JP2004317845A - Model data generation device, model data generation method, and method therefor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To generate a sound model from a currently existent sound model DB through statistical processing in specified environment. <P>SOLUTION: Each sound model DB is stored with phoneme HMM data. A sound model generating device 1 receives phoneme HMM data, needed to generate test data "Hello" for decision from respective sound model DBs and generates cancatenated phoneme data representing an array of "Hello" from the received phoneme data. A database such that the voice feature quantity of the concatenated phoneme data is similar to the voice feature quantity of the test data for decision is selected. Composite phoneme concatenated HMM data are generated by combining identical phoneme HMM data stored in the selected database at a certain ratio and also concatenated to generate composite phoneme concatenated HMM data, and a ratio at which an evaluation value becomes high is found. Phoneme HMM data as to all phonemes are received from the selected database and put together at the ratio to generate composite phoneme HMM data, which are regarded as a sound model in the specified environment. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a process for generating model data such as an acoustic model and a language model.
[0002]
Problems to be solved by the prior art and the invention
Conventionally, when generating an acoustic model, a predetermined sentence is spoken for each supposed speaker or environment, and the voice data is generated by statistical processing. For example, as an acoustic model for voice recognition in a car navigation system, under various driving conditions in a car, 100 men in their 20s to 50s and 120 women in their 20s to 50s are set in advance. A predetermined sentence is spoken, a feature amount is extracted from the voice data, and a statistical process is performed to generate an acoustic model. By using such a statistical method, the speech recognition rate of an unspecified speaker can be increased.
[0003]
However, such a method requires a large number of data collections for performing statistical processing in each environment in which speech recognition is to be performed.
[0004]
The above-mentioned problem becomes a problem not only for generation of an acoustic model but also for a language model which is a statistical model.
[0005]
An object of the present invention is to solve the above-described problem and to provide an acoustic model generation device or an acoustic model generation method that generates an acoustic model subjected to the statistical processing in a certain environment using already existing audio data. I do.
[0006]
It is another object of the present invention to provide a language model generation device or method for generating a language model that has been subjected to the statistical processing in a certain environment using existing language data.
[0007]
It is another object of the present invention to provide a model data generating apparatus or method for generating model data that has been subjected to the statistical processing under a certain environment using model data that already exists.
[0008]
Means for Solving the Problems and Effects of the Invention
1) An acoustic model generation method according to the present invention is an acoustic model generation method for generating phoneme HMM data under a specific environment. 1) A plurality of candidate data storage devices for storing phoneme HMM data are connected to a network. When the test data for judgment under the specific environment composed of the sound data and the notation character string is given, the phoneme corresponding to the notation character string in the stored phoneme HMM data is stored from each candidate data storage device. Receiving the HMM data, determining the similarity between the phoneme-connected HMM data generated by connecting the phoneme HMM data, and the sound data of the test data for determination, and 3) based on the obtained similarity, A plurality of selected candidate data storage devices are specified from the candidate data storage devices; and 4) sounds stored in each of the selected candidate data storage devices. When two or more HMM data are combined for each phoneme, a synthesis ratio with a higher similarity is obtained for the phoneme-connected HMM data corresponding to the written character string of the test data for determination. The synthesized phoneme HMM data obtained by synthesizing the phoneme HMM data stored in the completed candidate data storage device is generated as an acoustic model under the specific environment.
[0009]
In this way, the similarity between the phoneme-connected HMM data generated by connecting the received phoneme HMM data and the sound data of the test data for determination is determined, and the selected candidate data storage device is determined. When the phoneme HMMs extracted from the plurality of selected candidate data storage devices are synthesized at a certain synthesis ratio, a synthesis ratio with which the similarity with the judgment test data is higher is obtained, and a new synthesis ratio is calculated using the ratio. A simple acoustic model is generated. Therefore, by collecting the acoustic data under the specific environment to the extent necessary for the determination, it is possible to obtain a statistically processed acoustic model using the phoneme HMM data already stored in the candidate data storage device.
[0010]
2) An acoustic model generation method according to the present invention is an acoustic model generation method for generating a statistically processed acoustic model under a specific environment, and includes the following steps. 1) a step of specifying a selected candidate data storage device from a plurality of data storage devices connected to a network, the test data for determination under the specific environment being composed of acoustic data and its notation character string; A selection step of specifying a selected candidate data storage device based on the degree of conformity with acoustic model unit data corresponding to a written character string of the determination test data stored in the data storage device; 2) the selected candidate data storage device A combination in which the degree of similarity with the combined acoustic model unit data corresponding to the written character string of the test data for judgment is higher for the combined acoustic model unit data obtained by combining two or more acoustic model unit data stored for each unit data. A combination ratio determining step of determining a ratio; and 3) all sound modes from the selected candidate data storage device. Receive Le unit data generating step of generating a composite acoustic model unit data is synthesized with a combination ratio with the determined as statistical processed acoustic model under the specific environment.
[0011]
In this way, the sound model unit concatenated data generated by concatenating the received sound model unit data and the similarity between the sound data of the test data for judgment are determined to determine the selected candidate data storage device, Further, when the acoustic model unit data extracted from the plurality of selected candidate data storage devices is combined at a certain combination ratio, a combination ratio that gives a higher similarity with the test data for determination is obtained, and the ratio is calculated. A new acoustic model is generated using this. Therefore, a statistically processed acoustic model can be obtained by collecting acoustic data in the specific environment to an extent necessary for the determination.
[0012]
3) In the acoustic model generation method according to the present invention, the data storage device stores basic data for generating acoustic model unit data, and generates the acoustic model unit data generation request when given. Remember. Therefore, if the acoustic model unit data has not been converted into data, only the requested acoustic model unit data can be generated.
[0013]
4) In the acoustic model generation method according to the present invention, the acoustic model unit data includes at least an average value and a variance value of a feature vector for the phoneme, and the combination ratio determination step corresponds to the written character string. The average value and the variance value are synthesized for each feature vector of the phoneme to be processed. Therefore, statistically processed data using the average value and the variance value of the feature vector can be obtained.
[0014]
5) An acoustic model generation device according to the present invention is an acoustic model generation device that generates a statistically processed acoustic model under a specific environment, and 1) a selected candidate from a plurality of candidate data storage devices connected to a network. Means for specifying a data storage device, the acoustic model and the acoustic model unit data stored in each of the candidate data storage device, the degree of conformity of the test data for determination under the specific environment composed of the sound data and its notation character string Selected candidate determining means for specifying a data storage device to be selected based on the degree of suitability when the calculation is performed by using the 2nd combination of acoustic model unit data stored in each candidate data storage device for each unit data In this case, the sound data corresponding to the written character string of the test data for determination is stored in the selected candidate data storage device. Combination ratio determining means for determining the ratio of combinations of selected candidate data storage devices that are higher than the fitness calculated by the stored acoustic model unit data. 3) All acoustic model unit data from the selected candidate data storage device. And acoustic model generation means for generating synthetic acoustic model unit data synthesized at the determined combination ratio as a statistically processed acoustic model under the specific environment.
[0015]
As described above, when the acoustic model unit data extracted from each candidate data storage device is synthesized at a certain synthesis ratio, the ratio of the combination of the selected candidate data storage devices that has a higher degree of suitability with the test data for judgment is higher. Is determined, and a new acoustic model is generated using the ratio. By collecting acoustic data under the specific environment to an extent necessary for the determination, a statistically processed acoustic model can be obtained. .
[0016]
6) An acoustic model generating apparatus according to the present invention is an acoustic model generating apparatus that generates acoustic model unit data under a specific environment, and 1) determination under the specific environment composed of acoustic data and its notation character string. Model data storing means for storing test data for use, and acoustic model unit data obtaining means for obtaining, from a plurality of selected candidate data storage devices, acoustic model unit data corresponding to the written character string of the test data for determination. 3) With respect to the acoustic model unit data acquired from the plurality of selected candidate data storage devices, the acoustic model unit data is synthesized at a predetermined synthesis ratio in the same acoustic model unit to generate synthesized acoustic model unit data. For the connected acoustic model unit data generated by the connection, the degree of conformity with the acoustic data of the test data for judgment is determined. Means for determining a synthesis rate at which the degree of adaptation is higher; 4) receiving all acoustic model unit data from the selected candidate data storage device based on the synthesis rate, and A generating unit configured to generate the synthesized acoustic model unit data as an acoustic model under the specific environment.
