JP6367993B1 - Learning device, noise suppression parameter set switching rule learning device, speech recognition device, learning method, noise suppression parameter set switching rule learning method, speech recognition method, program - Google Patents

Abstract

A learning apparatus capable of setting a noise suppression parameter set based on a sound collection condition is provided. A speech database that stores a correct sentence and sound collection conditions corresponding to a sound file in association with each other and stores them as a first entry, and associates feature amounts, sound collection conditions, noise suppression parameter sets, and accuracy information. A speech recognition result database creating unit for storing the second entry in the speech recognition result database, a search for the speech recognition result database based on the designated sound pickup condition, and a group based on the feature value for the searched second entry It includes a noise suppression parameter set switching rule learning unit that performs grouping according to the grouping criterion and stores the optimum noise suppression parameter set, the grouping criterion, and the sound collection condition in the switching rule database as a third entry in association with each other. [Selection] Figure 1

Description

The present invention relates to a speech recognition service for writing speech contents to a smartphone or the like into characters, and in order to provide the speech recognition service in various environments, a noise suppression parameter set suitable for each environment is determined in advance. learning device used to, noise suppression parameter set changeover rule learning device, a speech recognition apparatus, a learning method, a noise suppression parameter set changeover rule learning process, speech recognition method, and a program.

  For example, Patent Document 1 is a background art in this technical field. The parameter determination device of Patent Literature 1 performs grouping on a speech data set using a noise level for each band as a feature amount, and selects a noise suppression parameter set that maximizes the speech recognition rate for each group. .

Japanese Patent Laying-Open No. 2006-139025 Japanese Patent No. 3309895

Y. Uemura, Y. Takahashi, H. Saruwatari, K. Shikano, K. Kondo, “AUTOMATIC OPTIMIZATION SCHEME OF SPECTRAL SUBTRACTION BASED ON MUSICAL NOISE ASSESSMENT VIA HIGHER-ORDER STATISTICS,” Seattle, 2008. M. Vondrasek, P. Pollak, “Methods for Speech SNR estimation: Evaluation Tool and Analysis of VAD Dependency,” Radioengineering, Vol 14, 14, pp. 6-11, 2005.

  The optimal noise suppression parameter set may vary depending on the sound signal collection conditions. For example, if the terminal type that picks up the sound is different, the microphone element and its configuration are different, and therefore the parameter value optimal for the noise suppression processing may be greatly different. Since the conventional technology does not consider the sound collection condition, there is a possibility that the optimum noise suppression parameter set may not be selected.

  Therefore, an object of the present invention is to provide a learning device that can set a noise suppression parameter set based on sound collection conditions.

  The learning device of the present invention includes a speech recognition result database creation unit and a noise suppression parameter set switching rule learning unit.

  The speech recognition result database creation unit stores a speech file, a correct sentence corresponding to the speech file, and a sound collection condition that is a label that defines a condition at the time of sound collection of the speech file in association with each other, and stores them as a first entry. Using the speech database, features that define the noise characteristics of speech files, sound collection conditions, a noise suppression parameter set that is a set of parameters used for noise suppression, and accuracy that evaluates the accuracy of speech recognition results The process of associating information and storing it as a second entry in the speech recognition result database is executed for each noise suppression parameter set of the noise suppression parameter set group consisting of a plurality of noise suppression parameter sets.

  The noise suppression parameter set switching rule learning unit searches the speech recognition result database based on the specified sound pickup conditions, groups the searched second entries according to the grouping criterion based on the feature amount, As the third entry, the optimum noise suppression parameter set, which is the noise suppression parameter set selected in each group so that the accuracy information satisfies a predetermined condition, the grouping criterion, and the sound collection condition used for the search are associated with each other. Store in the switching rule database.

  According to the learning device of the present invention, it is possible to set a noise suppression parameter set based on sound collection conditions.

