 14 ₁ ′ (not shown) refers to the input / output symbol string arrangement storage unit 15 ′ (FIG. 13), and the input symbol stage number n of the stage number m = 2 is n = 1. The word string output by 11 ₁ ′ is selected and input to the second-stage hypothesis generation unit 11 ₂ ′. The hypothesis generation unit 11 ₂ ′ generates the class to which the word string belongs as the auxiliary hypothesis f _{2 in association} with the class conversion rule to which the word stored in the class conversion hypothesis generation rule storage unit 10 ₃ ′ belongs. The hypothesis narrowing-down unit 13 ₁ ′ associated with the hypothesis generation unit 11 ₂ ′ outputs the extracted auxiliary hypothesis f ₂ to the outside because the output information is True with reference to the input / output symbol string arrangement storage unit 15 ′.

Since the input symbol stage number n of m = 3 in the input / output symbol string arrangement storage unit 15 ′ is n = 2, the input symbol selection unit 14 ₂ ′ has three classes of class strings output from the hypothesis generation unit 11 ₂ ′. This is input to the hypothesis generation unit 11 ₃ ′ of the eye. The hypothesis generation unit 11 ₃ ′ assigns a weight (probability) to the class string in association with the class language model conversion rule that gives the ease of connection of the class to which the word stored in the class language model hypothesis generation rule storage unit 10 ₄ ′ belongs. ) to generate the class sequence as auxiliary hypothesis f ₃ imparted with.

Since the input symbol stage number n of m = 4 in the input / output symbol string arrangement storage section 15 ′ is n = 1, the input symbol selection section 14 ₃ ′ has four stages of word strings output from the hypothesis generation section 11 ₁ ′. Input to the hypothesis generation unit 11 ₄ ′ of the eye. The hypothesis generation unit 11 ₄ ′ uses the paraphrase word string corresponding to the word string as the auxiliary hypothesis f _{4 in association} with the rule for converting the word stored in the paraphrase hypothesis generation rule storage unit 10 ₄ ′ to the paraphrase word. Generate.

Since the input symbol stage number n of m = 5 in the input / output symbol string arrangement storage unit 15 ′ is n = 4, the input symbol selection unit 14 ₄ ′ uses the paraphrase word string output by the hypothesis generation unit 11 ₄ ′ as 5 The data is input to the hypothesis generation unit 11 ₅ ′ at the stage. The hypothesis generation unit 11 ₅ ′ is rephrased in association with a rule for converting the word stored in the paraphrase destination word language model hypothesis generation rule storage unit 10 ₅ ′ into a paraphrase destination language model that provides easy connection of words. the paraphrase word sequence assigned weight to words is generated as an auxiliary hypothesis f _5. The hypothesis narrowing-down unit 13 ₅ ′ associated with the hypothesis generation unit 11 ₅ ′ outputs the extracted auxiliary hypothesis f ₅ to the outside because the output information is True with reference to the input / output symbol string arrangement storage unit 15 ′.

As described above, since the hypothesis generation rule for paraphrasing takes a word string as input, it cannot be cascaded after the hypothesis generation rule for the class language model in the conventional method. Cascade connection is made possible by the functions of the symbol selector 11 'and the input / output symbol string arrangement storage unit 15', and paraphrasing is also possible in the use of a class language model. In addition, _two word strings of the auxiliary hypothesis f ₂ (word before paraphrase) and f ₅ (word after paraphrase) can be output simultaneously as output symbol strings.

In the above description, the hypothesis generation unit 11 ₁ ′ in the first stage reads and processes two hypothesis generation rules, but the number of hypothesis generation units corresponding to the number of hypothesis generation rules is used for speech. A recognition device may be configured.

  In addition, by replacing the hypothesis generation rule for the paraphrase destination word language model with the hypothesis generation rule for the paraphrase destination class language model in the above, the hypothesis generation rule for the paraphrase destination class conversion is inserted immediately before, so that the class model for the paraphrase destination language model is also replaced. Language models can be used.