[0017]
As described above, when the acoustic model unit data extracted from the plurality of selected candidate data storage devices is combined at a certain combination ratio, a combination ratio at which the similarity with the test data for determination becomes higher is calculated. A new acoustic model is generated using the ratio. Therefore, a statistically processed acoustic model can be obtained by collecting acoustic data in the specific environment to an extent necessary for the determination.
[0018]
7) An acoustic model generation device according to the present invention is generated by connecting acoustic model unit data for forming a notation character string of the test data for determination from a plurality of candidate data storage devices connected to a network. The selected candidate data storage device is specified based on the degree of matching with the connected acoustic model unit data. Therefore, the selected candidate data storage device can be used according to the degree of conformity with the test data for determination.
[0019]
8) In the acoustic model generation device according to the present invention, the attribute data indicating the attribute is added to the judgment test data, and the acoustic model stored in the plurality of candidate data storage devices connected to the network. The selected candidate data storage device is specified based on the degree of matching with the attribute data indicating the attribute. Therefore, it is possible to quickly specify the selected candidate data storage device.
[0020]
9) In the acoustic model generation device according to the present invention, acoustic model unit data for forming a notation character string of the test data for determination is received from the plurality of candidate data storage devices, and these are coupled and coupled. Generate acoustic model unit data. Therefore, the candidate data storage device can determine whether or not to select only by outputting a part of the acoustic model unit data necessary for the determination.
[0021]
10) In the acoustic model generation device according to the present invention, the test data for determination is provided to the plurality of candidate data storage devices, and the evaluation is performed using the acoustic model unit data stored in each candidate data storage device. Upon receiving the evaluation value, the selected candidate data storage device is specified. Therefore, at the stage of determining whether or not the candidate data storage device is to be selected, the candidate data storage device can determine this without outputting acoustic model unit data.
[0022]
11) A language model generation device according to the present invention is a language model generation device that generates a statistically processed language model under a specific environment, and 1) a determination test for storing the determination test data under the specific environment. Data storage means, 2) language model unit data acquisition means for acquiring language model unit data included in the test data for judgment from a plurality of selected candidate data storage devices, and 3) acquisition from the selected candidate data storage devices. For the language model unit data, the same language model unit data is synthesized at a certain synthesis ratio to generate synthesized language model unit data, the frequency of appearance in the test data for judgment is calculated, and the synthesis ratio at which the fitness is higher is determined. Means for determining the combination ratio to be used; 4) all languages from the selected candidate data storage device based on the combination ratio; Receive Dell unit data to generate a composite language model unit data is combined by the combining ratio with the determined as the language model under the specific environment.
[0023]
As described above, when the language model unit data extracted from the plurality of selected candidate data storage devices is synthesized at a certain synthesis ratio, a synthesis ratio at which the similarity with the test data for determination becomes higher is calculated. Generate a new language model using the ratio. Therefore, a statistically processed language model can be obtained by collecting language data under the specific environment to an extent necessary for the determination.
[0024]
12) A model data generation device according to the present invention is a model data generation device that generates statistically processed model data under a specific environment, and 1) a determination test for storing the determination test data under the specific environment. Data storage means, 2) model unit data acquisition means for acquiring model unit data included in the test data for determination from the selected candidate data storage device, and 3) model unit data acquired from the selected candidate data storage device. A synthesis ratio determining unit that generates the synthesized model unit data by synthesizing the same model unit data at a certain synthesis ratio, calculates a degree of conformity with the test data for determination, and determines a combination ratio at which the degree of conformity is higher; 4) receiving all model unit data from the selected candidate data storage device based on the combining ratio; Generating a composite model unit data is combined by the combining proportions were serial determined as a statistical processed model data under the specific environment.
[0025]
As described above, when the model unit data extracted from the plurality of selected candidate data storage devices is combined at a certain combination ratio, a combination ratio at which the similarity with the test data for determination becomes higher is determined. To generate a new statistically processed model. Therefore, statistically processed model data can be obtained by collecting data under the specific environment to the extent necessary for the determination.
[0026]
13) A language model generation device according to the present invention is a language model generation device that generates a statistically processed language model under a specific environment, and 1) a determination test for storing the determination test data under the specific environment. Data storage means, 2) language model unit data acquisition means for acquiring language model unit data included in the test data for judgment from the selected candidate data storage device, 3) language model for acquiring the acquired language model unit data As the unit data, an appearance frequency for the same language model unit data is obtained from the selected candidate data storage device, and for each language model unit data, these are combined at a certain combination ratio to generate the appearance of the combined language model unit data. The frequency is calculated, and the degree of conformity with the frequency of appearance in the test data for judgment is calculated. A combining ratio determining means for determining a higher combining ratio; 4) a combining language model obtained by receiving all language model unit data from the selected candidate data storage device based on the combining ratio and combining the unit data at the determined combining ratio. The unit data is generated as a language model under the specific environment.
[0027]
As described above, when the language model unit data extracted from the plurality of selected candidate data storage devices is combined at a certain combination ratio, a combination ratio that provides a higher degree of conformity with the test data for determination is obtained. Generate a new language model using the ratio. Therefore, a statistically processed language model can be obtained by collecting language data under the specific environment to an extent necessary for the determination.
[0028]
14) A program according to the present invention is an acoustic model generation program for generating statistically processed model data under a specific environment, and is a program for causing a computer to execute the following processing. 1) For the test data for judgment under the specific environment composed of the sound data and the notation character string, acoustic model unit data corresponding to the notation character string of the judgment test data from a plurality of selected candidate data storage devices. 2) With respect to the acoustic model unit data acquired from the plurality of selected candidate data storage devices, the acoustic model unit data is combined at the same acoustic model unit at a certain combination ratio to generate synthetic acoustic model unit data. For the connected acoustic model unit data generated by concatenating the test data, the degree of conformity with the acoustic data of the test data for judgment is calculated, and a synthesis ratio at which the degree of conformity is higher is determined. 3) Based on the synthesis ratio, Synthesized sound obtained by receiving all sound model unit data from the selected candidate data storage device and synthesizing them at the determined synthesis ratio Generating a Dell unit data as the acoustic model under the specific environment.
[0029]
As described above, when the acoustic model unit data extracted from the plurality of selected candidate data storage devices is combined at a certain combination ratio, a combination ratio that provides a higher degree of matching with the test data for determination is determined. A new acoustic model is generated using the ratio. Therefore, a statistically processed acoustic model can be obtained by collecting acoustic data in the specific environment to an extent necessary for the determination.
[0030]
15) A program according to the present invention is a model data generation program for generating statistically processed model data under a specific environment, and is a program for causing a computer to execute the following processing: 1) Selected candidate data storage device From the stored model unit data included in the judgment test data under the specific environment, and 2) combining the same model unit data at a certain combination ratio with respect to the model unit data acquired from the selected candidate data storage device. To generate the combined model unit data, calculate the degree of conformity with the test data for judgment, determine the combination ratio at which the degree of conformity is higher, and 3) based on the combination ratio, the selected candidate data storage device. And the combined model unit data obtained by combining with the determined combining ratio. The generated as statistical processed model data under the specific environment.
[0031]
As described above, when the language model unit data extracted from the plurality of selected candidate data storage devices is synthesized at a certain synthesis ratio, a synthesis ratio at which the similarity with the test data for determination becomes higher is calculated. Generate a new language model using the ratio. Therefore, a statistically processed language model can be obtained by collecting language data under the specific environment to an extent necessary for the determination.
[0032]
16) A program according to the present invention is an acoustic model generation program for generating statistically processed model data under a specific environment, and is a program for causing a computer to execute the following processing. Acquiring acoustic model unit data corresponding to the notation character string of the judgment test data from a plurality of candidate data storage devices for the judgment test data under the specific environment composed of the notation character string, For the acoustic model unit data acquired from the candidate data storage device, the combined acoustic model unit data is synthesized at a certain synthesis ratio in the same acoustic model unit to generate synthetic acoustic model unit data, and the combined acoustic model unit data is generated by linking the combined acoustic model unit data. For the unit data, the degree of conformity with the acoustic data of the test data for judgment is calculated, and the degree of conformity is calculated. And 3) receiving all acoustic model unit data from the candidate data storage device based on the composite ratio and combining the combined acoustic model unit data with the determined composite ratio based on the specific environment. Generated as an acoustic model below. As described above, when the acoustic model unit data extracted from the plurality of candidate data storage devices is synthesized at a certain synthesis ratio, a synthesis ratio that provides a higher degree of conformity with the test data for determination is obtained, and the ratio is calculated. To generate a new acoustic model. Therefore, a statistically processed acoustic model can be obtained by collecting acoustic data in the specific environment to an extent necessary for the determination.