1 is a block diagram illustrating a configuration of a voice recognition system according to Embodiment 1. FIG. The block diagram which shows the structure of the speech recognition system of a modification. The block diagram which shows the structure of the speech recognition apparatus of a modification. FIG. 3 is a sequence diagram illustrating the operation of the learning device according to the first embodiment. FIG. 3 is a sequence diagram showing the operation of the noise suppression / voice recognition device of the first embodiment. The block diagram which shows the structure of a speech recognition result database preparation part. The flowchart which shows operation | movement of the speech recognition result database preparation part. The block diagram which shows the structure of a noise suppression parameter set switching rule learning part. The flowchart which shows operation | movement of a noise suppression parameter set switching rule learning part. The block diagram which shows the structure of a noise suppression / speech recognition apparatus. The flowchart which shows operation | movement of a noise suppression / voice recognition apparatus.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

<Voice recognition system 1>
With reference to FIG. 1, the structure of the speech recognition system 1 of Example 1 is demonstrated. As shown in the figure, the speech recognition system 1 of the present embodiment includes a learning device 10, a speech database 11, a speech recognition result database 13, a switching rule database 15, and a noise suppression / speech recognition device 16. The apparatus 10 includes a speech recognition result database creation unit 12 and a noise suppression parameter set switching rule learning unit 14.

  In addition, as shown in FIG. 2, the learning apparatus 10 can be divided into two apparatuses. Specifically, it is an independent device having the same function as the speech recognition result database creating device 12 and the noise suppression parameter set switching rule learning unit 14 which are independent devices having the same function as the speech recognition result database creating unit 12. You may comprise as the speech recognition system 1a containing the certain noise suppression parameter set switching rule learning apparatus 14. FIG.

  Note that the entire system may be configured as one apparatus as shown in FIG. Specifically, a speech database 11, a speech recognition result database creation unit 12, a speech recognition result database 13, a noise suppression parameter set switching rule learning unit 14, a switching rule database 15, and a noise suppression / speech recognition device 16 You may comprise as the speech recognition apparatus 1b containing the noise suppression and the speech recognition part 16 which has the same function. In this case, the three databases may be arranged outside the voice recognition device 1b. In the following, the description will proceed according to the configuration shown in FIG.

  In this embodiment, a speech recognition service used on-line is assumed. A person who provides and operates a voice recognition service is called an operator, and a person who uses the voice recognition service from a terminal such as a smartphone is called a user. The operator is illustrated and referred to as operator a. The user is illustrated and referred to as user b.

  For example, the learning device 10 is assumed to be a device that operates in response to the operation of the operator a. The noise suppression / speech recognition device 16 is assumed to be an online processing device in a speech recognition service that operates when the user b operates. The speech recognition result database creation unit 12 is a processing unit for preparing a speech recognition result database 13 used for noise suppression parameter set switching rule learning. The noise suppression parameter switching rule learning unit 14 determines in advance a rule as to what kind of noise suppression parameter set is assigned to what kind of voice file when the noise suppression / voice recognition device 16 is used. It is a processing part.

<Definition of terms>
The terms used in this example are defined in the following table.

<Operation of Learning Device 10>
Hereinafter, step S12 and step S14 which are operation | movement of the learning apparatus 10 are demonstrated with reference to FIG.

<Operation of Speech Recognition Result Database Creation Unit 12 (S12)>
Prior to this step, the operator a prepares a noise suppression parameter set group in advance and inputs it to the speech recognition result database creation unit 12. In addition to this, the voice database 11 is referred to in this step. In the voice database 11, a voice file, a correct sentence, and a sound collection condition are associated with each other and stored in advance as a first entry. The voice recognition result database creation unit 12 calculates a feature amount from a target voice file. The speech recognition result database creation unit 12 performs noise suppression on the target speech file based on the noise suppression parameter set. For the noise suppression method and the noise suppression parameter set, the same method as in Patent Document 1 can be used. The speech recognition result database creation unit 12 recognizes the speech file after noise suppression. The speech recognition result database creation unit 12 calculates accuracy information by comparing the speech recognition result and the correct sentence.

  The speech recognition result database creation unit 12 associates the feature amount, the sound collection condition, the noise suppression parameter set, and the accuracy information, and stores the processing in the speech recognition result database 13 as a second entry in the noise suppression parameter set group. For each noise suppression parameter set.