  At that time, the input of the paraphrase destination class conversion hypothesis generation rule is a word string that is the output of the word utterance dictionary hypothesis generation rule, and the input of the paraphrase destination class language model hypothesis generation rule is the generation of the paraphrase destination class conversion hypothesis generation A class sequence that is the output of the rule.

Class language models are described, for example, in Peter F. Brown et al. “Classbased ngram models of natural language”, Computational Linguistics 18 (4), 467-479, 1992.
A method of describing a word pronunciation dictionary and a language model for speech recognition using weighted finite state transition rules is described in, for example, “Weighted finite-state transducers in speech” by M. Mohri, F. Pereira, M. Riley at the international conference ASR2000. recognition ", Proceeding of ASR2000, pp.97-106,2000.

  As an acoustic model that represents a set of standard acoustic pattern sequences of various speech fixed units (for example, phonemes), for example, a hidden Markov model method that models a set of these acoustic pattern sequences based on probability / statistical theory ( Hidden Markov Model (hereinafter referred to as HMM) is the mainstream. Details of the HMM method are disclosed in, for example, “Recognition of Speech by Stochastic Model” by Seiichi Nakagawa, edited by the Institute of Electronics, Information and Communication Engineers.

  Acoustic patterns include mel cepstrum (referred to as MFCC), delta MFCC, LPC cepstrum, logarithmic power, and the like obtained by analyzing an audio signal every short time (for example, 10 milliseconds).

  Descriptions of individual hypothesis generation units used to simultaneously convert and output sentence style while recognizing speech can be found in, for example, Reference Document (Japanese Patent Laid-Open No. 2004-110673 “Sentence Style Conversion Method, Sentence Style Conversion Device, Sentence style conversion program, storage medium storing sentence style conversion program ").

  When the processing means in the above apparatus is realized by a computer, the processing contents of the 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., magneto-optical recording media, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  This 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.

Claims (8)