[0033]
17) An acoustic model generation device according to the present invention is an acoustic model generation device that generates acoustic model unit data under a specific environment, and 1) for determination under the specific environment, which is composed of acoustic data and a character string thereof. Determination test data storage means for storing test data; 2) acoustic model unit data acquisition means for acquiring, from a plurality of selected candidate data storage devices, acoustic model unit data corresponding to the written character string of the determination test data; 3) With respect to the acoustic model unit data acquired from the plurality of selected candidate data storage devices, the acoustic model unit data is combined at a predetermined combination ratio in the same acoustic model unit to generate combined acoustic model unit data, and these are combined. For the connected acoustic model unit data generated as described above, the degree of conformity with the acoustic data of the test data for judgment is determined. Means for determining a synthesis rate at which the degree of adaptation is higher; 4) receiving all acoustic model unit data from the selected candidate data storage device based on the synthesis rate, and Generating means for generating the synthesized acoustic model unit data as an acoustic model under the specific environment; 5) generated acoustic model storing means for storing the generated acoustic model; and 6) newly provided test data for judgment. , The degree of matching with the generated acoustic model stored in the generated acoustic model storage means is calculated, and if the degree exceeds a predetermined threshold value, the generated acoustic model is output as data of an acoustic model under the specific environment. An acoustic model output unit is provided. Therefore, the already generated acoustic model data can be directly output as acoustic model data under the specific environment.
[0034]
18) An acoustic model generating method according to the present invention is an acoustic model generating method for generating acoustic model unit data under a specific environment, and 1) determination under the specific environment composed of acoustic data and a character string thereof. Given test data, predetermined parameter data or parameter data extracted from acoustic data stored in each candidate data storage device is statistically generated to generate acoustic model unit data corresponding to the written character string. Obtaining the processed statistical data, determining the similarity between the connected acoustic model unit data generated by connecting the acoustic model unit data and the acoustic data of the test data for determination, and storing the selected candidate data. Specifying a plurality of devices; and 2) storing selected candidate data from the candidate data storage device based on the obtained similarity. 3) When two or more acoustic model unit data stored in each of the selected candidate data storage devices are combined for each phoneme, a connected acoustic model unit corresponding to the written character string of the test data for determination. For the data, a combination ratio with a higher similarity is obtained. 4) The predetermined parameter data or parameter data extracted from the acoustic data stored in each of the selected candidate data storage devices at this combination ratio is subjected to statistical processing. The obtained statistical data is obtained, the acoustic model unit data is generated, and synthetic acoustic model unit data obtained by synthesizing the acoustic model unit data is generated and output as an acoustic model under the specific environment.
[0035]
As described above, in the determination step as to whether or not to be a synthesis target, data for determining the degree of coincidence of some data is received. As a result, the acoustic model unit data is received only from the candidate data storage device that actually generates the acoustic model. As a result, billing according to use becomes possible. Also, the candidate data storage device is not the acoustic data but the predetermined parameter data extracted from the acoustic data or the statistical data obtained by performing a statistical process on the parameter data. As a result, privacy can be secured, and the problem of unintended use of data stored in the data storage device can be prevented.
[0036]
19) A language model generating method according to the present invention is a language model generating method for generating a statistically processed language model under a specific environment. 1) When the judgment test data under the specific environment is given, Obtaining predetermined parameter data or statistical data obtained by statistically processing parameter data for generating language model unit data included in the determination test data from the candidate data storage device; and 2) generating language model unit data. And determining the similarity of the appearance frequency of the language model unit data present in the test data for determination to specify a plurality of selected candidate data storage devices, and 3) data extracted from each of the selected candidate data storage devices. When two or more language model unit data generated from are combined for each language model unit data, A synthesis ratio at which the degree of similarity with the language model unit data existing in the test data is higher is determined. 4) Predetermined parameter data for all data stored in each selected candidate data storage device or these parameter data. Is received and statistical data obtained by performing statistical processing is generated, and synthesized language model unit data obtained by synthesizing these at the synthesis ratio is generated as a language model under the specific environment.
[0037]
As described above, at the stage of determining whether or not to be a synthesis target, data for determining the degree of coincidence of some data is received, and when it is determined to be a synthesis target, all language model unit data is read from the candidate data storage device. Upon receipt, the language model unit data is received only from the candidate data storage device that actually generates the language model. As a result, billing according to use becomes possible. The candidate data storage device is not language data but rather predetermined parameter data extracted from the language data or statistical data obtained by performing statistical processing on the parameter data. As a result, privacy can be secured, and the problem of unintended use of data stored in the data storage device can be prevented.
[0038]
20) A method for generating statistically processed model data according to the present invention is a model data generating method for generating statistically processed model data under a specific environment, wherein 1) the test data for judgment under the specific environment is When given, only a part of data necessary for calculating the degree of coincidence with the test data for judgment is determined from each candidate data storage device in order to determine whether or not the data is to be synthesized in a predetermined parameter format. Or the data in the parameter format as statistically processed data, calculates the degree of coincidence with the test data for determination, specifies a plurality of selected candidate data storage devices, and 2) selects each of the selected candidate data. When two or more model unit data generated from data extracted from the storage device are combined for each model unit data, The synthesis ratio at which the similarity of the strike data to the model unit data is higher is determined. 3) The predetermined parameter data of all the data stored in each of the selected candidate data storage devices or these parameter data are statistically calculated. The processing unit receives the processed statistical data, and generates synthetic model unit data obtained by synthesizing the statistical data at the synthesizing ratio as model data under the specific environment.
[0039]
As described above, in the step of determining whether or not to be a synthesis target, data for determining the degree of matching of some data is received, and when it is determined to be a synthesis target, all model unit data is received from the candidate data storage device. Thus, the model unit data is received only from the candidate data storage device that actually generates the model data. As a result, billing according to use becomes possible. From the candidate data storage device, not the raw data that is the basis of the model data, but the predetermined parameter data extracted from the raw data or the statistical data obtained by performing statistical processing on the parameter data. As a result, privacy can be secured, and the problem of unintended use of data stored in the data storage device can be prevented.
[0040]
21) An acoustic model output device according to the present invention is an acoustic model output device for generating acoustic model unit data under a specific environment. 1) Acquisition of the plurality of acquired acoustic model unit data in the same acoustic model unit The model unit data is synthesized at a certain ratio to generate synthetic acoustic model unit data, and these are connected in the order of the notation character string to generate connected acoustic model unit data. Means for calculating a degree of conformity with test data; 2) rate determining means for changing the ratio based on a predetermined ratio determining rule to determine a more suitable ratio; Receiving all the acoustic model unit data stored in the selected candidate data storage device or basic data for generating the same. Te, and generates a synthesized acoustic model unit data is combined by the combining ratio with the determined, and an output means for outputting the acoustic model under the specific environment.
[0041]
As described above, when the acoustic model unit data extracted from the plurality of selected candidate data storage devices is combined at a certain combination ratio, a combination ratio at which the similarity with the test data for determination becomes higher is calculated. A new acoustic model is generated using the ratio. Therefore, a statistically processed acoustic model can be obtained by collecting acoustic data in the specific environment to an extent necessary for the determination.
[0042]
Hereinafter, the meaning of the terms used in the present specification will be described.
[0043]
The “acoustic model unit data” corresponds to phoneme HMM data in the embodiment, but includes data in subword units. The subword unit includes, for example, data expressed in syllable units, phoneme pairs, semisyllables, and the like. Also, in addition to the phoneme HMM data, unit data that can generate acoustic data by combining individual data is included.
[0044]
“Synthesized acoustic model unit data” refers to acoustic model unit data obtained by synthesizing one or more acoustic model unit data for each unit. Here, the combination includes the case where the combination ratio is 1: 0. That is, if the combination ratio of two or more databases is 1: 0 in the calculation of the combination ratio, the model unit data itself extracted from the database corresponding to “1” becomes the combined model unit data. The same applies to the language model.