  That is, noise suppression is performed using a noise suppression parameter set group including N (N is an integer of 2 or more) noise suppression parameter sets prepared in advance by the operator a. N noise-suppressed audio files are generated for one audio file by noise suppression, and a corresponding speech recognition result is generated. The generated N speech recognition results are compared with the correct sentence, and N pieces of accuracy information are calculated for the N speech recognition results. Accordingly, N second entries in which these are associated are stored in the speech recognition result database 13. The above is step S12. Step S <b> 12 is executed for each audio file in the audio database 11.

The feature amount may be configured to include one or more of the following values, for example.
-Noise power for each band (see Patent Document 2)
-Noise kurtosis (see Non-Patent Document 1)
・ Signal-to-noise ratio (see Non-Patent Document 2)
As for the sound collection conditions, those stored in the voice database 11 in association with the voice files are passed on to the voice recognition result database 13 as they are. As the sound collection conditions, for example, the items and labels shown in the table below can be considered. It is preferable if the terminal name is included as an item of the sound pickup condition.

The accuracy information can be calculated from the following information.
・ Number of characters or words in correct sentences ・ Number of correct characters or words in speech recognition results ・ Number of erroneous characters or words in speech recognition results

<Operation of Noise Suppression Parameter Set Switching Rule Learning Unit 14 (S14)>
Prior to this step, the operator a designates a sound collection condition for learning the switching rule and inputs it to the noise suppression parameter set switching rule learning unit 14. The noise suppression parameter set switching rule learning unit 14 searches the speech recognition result database 13 based on the sound collection conditions specified by the operator a. The noise suppression parameter set switching rule learning unit 14 groups the searched second entries according to a grouping criterion based on the feature amount. The noise suppression parameter set switching rule learning unit 14 selects a noise suppression parameter set in each group so that accuracy information in each group satisfies a predetermined condition, and sets it as an optimal noise suppression parameter set. The noise suppression parameter set switching rule learning unit 14 stores the optimum noise suppression parameter set, the grouping criterion, and the sound collection condition used for the search in association with each other in the switching rule database 15 as a third entry. The above is step S14. This embodiment is characterized in that noise characteristics that are indistinguishable from the feature amount alone are distinguished by a search based on sound collection conditions, and a noise suppression parameter set suitable for each is selected.

A method for specifying the sound pickup condition will be described. The sound collection condition is designated by the operator a, but in order to reduce the work, any one of the following designation methods may be used for one item of the sound collection condition.
・ Specify a single condition.
・ Specify multiple conditions with OR condition.
-No condition is specified (OR condition for all conditions).

  When the sound collection condition includes a plurality of items, an AND condition for all items is designated. If the number of items and the number of conditions for each item are large, the total number of combinations is considered to be very large because the number of sound collecting conditions is a combination thereof. Therefore, the noise suppression parameter set switching rule may be learned for all sound collection conditions, or may be learned only for some sound collection conditions. When learning only for some sound collection conditions, there is a possibility that the sound collection conditions specified when the user b uses the speech recognition service do not exist in the noise suppression parameter switching rule learning database 15. Therefore, it is preferable to specify one of the learned noise suppression parameter set switching rules on the switching rule database 15 as a default rule used in the above case.

  For grouping and selection of the optimal noise suppression parameter set, the same method as in Patent Document 1 can be used. The grouping criterion and the optimum noise suppression parameter set for each group are repeatedly and alternately determined so that the voice recognition accuracy (accuracy information) of the entire transferred data is maximized.

<Operation of Noise Suppression / Voice Recognition Device 16>
Hereinafter, step S16, which is the operation of the noise suppression / voice recognition device 16, will be described with reference to FIG. Prior to this step, it is assumed that the user b designates the voice file to be voice-recognized and the sound collection conditions of the voice file and inputs them to the noise suppression / voice recognition device 16.

  As a method for specifying the sound pickup condition, for example, the user b may manually specify the current sound pickup condition when using the voice recognition service. Further, the terminal used by the user b may specify the current sound pickup conditions through software or the like. As a method of specifying / acquiring the sound collection condition, there are a case where at least one item of the sound collection condition is specified / acquired and a case where none of the items is specified / acquired. In a search for an item that is not designated / acquired for sound pickup condition search, the OR condition of all conditions is hit.