M (M ≧ 2) cascaded hypothesis generation units, M hypothesis correction units associated with each of the hypothesis generation units, and M hypothesis narrowing units associated with each of the hypothesis correction units, The input symbol string of the first stage hypothesis generation section provided between the cascaded hypothesis generation sections and the output symbol strings output from the preceding and previous hypothesis generation sections are input to the next stage. An input symbol string selection unit that selects an input symbol string of the hypothesis generation unit, and an output selection of whether or not to output the input symbol string selection information selected by the input symbol string selection unit and the hypotheses extracted by the hypothesis narrowing unit A symbol string conversion device comprising an input / output symbol string arrangement storage unit storing information,
The hypothesis generation unit of the first stage, the hypothesis generation process for the hypothesis f ₁ generated according hypothesis generation rule is a state transition process capable state transitions for an input symbol sequence input,
The input signal selection section refers to the input / output symbol string arrangement storage section described above, and the input symbol string input to the first-stage hypothesis generation section and the output symbol strings output from the preceding and previous hypothesis generation sections Is selected as an input symbol string of the m-th hypothesis generation unit (m is an integer from 1 to M), and the m-th hypothesis generation unit generates the auxiliary hypothesis f _m . An auxiliary hypothesis generation process that repeats the auxiliary hypothesis generation process up to the M-th stage while incrementing m;
The hypothesis correction unit, decrements the m until the weight of the auxiliary hypothesis f _m-1 of the previous stage in the cumulative weight of the auxiliary hypothesis f _m corrected, corrects the cumulative weight of the hypothesis f ₁ the cumulative weight of the auxiliary hypothesis While repeating the hypothesis correction process,
The hypothesis narrowing unit extracts a hypothesis or auxiliary hypothesis in which the corrected cumulative weight exhibits a singular value, and uses the extracted hypothesis or auxiliary hypothesis as an input symbol string of the next hypothesis narrowing unit;
A symbol string conversion method comprising: In the symbol string conversion method according to claim 1,
The above hypothesis narrowing process is
Each hypothesis narrowing unit is a hypothesis narrowing process that also performs processing for selecting whether to output the extracted hypothesis f ₁ or auxiliary hypothesis f _m to the outside with reference to the input / output symbol string arrangement storage unit. A character string conversion method characterized by the above. A voice feature symbol string extraction process for extracting an acoustic feature symbol string from a time series of a short-time acoustic pattern of a voice with an audio signal as an input;
A symbol string representing an acoustic feature converted in the acoustic feature symbol string conversion process is converted into a symbol string corresponding to the audio signal by the symbol string conversion method according to claim 1 or 2,
The first-stage hypothesis generation unit receives as input the symbol string representing the acoustic features extracted by the speech feature symbol string extraction unit, the hypothesis generation rules for the acoustic model having standard features of fixed speech units, and the pronunciation of various words And a word pronunciation dictionary hypothesis generation rule that corresponds to a sequence of fixed speech units to generate a hypothesis that uses a word-based symbol string as an output symbol string. M (M ≧ 2) cascaded hypothesis generation units, M hypothesis correction units associated with each of the hypothesis generation units, and M hypothesis narrowing units associated with each of the hypothesis correction units, The input symbol string of the first stage hypothesis generation section provided between the cascaded hypothesis generation sections and the output symbol strings output from the preceding and previous hypothesis generation sections are input to the next stage. An input symbol string selection unit that selects an input symbol string of the hypothesis generation unit, and an output selection of whether or not to output the input symbol string selection information selected by the input symbol string selection unit and the hypotheses extracted by the hypothesis narrowing unit A symbol string conversion device including an input / output symbol string arrangement storage unit storing information,
The hypothesis generation unit of the first stage, the hypothesis f ₁ is a state transition process capable state transitions for an input symbol sequence input generated in accordance with the hypothesis generation rule,
The input signal selection section refers to the input / output symbol string arrangement storage section described above, and the input symbol string input to the first-stage hypothesis generation section and the output symbol strings output from the preceding and previous hypothesis generation sections Is selected as an input symbol string of the m-th hypothesis generation unit (m is an integer from 1 to M), and the m-th hypothesis generation unit generates the auxiliary hypothesis f _m . The auxiliary hypothesis generation process is repeated up to the M-th stage while incrementing m.
The hypothesis correction unit, decrements the m until the weight of the auxiliary hypothesis f _m-1 of the previous stage in the cumulative weight of the auxiliary hypothesis f _m corrected, corrects the cumulative weight of the hypothesis f ₁ the cumulative weight of the auxiliary hypothesis While repeating,
The hypothesis narrowing unit extracts a hypothesis or auxiliary hypothesis in which the corrected cumulative weight exhibits a singular value, and uses the extracted hypothesis or auxiliary hypothesis as an input symbol string of the hypothesis narrowing unit in the next stage.
A symbol string converter characterized by that. In the symbol string conversion device according to claim 4,
Each of the hypothesis narrowing units also performs processing for selecting whether to output the extracted hypothesis f ₁ or auxiliary hypothesis f _m to the outside with reference to the input / output symbol string arrangement storage unit. Symbol string converter. A voice feature symbol string extraction unit that extracts a voice feature symbol string from a time series of a short-time acoustic pattern of a voice by inputting a voice signal;
The symbol string representing the acoustic feature converted by the acoustic feature symbol string converter is converted into a symbol string corresponding to the voice signal by the symbol string conversion method according to claim 1 or 2,
The first-stage hypothesis generation unit receives as input the symbol string representing the acoustic features extracted by the speech feature symbol string extraction unit, the hypothesis generation rules for the acoustic model having standard features of fixed speech units, and the pronunciation of various words A hypothesis that generates a word-based symbol string as an output symbol string by referring to a word pronunciation dictionary hypothesis generation rule that corresponds to a sequence of fixed speech units.   A program for causing a computer to function as the symbol string conversion device according to claim 4 or the voice recognition device according to claim 6.   A computer-readable recording medium on which any one of the programs according to claim 7 is recorded.

Images