[0045]
The “phoneme HMM data” described in the embodiment is a case where the feature amount of each phoneme is converted into statistical data in the triphone format, but includes a statistical representation of the feature amount of the phoneme. .
[0046]
The “test data for judgment” is typical data collected under a target environment, and refers to typical data that does not reach statistical data.
[0047]
The “phoneme linked HMM data” refers to data obtained by linking phoneme HMM data.
[0048]
The “selected candidate data storage device” refers to a candidate data storage device to be subjected to a calculation process of a coefficient for weighting.
[0049]
“Combining two or more phoneme HMM data for each phoneme” means to synthesize statistical data on phoneme HMM data for the same phoneme from different candidate databases.
[0050]
"Synthesized phoneme HMM data" refers to phoneme HMM data obtained by combining two or more phoneme HMM data for each phoneme.
[0051]
“Connected acoustic model unit data” refers to aggregate data of acoustic model unit data generated by connecting acoustic model unit data.
[0052]
“Basic data for generating acoustic model unit data” refers to raw speech data as well as data that includes feature amounts extracted from them and has not been subjected to statistical processing.
[0053]
The “attribute data” refers to, for example, 50 males or attribute data relating to acoustic model data or its basic data stored in the database.
[0054]
“Acquiring acoustic model unit data” means not only generating and receiving acoustic model unit data from the data stored in the data storage device, but also receiving the data stored in the data storage device. Based on this, the case where the acoustic model unit data is generated by itself is also included. In the latter case, if the data is audio data stored in the data storage device, it can be received as a value of a predetermined parameter for the data or data obtained by statistically processing these parameters.
[0055]
BEST MODE FOR CARRYING OUT THE INVENTION
[1. Overview and description of functional blocks]
An embodiment of the present invention will be described with reference to the drawings. FIG. 1 shows a functional block diagram of an acoustic model generation device 1 according to the present invention. The acoustic model generation device 1 is a device that generates a statistically processed acoustic model under a specific environment. As shown in FIG. 1, it is connected to a plurality of acoustic model databases 9a to 9n, and includes an extraction destination determining unit 3, a combination ratio determining unit 5, and an acoustic model generating unit 7.
[0056]
In each of the acoustic model databases 9a to 9n, acoustic model data subjected to statistical processing for each phoneme is stored.
[0057]
The extraction destination determining unit 3 determines the conformity of the test data for specific environment judgment composed of the acoustic data and the character string thereof using the acoustic model unit data stored in the acoustic model databases 9a to 9n. Is calculated, and an acoustic model database as a selection candidate is specified. When two or more acoustic model unit data stored in each candidate acoustic model database are combined for each unit data, the combination ratio determining means 5 performs the calculation on the acoustic data corresponding to the written character string of the test data for determination. The ratio of the combination of the acoustic model databases that is higher than the obtained degree of fit is determined. The acoustic model generation means 7 receives all the acoustic model unit data from the acoustic model database of the combination candidate, and converts the acoustic model unit data synthesized at the determined combination ratio as a statistically processed acoustic model under the specific environment. Generate.
[0058]
[2. Hardware configuration)
The hardware configuration of the acoustic model generation device 1 shown in FIG. 1 will be described with reference to FIG. FIG. 2 is an example of a hardware configuration in which the acoustic model generation device 1 is configured using a CPU.
[0059]
The acoustic model generation device 1 includes a CPU 23, a memory 27, a hard disk 26, a CRT 30, a CDD (CD ROM drive) 25, an input device 28, a communication unit 32, and a bus line 29. The CPU 23 controls each unit via the bus line 29 according to a program stored in the hard disk 26. Note that, for example, Linux (trademark) or the like may be adopted as the operating system.
[0060]
The hard disk 26 stores a generation program described later. This program is read from the CDROM 25a storing the program via the CDD 25 and installed on the hard disk 26. In addition to the CDROM, a program such as a flexible disk (FD) or an IC card may be installed on a hard disk from a computer-readable recording medium. Furthermore, you may make it download using a communication line.
[0061]
In the present embodiment, by installing the program from the CDROM to the hard disk 26, the computer indirectly executes the program stored in the CDROM. However, without being limited to this, the program stored in the CDROM may be directly executed from the CDD 25. Note that, as a program executable by a computer, not only a program that can be directly executed by simply installing it as it is, but also a program that needs to be once converted into another form (for example, decompresses a data-compressed program) Etc.), and also includes those that can be executed in combination with other module parts.
[0062]
The hard disk 26 stores in advance a database specifying information storage unit 26t for specifying a database to be searched, a weight determination rule storage unit 26k, and a program storage unit 26p. In the present embodiment, a URL is adopted as the database specifying information. The programs and the weight determination rules stored in the program storage unit will be described later.
[0063]
The communication unit 32 performs data transmission processing with another computer connected to a network such as the Internet or an intranet.
[0064]
The acoustic model databases 9a to 9n connected to the acoustic model generation device 1 will be described. Each of the acoustic model databases 9a to 9n stores acoustic model data. In this embodiment, triphone format HMM format data (hereinafter referred to as phoneme HMM data) is employed for each phoneme as acoustic model data. The triphone format refers to a format in which the same phoneme “ko” has different feature values depending on the preceding and following phonemes, and is therefore classified as separate data for each of the preceding and following phonemes. For example, each phoneme in “ko”, “n”, “ni”, “chi”, and “ha” is “− / k / o”, “/ k / o / n”, “o / n / n”, “n / n / i "," n / i / ch "," i / ch / i "," ch / i / w "," i / w / a / ", and" w / a /-". Note that "-" in the phoneme "-/ k / o" indicates that there is no sound.
[0065]
The phoneme HMM data will be described with reference to FIG. The phoneme HMM data is represented by a vector having a predetermined feature amount as an element. In the present embodiment, as the elements of the feature vector, an amplitude spectrum obtained as a result of performing short-time spectrum analysis, a logarithmic amplitude spectrum which is a logarithmic value of the amplitude spectrum, and a cepstrum coefficient value which is an inverse cosine transform value of the logarithmic amplitude spectrum value. , And delta parameters obtained by regression analysis of their time changes. Each element is represented by its average, variance, self-loop probability, and state transition probability.
[0066]
In the present embodiment, the short-time spectrum analysis is performed every 10 ms for an analysis time of 20 ms, but the analysis time is not particularly limited.
[0067]
[3. Overview of processing)
1) In each database, phoneme HMM data is stored in advance for each phoneme in a triphone format.
[0068]
2) The acoustic model generation device gives, to each candidate database, a request to receive phoneme HMM data of phonemes constituting the test data for determination. For example, if the determination for the test data is "Hello", the phoneme HMM data "- / k / o", "/ k / o / n", "o / n / n", "n / n / i", Request phoneme HMM data of "n / i / ch", "i / ch / i", "ch / i / w", "i / w / a", "w / a /-". Acoustic model generation apparatus, to generate a consolidated phoneme data side-by-side with "Hello" from the phoneme HMM data received.
[0069]
3) The acoustic model generation device judges the similarity by comparing the audio feature amount of the test data for judgment with the audio feature amount of the connected phoneme data. Then, the data is stored as an evaluation of the data stored in the DB.
[0070]
4) The above processes 2) and 3) are performed for all candidate databases.
[0071]
5) A predetermined number is selected from among the candidate databases as the selected candidate databases.
[0072]
6) Combine the same phoneme HMM data stored in the selected candidate database at a certain ratio to generate synthesized phoneme HMM data, generate synthesized phoneme-connected HMM data by connecting these, and generate the generated synthesized phoneme-connected HMM. Obtain an evaluation value for the data. By changing the ratio, a ratio at which the evaluation value becomes higher is determined.
[0073]
7) Receiving the phoneme HMM data for all phonemes from the selection database, synthesizing each phoneme HMM data based on the ratio to generate synthesized phoneme HMM data, and obtaining this as an acoustic model under a specific environment.
[0074]
[4. flowchart〕
The processing in the acoustic model generation device 1 will be described with reference to FIG. Hereinafter, a case where an acoustic model for speech recognition in a car navigation system is generated will be described as an example.