  The noise suppression / speech recognition device 16 searches the switching rule database 15 based on the sound collection conditions designated by the user b. The noise suppression / voice recognition device 16 groups the voice files acquired from the user b according to the grouping criteria of the searched third entry. The noise suppression / voice recognition device 16 performs noise suppression on the voice file acquired from the user b based on the optimum noise suppression parameter set corresponding to the grouped group. The noise suppression / speech recognition device 16 recognizes the speech file after noise suppression and outputs a speech recognition result. The above is step S16.

<Details of Speech Recognition Result Database Creation Unit 12 and Step S12>
The details of the speech recognition result database creation unit 12 and step S12 according to the present embodiment will be described below with reference to FIGS. As shown in FIG. 6, the speech recognition result database creation unit 12 includes a feature amount calculation unit 121, a noise suppression unit 122, a speech recognition unit 123, and a speech recognition accuracy evaluation unit 124. The feature amount calculation unit 121 calculates a feature amount from the target audio file (S121). The noise suppression unit 122 performs noise suppression based on an arbitrary noise suppression parameter set in the noise suppression parameter set group designated by the operator a (S122). The voice recognition unit 123 recognizes the voice file after noise suppression (S123). The speech recognition accuracy evaluation unit 124 calculates accuracy information by comparing the speech recognition result and the correct sentence (S124). The speech recognition result database creation unit 12 specifies the processing for storing the feature amount, the sound collection condition, the noise suppression parameter set, and the accuracy information in the speech recognition result database 13 as the second entry in association with each other. It performs for each noise suppression parameter set of the noise suppression parameter set group. The above is the details of the operation in step S12.

<Details of Noise Suppression Parameter Set Switching Rule Learning Unit 14 and Step S14>
The details of the noise suppression parameter set switching rule learning unit 14 and step S14 of the present embodiment will be described below with reference to FIGS. As shown in FIG. 8, the noise suppression parameter set switching rule learning unit 14 includes a sound collection condition search unit 141, a grouping unit 142, an optimum parameter set selection unit 143, and a convergence determination unit 144. The sound collection condition search unit 141 searches the speech recognition result database 13 based on the sound collection condition designated by the operator a (S141). The grouping unit 142 groups the searched second entries according to a grouping criterion based on the feature amount (S142a). The optimum parameter set selection unit 143 selects the noise suppression parameter set of each group so that the accuracy information of each group is maximized in each group (S143a). The convergence determination unit 144 performs convergence determination on the noise suppression parameter set of each selected group (S144a), and determines that the optimization of the noise suppression parameter set has converged (S144b-Y), the optimal parameter An output command is output to the set selection unit 143 (S144c). On the other hand, when it is determined that the optimization of the noise suppression parameter set has not converged (S144b-N), the convergence determination unit 144 outputs a grouping command to the grouping unit 142 (S144d). The optimum parameter set selection unit 143 that has acquired the output command sets the latest noise suppression parameter set as the optimum noise suppression parameter set, and enters the switching rule database 15 as a third entry together with the grouping criteria and the sound collection conditions used for the search. Store (S143b). On the other hand, the grouping unit 142 that has acquired the grouping command changes the grouping criteria and executes grouping again (142b). The above is the details of the operation in step S14.

<Details of Noise Suppression / Voice Recognition Device 16 and Step S16>
Details of the noise suppression / speech recognition device 16 and step S16 of this embodiment will be described below with reference to FIGS. As shown in FIG. 10, the noise suppression / speech recognition device 16 includes a sound collection condition search unit 161, a feature amount calculation unit 162, a noise suppression parameter set derivation unit 163, a noise suppression unit 164, and a speech recognition unit 165. including. The sound collection condition search unit 161 searches the switching rule database 15 based on the sound collection condition designated by the user b (S161). The feature amount calculation unit 162 calculates a feature amount from the audio file acquired from the user b (S162). The noise suppression parameter set deriving unit 163 groups the feature amounts of the audio file acquired from the user b in accordance with the grouping criterion of the retrieved third entry, and derives an optimum noise suppression parameter set corresponding to this ( S163). The noise suppression unit 164 performs noise suppression on the voice file acquired from the user b based on the derived optimal noise suppression parameter set (S164). The voice recognition unit 165 recognizes the voice file after noise suppression and outputs a voice recognition result (S165). The above is the details of the operation in step S16.