[0075]
First, the operator of the acoustic model generation apparatus 1 records tens of sentences of utterances in the use environment (use place, user, utterance content, recording device) of the acoustic model desired to be generated, and generates typical voice data. And More specifically, in a car, a predetermined sentence, for example, “Search for a nearby convenience store” may be spoken by several people and stored. The CPU 23 extracts acoustic parameters from the typical voice data and stores them in association with the character strings of the text. Specifically, the speech feature amount is extracted and stored together with the data of the character string corresponding to the spoken word. Note that the audio feature has already been described, and a description thereof will be omitted.
[0076]
Next, the CPU 23 of the acoustic model generation device 1 performs a process of determining a selected candidate database (step S1 in FIG. 4).
[0077]
The process of determining the selected candidate database will be described with reference to FIG. The CPU 23 specifies a database to be accessed using the database specifying information stored in the database specifying information storage unit. Here, it is assumed that eleven databases of the acoustic model databases 9a to 9k have been specified.
[0078]
The CPU 23 initializes the process number i (step S11 in FIG. 5), and makes a request to the i-th database to acquire necessary phoneme HMM data (step S13). Specifically, in order to generate acoustic data of a character string “search for a nearby convenience store” for the acoustic model database 9a as the 0th database, phonemes “− / ch / i” and “ch / i” are used. / K "," i / k / a "," k / a / k "," a / k / u "," k / u / n "," u / n / o "," n / o / k " ”,“ O / k / o ”,“ k / o / n ”,“ o / n / b ”,“ n / b / i ”,“ b / i / n ”,“ i / n / i ”, “N / i / w”, “i / w / o”, “w / o / s”, “o / s / a”, “s / a / g”, “a / g / a”, “g / A / s "," a / s / u ", and" s / u /-". Then, by using the phoneme HMM data stored in the database, the sound data of “Search for a song” is generated, and the similarity between the sound data generated from the typical sound data and the generated sound data is calculated. (Step S15). In the present embodiment, the probability that typical voice data is output from acoustic data generated using phoneme HMM data is calculated, and the obtained value is used as the similarity.
[0079]
In addition, as a method of determining the similarity, various calculation methods conventionally used can be adopted. For example, the Euclidean distance between long-time spectra may be compared.
[0080]
The CPU 23 stores the obtained value as the evaluation of the acoustic model stored in the acoustic model database 9a (Step S17).
[0081]
Next, the CPU 23 determines whether or not the determination has been completed for all the candidate databases (step S19). In this case, since the processing has not been completed, the processing number i is incremented (step S21).
[0082]
Next, the CPU 23 repeats the processing from step S13 to step S17 for the first database. Specifically, the phonemes “− / ch / i”, “ch / i / k”, “i / k / a”, “k / a / k”, “a / k” are stored in the acoustic model database 9b. / U "," k / u / n "," u / n / o "," n / o / k "," o / k / o "," k / o / n "," o / n / b ""," N / b / i "," b / i / n "," i / n / i "," n / i / w "," i / w / o "," w / o / s ", "O / s / a", "s / a / g", "a / g / a", "g / a / s", "a / s / u", "s / u /-" The HMM data is received, and the evaluation value of the database is obtained.
[0083]
When determining in step S19 that the determination has been completed for all candidate databases, the CPU 23 determines a candidate database to be selected based on the evaluation value (step S23). In the present embodiment, a threshold value is set, and all databases having the threshold value or more are set as candidates.
[0084]
Note that an upper limit may be set for the maximum number of candidate databases.
[0085]
Next, the CPU 23 performs a combination ratio determination process (step S3 in FIG. 4). The combination ratio determination process is executed by obtaining a weighting coefficient for the selected candidate database. One coefficient λi for the selected candidate database Di is determined for each selected candidate database, and under the condition that the sum of the coefficients λ of all the selected candidate databases is 1, the combination having the highest similarity is determined. It is determined. The combination ratio determination process will be described with reference to FIG.
[0086]
Hereinafter, a case will be described in which the selected candidate databases determined in step S23 are the two acoustic model databases 9a and 9c.
[0087]
The CPU 23 determines the weight of the selected candidate database using the weight determination rules stored in the hard disk 26 (step S31 in FIG. 6). In the present embodiment, when the number of selected candidate databases is two, first, one weight is set to 0.01 and the other is set to 0.99.
[0088]
The CPU 23 synthesizes the same phoneme HMM data based on the determined weights (step S33). In this case, first, the phoneme HMM data of the first phoneme “− / ch / i” extracted from the acoustic model databases 9a and 9c is synthesized based on the weighting. In the present embodiment, each element of the feature vector for the phoneme is represented by the average value and the variance value, and these are combined at the weighting ratio. For example, FIG. 7 shows a state where the first elements of each feature vector extracted from the acoustic model databases 9a and 9c are weighted and combined. Such synthesis processing may be performed for all elements. In addition, the self-loop probability and the state transition probability are similarly combined with each other at the above ratio.
[0089]
As for the self-loop probability and the state transition probability, an average of the values extracted from both databases may be taken.
[0090]
Next, the CPU 23 determines whether or not all the phoneme HMM data necessary for generating the typical voice data has been processed (step S35). In this case, since there are still remaining phonemes, the other phonemes “ch / i / k”, “i / k / a”, “k / a / k”, “a / k / u”, “k / u / n” "," U / n / o "," n / o / k "," o / k / o "," k / o / n "," o / n / b "," n / b / i ", "B / i / n", "i / n / i", "n / i / w", "i / w / o", "w / o / s", "o / s / a", "s / A / g "," a / g / a "," g / a / s "," a / s / u ", and" s / u /-".
[0091]
When the CPU 23 determines that the synthesis phoneme HMM data has been generated for the necessary phoneme HMM data, the process proceeds from step S35 to step S37 to generate connected phoneme-connected HMM data (step S37).
[0092]
The CPU 23 generates the sound data of “Look for Search” using the phoneme HMM data generated by the synthesis, and calculates the similarity between the sound data generated from the typical sound data and the generated sound data. (Step S39). This process is the same as step S15.
[0093]
The CPU 23 determines whether or not the end condition is satisfied (step S41). In the present embodiment, the coefficient is changed by 0.01, and the process ends when the coefficient becomes 0.99.
[0094]
In this case, since the end condition is not met, the CPU 23 changes the weight based on the weight determination rule. As a result, the weight changes between 0.02 and 0.98.
[0095]
Hereinafter, the CPU 23 repeats the processing from step S33 to step S43 until the end condition is met. If the end condition is satisfied in step S41, the combining ratio is determined (step S45). For this, the coefficient may be determined so as to have the highest similarity.
[0096]
Note that the coefficient can be determined more quickly by using a genetic algorithm or the like.
[0097]
After determining the synthesis ratio, the CPU 23 performs an acoustic model generation process (Step S5 in FIG. 4). Specifically, the CPU 23 receives all the phoneme HMM data from all the selected candidate databases, and generates data obtained by synthesizing the phoneme HMM data of the selected candidate database using the coefficients obtained in step S3.
[0098]
As described above, based on information on the usage environment assumed by the speech recognition system, appropriate data is selected from a plurality of pieces of speech data that are stored in a distributed manner, and a statistic extracted from the selected data is used. Thus, an acoustic model having the same accuracy as when a large amount of speech is used as learning data can be created. In other words, it is possible to create a speech model suitable for speech recognition under a specific use environment without collecting a large amount of learning speech data under the use environment.
[0099]
In addition, when various audio data distributed and stored on a network are used for learning an acoustic model, it is possible to determine whether the data is appropriate for use in a desired use environment. Further, an appropriate audio model can be generated even if the data amount of the audio data distributed and stored on the network varies depending on the recording environment.
[0100]
Further, in the present embodiment, the case where the number of selected candidate databases determined in step S23 is two has been described, but the present invention can be similarly applied to the case where there are two or more. This is because each coefficient may be determined under the condition that the sum of the coefficients λ of all the selected candidate databases satisfies 1. When the number of selected candidate databases is large, the number of operations per round is drastically increased. In this case, the value to be changed may be increased.
[0101]
Note that the coefficient λi can take any value from 0 to 1. Therefore, some of the selected candidate databases may not be finally used for synthesis. Thus, only a plurality of selected candidate databases are specified, and only those necessary for generating an acoustic model in a desired environment are adopted.
[0102]
Note that the termination condition is not limited to the above, and a case where a similarity equal to or more than a predetermined value is obtained may be used as the termination condition.
[0103]
[5. Embodiment when applied to language model]
In the above embodiment, the case where the present invention is applied to an acoustic model has been described as an example. However, the present invention can be similarly applied to a natural language model. That is, a language model suitable for typical language data may be synthesized from a plurality of language models already existing. Hereinafter, a case will be described as an example where each language database stores a word trigram and its statistical processing data. Also, the hardware configuration is the same as that of FIG.
[0104]
First, a basic processing flowchart is shown in FIG. As described above, the analysis processing of the test data for determination (hereinafter referred to as typical data) (step S51), the processing of determining the language database of the selected candidate (step S53), the processing of determining the combination ratio (step S55), and the processing of generating the language model (step S55) Step S57) may be performed.
[0105]
The typical data analysis process will be described. When a test sentence for determination is given, the CPU 23 obtains statistical data of a word trigram for typical data. In this case, words may be extracted from the typical data, and statistical data of what words appear as words connected before and after each word may be calculated for each word. The linked data composed of the three words before and after a certain word is called a word trigram. This makes it possible to obtain a word trigram for typical data and its appearance frequency.
[0106]
Next, the CPU 23 performs a process of determining a selected candidate language database. This determination process will be described with reference to FIG. For the typical data, the evaluation value of each word trigram is expressed as a vector having the same number of dimensions as the number of word trigrams (step S61). Here, it is assumed that 1000-dimensional vector data is obtained.
[0107]
Next, the CPU 23 extracts the frequency of the same word trigram from each language database (step S63). Thereby, the same 1000-dimensional vector data can be obtained. In some cases, there is no word trigram in the language database. In this case, the frequency may be set to zero.
[0108]
The CPU 23 compares the similarity between the two sets of vector data (step S65). A known method can be used for the similarity between two vector data. In the present embodiment, the similarity is obtained by calculating the Euclidean distance. The Euclidean distance can be obtained by squaring the difference between the values of each dimension and calculating the sum thereof.
[0109]
For example, when the frequency in the test data and the frequency in the database Di are obtained for the word trigram as shown in FIG. 10, the similarity between the two is (0.01365-0.00175) ＾ 2 + (0.00455). −0.00000) ＾ 2 +...
[0110]
The CPU 23 determines whether or not the processing has been completed for all the language databases for which it is determined whether or not to be determined as the selected candidate (step S67), and determines a higher language database based on a predetermined selection criterion as the selected candidate language database. (Step S69).
[0111]
The CPU 23 performs a combination ratio determination process (Step S55 in FIG. 8). The outline of this processing is the same as that of the above acoustic model except for the calculation processing of the similarity. This will be briefly described with reference to FIG.
[0112]
Hereinafter, a case where the selected candidate databases determined in step S53 are two language model databases Di and Dj will be described.
[0113]
The CPU 23 determines the weight of the selected candidate database using the weight determination rule stored in the hard disk 26 (step S81 in FIG. 11). In the present embodiment, when the number of selected candidate databases is two, initially, the weight of the database Di is set to 0.01 and the weight of the database Dj is set to 0.99.
[0114]
The CPU 23 combines the appearance frequencies of the same word trigrams with the determined weights (step S83). For example, if the appearance frequency of the word trigram “を, suru, no” in the database Di is 0.00175 and the appearance frequency in the database Dj is 0.0001, the combination is 0.00175 * 0.01 + 0.00001 *. 0.99 = 0.0010075.
[0115]
The CPU 23 determines whether all the word trigrams necessary for evaluating the typical data have been processed (step S85). In this case, since there still remains, the process of step S83 is performed for another word trigram.
[0116]
When the CPU 23 determines that the combining process of step S83 is completed for the necessary word trigram, the process proceeds from step S85 to step S89, and calculates the similarity with the test data (step S89). This process is the same as step S65 in FIG.
[0117]
The CPU 23 determines whether or not the end condition is satisfied (step S91). In the present embodiment, the coefficient is changed by 0.01, and the process is terminated when the coefficient becomes 0.99.
[0118]
In this case, since the end condition is not met, the CPU 23 changes the weight based on the weight determination rule. As a result, the weight changes between 0.02 and 0.98.
[0119]
Hereinafter, the CPU 23 repeats the processing from step S83 to step S93 until the end condition is met. If the end condition is met in step S91, the combining ratio is determined (step S45). For this, the coefficient may be determined so as to have the highest similarity.
[0120]
After determining the combination ratio, the CPU 23 performs a language model generation process (step S57 in FIG. 8). Specifically, the CPU 23 receives all the word trigrams from all the selected candidate databases, and uses the coefficient obtained in step S55 to synthesize data obtained by synthesizing the appearance frequencies of the word trigrams in the selected candidate database. Generate.
[0121]
In this manner, as with the acoustic model, a language model suitable for the statistically processed test data can be generated from a plurality of language model databases. Note that the modification described in the acoustic model, such as when there are three or more selected candidate databases, can be similarly applied to a language model.
[0122]
[6. Other embodiments]
As described above, in the present embodiment, the target acoustic model is generated by combining the audio data that has already been stored at a predetermined ratio. Therefore, it is not necessary to collect a large amount of audio data to generate an acoustic model. In addition, since only the parameter or the statistical processing data is acquired from each database instead of the original sound data, there is no problem of using the original sound data for other purposes. It is also possible to charge according to the amount of data stored in each database.
[0123]
In the above embodiment, the case where the phoneme HMM data is stored in advance in each acoustic model database has been described, but the phoneme HMM data may be generated from the raw speech data or the feature amount extracted from the raw data each time. Good.
[0124]
In the above embodiment, for a voice model in an environment of a car, a database storing voice data in a car and a database storing voice data of a man are synthesized at a predetermined ratio, and an acoustic model is obtained. The case of generation is described as an example. However, the present invention focuses on that the method for generating the acoustic model for the specific person is a combination of the plurality of basic data at a predetermined ratio, and has a predetermined term spoken. It is to determine which database should be combined with which basic data in which database, and generate the acoustic model for the specific person. Therefore, any database under any environment can be combined as long as a plurality of databases from which statistical data can be extracted. For example, if there is no acoustic model of a man in his forties, a database of infant voices, a database of adult voices, and a database of old people's voices are appropriately combined to produce an acoustic model of a male in his forties. Can also be generated.
[0125]
In the above embodiment, information (URL, etc.) for specifying the candidate database is stored and searched, but the search may be performed using a search engine or the like.
[0126]
In the above embodiment, when determining whether or not to determine the selected candidate database, the phoneme HMM data is received from each acoustic model database, and the similarity is determined by the language model generation device. The typical phonetic data is given to the acoustic model database, only the fitness when using the acoustic model stored in the acoustic model database is received, and after determining as the selected candidate, the desired phoneme HMM data is obtained from each database. You may make it receive. That is, when determining whether or not to select as a selected candidate, the phoneme HMM data stored in each database is not particularly required, and an evaluation value (score) based on the acoustic model stored in the database may be obtained.
[0127]
Further, in the above embodiment, a database having a predetermined degree of matching from the candidate database is set as a selected candidate database, and a weighting coefficient is obtained for the selected candidate database. Instead, the coefficients may be obtained for all the candidate databases.
[0128]
In the above embodiment, for simplicity of description, the case where each vector element of each phoneme is represented by one normal distribution has been described. However, in reality, in an acoustic model, the feature amount of each phoneme is often represented by combining a plurality of normal distributions. In this case, the expected value of the number of distribution element selection in the composite distribution, the expected value of the acoustic parameter value for each distribution element in the composite distribution, the expected value of the square value of the acoustic parameter for each distribution element in the composite distribution, and the like are further considered. What should I do?
[0129]
In the above embodiment, a case has been described in which each database already stores statistically processed model data. However, basic data for calculating the model data may be stored.
[0130]
In the above-described embodiment, the degree of matching between the model data stored in the database and the test data is calculated from the candidate database, and a predetermined number of selected candidate databases are obtained. Is not limited to this, and for example, one having a matching degree equal to or higher than a predetermined value may be selected.
[0131]
If attribute data is added to data stored in the candidate database instead of such selection criteria, this may be used.