<Effects produced by the speech recognition system 1 of this embodiment>
According to the speech recognition system 1 of the present embodiment, it is possible to automatically select a noise suppression parameter set suitable for each sound collection condition even when the sound collection conditions are various.

<Supplementary note>
The apparatus of the present invention includes, for example, a single hardware entity as an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity. Can be connected to a communication unit, a CPU (Central Processing Unit, may include a cache memory or a register), a RAM or ROM that is a memory, an external storage device that is a hard disk, and an input unit, an output unit, or a communication unit thereof , A CPU, a RAM, a ROM, and a bus connected so that data can be exchanged between the external storage devices. If necessary, the hardware entity may be provided with a device (drive) that can read and write a recording medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device of the hardware entity stores a program necessary for realizing the above functions and data necessary for processing the program (not limited to the external storage device, for example, reading a program) It may be stored in a ROM that is a dedicated storage device). Data obtained by the processing of these programs is appropriately stored in a RAM or an external storage device.

  In the hardware entity, each program stored in an external storage device (or ROM or the like) and data necessary for processing each program are read into a memory as necessary, and are interpreted and executed by a CPU as appropriate. . As a result, the CPU realizes a predetermined function (respective component requirements expressed as the above-described unit, unit, etc.).

  The present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the above embodiment may be executed not only in time series according to the order of description but also in parallel or individually as required by the processing capability of the apparatus that executes the processing. .

  As described above, when the processing functions in the hardware entity (the apparatus of the present invention) described in the above embodiments are realized by a computer, the processing contents of the functions that the hardware entity should have are described by a program. Then, by executing this program on a computer, the processing functions in the hardware entity are 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 medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  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. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

In this embodiment, a hardware entity is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

Claims (7)