[0132]
In the above embodiment, the required model unit data is received from each candidate database, and the combined data and the test data are compared. However, the test data is given to each candidate database, and the comparison result (evaluation value) is given. May be received and sorted.
[0133]
In the above embodiment, the similarity is determined based on one typical voice data. However, the similarity may be determined based on an average evaluation value of a plurality of typical voice data.
[0134]
The term “network connection” is a concept that includes not only a case where each device is connected via a so-called network such as the Internet or an intranet, but also a case where each device is connected via some kind of data transfer path. Further, the case where even one device physically stores acoustic model data and language model data is stored in a plurality of areas.
[0135]
In the above-described embodiment, the case where acoustic model data and language model data are generated each time has been described. However, once generated acoustic model data and language model data are stored, new test data is provided. In this case, it may be determined whether or not there is a match in the model data that has already been accumulated, and if so, it may be output as it is. As a result, the existing model data can be effectively used. Whether or not there is a match may be determined, for example, in advance by a threshold value.
[0136]
In the present embodiment, in order to realize the functions shown in FIG. 1, this is realized by software using a CPU. However, some or all of them may be realized by hardware such as a logic circuit.
[0137]
Note that a part of the processing of the program may be further performed by an operating system (OS).
[Brief description of the drawings]
FIG. 1 is a functional block diagram of an acoustic model generation device 1 according to the present invention.
FIG. 2 is a diagram illustrating an example of a hardware configuration in which the acoustic model generation device 1 is implemented using a CPU.
FIG. 3 is a diagram showing a data structure of an acoustic model.
FIG. 4 is an overall flowchart of an acoustic model generation process.
FIG. 5 is a detailed flowchart for determining a selected candidate database.
FIG. 6 is a detailed flowchart of determining a combination ratio.
FIG. 7 is a diagram for explaining a combining process.
FIG. 8 is an overall flowchart of a language model generation process.
FIG. 9 is a detailed flowchart for determining a selected candidate database.
FIG. 10 shows a data structure of a word trigram.
FIG. 11 is a detailed flowchart of determining a combination ratio.
[Explanation of symbols]
1 ... Acoustic model generation device
23 ... CPU
27 ... Memory

Claims (21)

An acoustic model generation method for generating phoneme HMM data under a specific environment,
Network connection with a plurality of candidate data storage devices for storing phoneme HMM data,
Given the test data for determination in the specific environment composed of acoustic data and its notation character string, the phoneme HMM corresponding to the notation character string among the stored phoneme HMM data is stored from each of the candidate data storage devices. Receiving the data, determining the similarity between the phoneme-connected HMM data generated by connecting the phoneme HMM data, and the sound data of the test data for determination,
Based on the obtained similarity, specify a plurality of selected candidate data storage devices from the candidate data storage device,
When two or more phoneme HMM data stored in each of the selected candidate data storage devices are combined for each phoneme, a synthesis having a higher similarity is obtained for phoneme-connected HMM data corresponding to the written character string of the test data for determination. Find the percentage,
Generating synthesized phoneme HMM data obtained by synthesizing the phoneme HMM data stored in each of the selected candidate data storage devices at the synthesis ratio as an acoustic model under the specific environment;
An acoustic model generation method characterized by the following. An acoustic model generation method for generating a statistically processed acoustic model under a specific environment, comprising the following steps:
Identifying a selected candidate data storage device from the plurality of data storage devices connected to the network, and determining a degree of conformity between the acoustic data and the test data for determination in the specific environment, which is composed of notation character strings thereof. A selecting step of specifying a selected candidate data storage device from the fitness when the calculation is performed using the acoustic model unit data stored in each data storage device;
Regarding synthesized acoustic model unit data obtained by combining two or more acoustic model unit data stored in the selected candidate data storage device for each unit data, the combined acoustic model unit data corresponding to the notation character string of the test data for determination is used. A combination ratio determination step of determining a combination ratio at which the similarity is higher,
A generating step of receiving all acoustic model unit data from the selected candidate data storage device and generating synthetic acoustic model unit data synthesized at the determined combination ratio as a statistically processed acoustic model under the specific environment;
An acoustic model generation method characterized by the following. The acoustic model generation method according to claim 2,
The data storage device stores basic data for generating acoustic model unit data, and when an acoustic model unit data generation request is given, generating and storing the request,
What is characterized by. The acoustic model generation method according to claim 2,
The acoustic model unit data includes at least an average value and a variance value of a feature vector for the phoneme,
In the combination ratio determination step, combining an average value and a variance value for each feature vector of a phoneme corresponding to the written character string,
What is characterized by. An acoustic model generation device that generates a statistically processed acoustic model under a specific environment,
A means for specifying a selected candidate data storage device from a plurality of candidate data storage devices connected to a network, comprising: A selected candidate determining unit that specifies a data storage device to be selected from the fitness when the calculation is performed using the acoustic model unit data stored in each of the candidate data storage devices,
When two or more acoustic model unit data stored in each of the candidate data storage devices are combined for each unit data, the acoustic data corresponding to the written character string of the test data for determination is stored in the selected candidate data storage device. Combination ratio determining means for determining a ratio of a combination of the selected candidate data storage devices that is higher than the fitness calculated by the stored acoustic model unit data,
Acoustic model generation means for receiving all acoustic model unit data from the selected candidate data storage device and generating synthetic acoustic model unit data synthesized at the determined combination ratio as a statistically processed acoustic model under the specific environment ,
An acoustic model generation device comprising: An acoustic model generation device that generates acoustic model unit data under a specific environment,
Test data storage means for storing test data for determination under the specific environment composed of sound data and its notation character string,
From a plurality of selected candidate data storage device, acoustic model unit data acquisition means for acquiring acoustic model unit data corresponding to the written character string of the test data for determination,
For the acoustic model unit data obtained from the plurality of selected candidate data storage devices, the acoustic model unit data is synthesized in the same acoustic model unit at a predetermined synthesis ratio to generate synthesized acoustic model unit data, and these are connected. For the generated connected acoustic model unit data, a composition ratio determining unit that calculates a degree of conformity with the acoustic data of the test data for determination and determines a composition ratio at which the degree of conformity is higher,
Generating means for receiving all acoustic model unit data from the selected candidate data storage device based on the synthesis ratio and generating synthetic acoustic model unit data synthesized at the determined synthesis ratio as an acoustic model under the specific environment ,
An acoustic model generation device comprising: The acoustic model generation device according to claim 6,
From a plurality of candidate data storage devices connected to the network, from the degree of conformity with the connected acoustic model unit data generated by connecting the acoustic model unit data to form the notation character string of the test data for determination, Identifying the selected candidate data store;
What is characterized by. The acoustic model generation device according to claim 6,
Attribute data indicating the attribute is added to the judgment test data,
Identifying the selected candidate data storage device based on the degree of matching with the attribute data indicating the attribute of the acoustic model stored in the plurality of candidate data storage devices connected to the network,
What is characterized by. In the acoustic model generation device according to claim 7 or 8,
From the plurality of candidate data storage device, receiving acoustic model unit data for constituting a notation character string of the test data for determination, and linking them to generate connected acoustic model unit data,
What is characterized by. In the acoustic model generation device according to claim 7 or 8,
The plurality of candidate data storage devices, providing the test data for determination, receiving an evaluation value when evaluated using acoustic model unit data stored in each candidate data storage device, the selected candidate data storage Identifying the device,
What is characterized by. A language model generation device that generates a statistically processed language model under a specific environment,
Judgment test data storage means for storing the judgment test data under the specific environment,
From a plurality of selected candidate data storage devices, language model unit data acquisition means for acquiring language model unit data included in the test data for determination,
For the language model unit data acquired from the selected candidate data storage device, the same language model unit data is synthesized at a certain synthesis ratio to generate synthesized language model unit data, and the appearance frequency in the test data for determination is calculated, Means for determining a combination ratio at which the degree of conformity is higher,
Receiving all language model unit data from the selected candidate data storage device based on the combination ratio, and generating combined language model unit data combined at the determined combination ratio as a language model under the specific environment;
A language model generation device characterized by the following. A model data generation device that generates statistically processed model data under a specific environment,
Judgment test data storage means for storing the judgment test data under the specific environment,
From the selected candidate data storage device, model unit data obtaining means for obtaining model unit data included in the test data for determination,
For the model unit data obtained from the selected candidate data storage device, the same model unit data is synthesized at a certain synthesis ratio to generate synthetic model unit data, and the degree of conformity with the test data for judgment is calculated. Composition ratio determining means for determining a composition ratio at which
On the basis of the combination ratio, all model unit data is received from the selected candidate data storage device, and combined model unit data combined at the determined combination ratio is generated as statistically processed model data under the specific environment. thing,
A model data generation device characterized by the following. A language model generation device that generates a statistically processed language model under a specific environment,
Judgment test data storage means for storing the judgment test data under the specific environment,
From the selected candidate data storage device, a language model unit data acquisition unit for acquiring language model unit data included in the test data for determination,
With the acquired language model unit data as acquisition target language model unit data, an appearance frequency for the same language model unit data is acquired from the selected candidate data storage device, and for each language model unit data, A composition ratio determining unit that computes the frequency of appearance of the synthesized language model unit data by combining with the ratio, calculates the degree of conformity with the frequency of appearance in the test data for determination, and determines the composition ratio at which the degree of conformity is higher;
Receiving all language model unit data from the selected candidate data storage device based on the combination ratio, and generating combined language model unit data combined at the determined combination ratio as a language model under the specific environment;
A language model generation device characterized by the following. An acoustic model generation program that generates statistically processed model data under a specific environment, and is a program for causing a computer to execute the following processing.