Using a speech database that stores a sound file, a correct sentence corresponding to the sound file, and a sound collection condition that is a label that defines a condition at the time of sound collection of the sound file as a first entry in association with each other, Corresponding the feature quantity that defines the noise characteristics of the audio file, the sound collection condition, the noise suppression parameter set that is a set of parameters used for noise suppression, and accuracy information that is a value for evaluating the accuracy of the speech recognition result A speech recognition result database creation unit that executes processing for storing the second entry in the speech recognition result database for each noise suppression parameter set of the noise suppression parameter set group including the plurality of noise suppression parameter sets; ,
The speech recognition result database is searched based on the specified sound pickup condition, the searched second entries are grouped according to a grouping criterion based on the feature amount, and the accuracy information in each group is a predetermined value. A switching rule as a third entry by associating the optimum noise suppression parameter set, which is the noise suppression parameter set selected in each group so as to satisfy the condition, the grouping criterion, and the sound collection condition used for the search A learning apparatus including a noise suppression parameter set switching rule learning unit stored in a database. A feature amount that defines noise characteristics of an audio file; a sound collection condition that is a label that defines conditions at the time of sound collection of the audio file; a noise suppression parameter set that is a set of parameters used for noise suppression; and speech recognition A speech recognition result database in which an entry associated with accuracy information, which is a value for evaluating the accuracy of the result, is stored for each noise suppression parameter set of the noise suppression parameter set group including the plurality of noise suppression parameter sets. A noise suppression parameter set switching rule learning device using
A sound collection condition search unit for searching the speech recognition result database based on the specified sound collection condition;
A grouping unit that groups the retrieved entries according to a grouping criterion based on the feature amount;
Corresponding to the optimum noise suppression parameter set, which is the noise suppression parameter set selected in each group so that the accuracy information in each group satisfies a predetermined condition, the grouping criterion, and the sound collection condition used for the search A noise suppression parameter set switching rule learning device including an optimum parameter set selection unit that stores an entry as an entry in a switching rule database. An audio database that stores a first entry in association with an audio file, a correct sentence corresponding to the audio file, and a sound collection condition that is a label that defines a condition at the time of sound collection of the audio file;
Corresponds to features that define noise characteristics of the voice file, the sound collection conditions, a noise suppression parameter set that is a set of parameters used for noise suppression, and accuracy information that is a value that evaluates the accuracy of the speech recognition result A speech recognition result database creation unit that executes the process as the second entry for each noise suppression parameter set of the noise suppression parameter set group including the plurality of noise suppression parameter sets;
A speech recognition result database for storing the second entry;
The speech recognition result database is searched based on the sound collection condition specified by the operator, the searched second entries are grouped according to a grouping criterion based on the feature amount, and the accuracy information in each group Is a third entry by associating the optimum noise suppression parameter set, which is the noise suppression parameter set selected in each group so as to satisfy a predetermined condition, the grouping criterion, and the sound collection condition used for the search. A noise suppression parameter set switching rule learning unit, and
A switching rule database for storing the third entry;
The switching rule database is searched based on the sound pickup conditions specified by the user, and the audio files obtained from the user are grouped according to the grouping criteria of the searched third entry and grouped. A noise suppression / voice recognition unit that suppresses noise of a voice file acquired from the user based on the optimum noise suppression parameter set corresponding to a group, recognizes the voice file after noise suppression, and outputs a voice recognition result; Including speech recognition device. A learning method executed by a learning device,
Using a speech database that stores a sound file, a correct sentence corresponding to the sound file, and a sound collection condition that is a label that defines a condition at the time of sound collection of the sound file as a first entry in association with each other, Corresponding the feature quantity that defines the noise characteristics of the audio file, the sound collection condition, the noise suppression parameter set that is a set of parameters used for noise suppression, and accuracy information that is a value for evaluating the accuracy of the speech recognition result A process of storing in the speech recognition result database as a second entry for each noise suppression parameter set of the noise suppression parameter set group consisting of a plurality of the noise suppression parameter sets;
The speech recognition result database is searched based on the specified sound pickup condition, the searched second entries are grouped according to a grouping criterion based on the feature amount, and the accuracy information in each group is a predetermined value. A switching rule as a third entry by associating the optimum noise suppression parameter set, which is the noise suppression parameter set selected in each group so as to satisfy the condition, the grouping criterion, and the sound collection condition used for the search A learning method comprising the step of storing in a database. A feature amount that defines noise characteristics of an audio file; a sound collection condition that is a label that defines conditions at the time of sound collection of the audio file; a noise suppression parameter set that is a set of parameters used for noise suppression; and speech recognition A speech recognition result database in which an entry associated with accuracy information, which is a value for evaluating the accuracy of the result, is stored for each noise suppression parameter set of the noise suppression parameter set group including the plurality of noise suppression parameter sets. A noise suppression parameter set switching rule learning method using
Searching the speech recognition result database based on the designated sound pickup conditions;
Grouping the retrieved entries according to a grouping criterion based on the features;
Corresponding to the optimum noise suppression parameter set, which is the noise suppression parameter set selected in each group so that the accuracy information in each group satisfies a predetermined condition, the grouping criterion, and the sound collection condition used for the search And storing it as an entry in the switching rule database.
A noise suppression parameter set switching rule learning method executed by a noise suppression parameter set switching rule learning device. A speech recognition method executed by a speech recognition apparatus,
Using a speech database that stores a sound file, a correct sentence corresponding to the sound file, and a sound collection condition that is a label that defines a condition at the time of sound collection of the sound file as a first entry in association with each other, Corresponding the feature quantity that defines the noise characteristics of the audio file, the sound collection condition, the noise suppression parameter set that is a set of parameters used for noise suppression, and accuracy information that is a value for evaluating the accuracy of the speech recognition result A process of storing in the speech recognition result database as a second entry for each noise suppression parameter set of the noise suppression parameter set group consisting of a plurality of the noise suppression parameter sets;
The speech recognition result database is searched based on the sound collection condition specified by the operator, the searched second entries are grouped according to a grouping criterion based on the feature amount, and the accuracy information in each group Is a third entry by associating the optimum noise suppression parameter set, which is the noise suppression parameter set selected in each group so as to satisfy a predetermined condition, the grouping criterion, and the sound collection condition used for the search. Storing in the switching rule database as
The switching rule database is searched based on the sound pickup conditions specified by the user, and the audio files obtained from the user are grouped according to the grouping criteria of the searched third entry and grouped. A speech recognition method comprising the steps of: noise-suppressing a speech file acquired from the user based on the optimum noise suppression parameter set corresponding to a group; speech recognition of the speech file after noise suppression; and outputting a speech recognition result.   A program for causing a computer to function as the apparatus according to any one of claims 1 to 3.

Images