Acquisition of acoustic model unit data corresponding to the notation character string of the test data for judgment, from a plurality of selected candidate data storage devices, for the test data for judgment in the specific environment composed of sound data and its notation character string And
With respect to the acoustic model unit data obtained from the plurality of selected candidate data storage devices, the acoustic model unit data is synthesized at a certain synthesis ratio in the same acoustic model unit to generate synthetic acoustic model unit data, and the combined acoustic model unit data is generated. For the connected acoustic model unit data, the degree of conformity with the acoustic data of the test data for judgment is calculated, and a synthesis ratio at which the degree of conformity is higher is determined,
On the basis of the combination ratio, receiving all acoustic model unit data from the selected candidate data storage device, and generating combined acoustic model unit data combined at the determined combination ratio as an acoustic model under the specific environment,
An acoustic model generation program characterized by the following. A model data generating program for generating statistically processed model data under a specific environment, the program causing a computer to execute the following processing,
From the selected candidate data storage device, to obtain the model unit data included in the stored test data for determination under the specific environment,
For the model unit data obtained from the selected candidate data storage device, the same model unit data is synthesized at a certain synthesis ratio to generate synthetic model unit data, and the degree of conformity with the test data for judgment is calculated. Is determined to be higher,
On the basis of the combination ratio, all model unit data is received from the selected candidate data storage device, and combined model unit data combined at the determined combination ratio is generated as statistically processed model data under the specific environment. ,
A model data generation program characterized by the following. An acoustic model generation program that generates statistically processed model data under a specific environment, and is a program for causing a computer to execute the following processing.
For the test data for judgment under the specific environment composed of acoustic data and its notation character string, from a plurality of candidate data storage devices, acquire acoustic model unit data corresponding to the notation character string of the test data for judgment,
For the acoustic model unit data obtained from the plurality of candidate data storage devices, the acoustic model unit data is synthesized at a certain synthesis ratio in the same acoustic model unit to generate synthetic acoustic model unit data, and the combined acoustic model unit data is generated. For the acoustic model unit data, the degree of conformity with the acoustic data of the test data for judgment is calculated, and the synthesis ratio at which the degree of conformity is higher is determined,
On the basis of the synthesis ratio, receiving all acoustic model unit data from the candidate data storage device, and generating synthetic acoustic model unit data synthesized at the determined synthesis ratio as an acoustic model under the specific environment,
An acoustic model data generation program characterized by the following. An acoustic model generation device that generates acoustic model unit data under a specific environment,
Test data storage means for storing test data for determination under the specific environment composed of sound data and its notation character string,
From a plurality of selected candidate data storage device, acoustic model unit data acquisition means for acquiring acoustic model unit data corresponding to the written character string of the test data for determination,
For the acoustic model unit data obtained from the plurality of selected candidate data storage devices, the acoustic model unit data is synthesized in the same acoustic model unit at a predetermined synthesis ratio to generate synthesized acoustic model unit data, and these are connected. For the generated connected acoustic model unit data, a composition ratio determining unit that calculates a degree of conformity with the acoustic data of the test data for determination and determines a composition ratio at which the degree of conformity is higher,
Generating means for receiving all acoustic model unit data from the selected candidate data storage device based on the synthesis ratio and generating synthetic acoustic model unit data synthesized at the determined synthesis ratio as an acoustic model under the specific environment ,
Generated acoustic model storage means for storing the generated acoustic model,
For the newly given test data for judgment, the degree of matching with the generated acoustic model stored in the generated acoustic model storage means is calculated, and if it exceeds a predetermined threshold, the generated acoustic model is converted to the specific environment. An acoustic model output means for outputting as data of an acoustic model below,
An acoustic model generation device comprising: An acoustic model generation method for generating acoustic model unit data under a specific environment,
Given the test data for determination under the specific environment composed of the acoustic data and its notation character string, it is stored in each candidate data storage device to generate acoustic model unit data corresponding to the notation character string. Predetermined parameter data extracted from the obtained acoustic data or statistical data obtained by statistically processing the parameter data, and connected acoustic model unit data generated by connecting the acoustic model unit data, and the test data for determination. Judgment of the similarity to the acoustic data of, to specify a plurality of selected candidate data storage device,
Based on the obtained similarity, specify a plurality of selected candidate data storage devices from the candidate data storage device,
When two or more acoustic model unit data stored in each of the selected candidate data storage devices are combined for each phoneme, the similarity of the connected acoustic model unit data corresponding to the written character string of the test data for determination is higher. Is calculated,
At this combination ratio, predetermined parameter data extracted from the acoustic data stored in each of the selected candidate data storage devices or statistical data obtained by performing statistical processing on the parameter data is obtained, and the acoustic model unit data is generated. Along with generating synthetic acoustic model unit data obtained by synthesizing these acoustic model unit data, outputting as an acoustic model under the specific environment,
An acoustic model generation method characterized by the following.
In this way, at the stage of the tentative selection, data for determining the degree of coincidence of some data is received, and when it is determined that the data is to be used for synthesis, by receiving all acoustic model unit data from the candidate data storage device, The acoustic model unit data is received only from the candidate data storage device that actually generates the acoustic model. As a result, billing according to use becomes possible. Also, the candidate data storage device is not the acoustic data but the predetermined parameter data extracted from the acoustic data or the statistical data obtained by performing a statistical process on the parameter data. Therefore, privacy can be secured and the problem of unintended use of data stored in the data storage device can be prevented. A language model generation method for generating a statistically processed language model under a specific environment,
When the test data for determination under the specific environment is given, statistical processing is performed on predetermined parameter data or parameter data for generating language model unit data included in the test data for determination from each candidate data storage device. Statistics data,
The generated language model unit data and the similarity of the appearance frequency of the language model unit data present in the test data for determination are determined, and a plurality of selected candidate data storage devices are specified.
When two or more language model unit data generated from the data extracted from each of the selected candidate data storage devices are combined for each language model unit data, the similarity with the language model unit data present in the test data for determination Finds a higher synthesis ratio,
Receives predetermined parameter data for all data stored in each of the selected candidate data storage devices or statistical data obtained by performing statistical processing on these parameter data, and synthesizes language data obtained by synthesizing them at the synthesis ratio. As a language model under the specific environment,
A language model generation method characterized by the following. A model data generation method for generating statistically processed model data under a specific environment,
Given the test data for determination under the specific environment, it is necessary to calculate the degree of matching with the test data for determination from each candidate data storage device to determine whether or not to perform the synthesis. Receives only a part of the data as data in a predetermined parameter format or data obtained by statistically processing the data in the parameter format, calculates the degree of matching with the test data for determination, and specifies a plurality of selected candidate data storage devices. And
When two or more model unit data generated from the data extracted from the selected candidate data storage devices are combined for each model unit data, the similarity of the test data for judgment with the model unit data becomes higher. Find the composition ratio,
Receiving predetermined parameter data for all data stored in each of the selected candidate data storage devices or statistical data obtained by performing statistical processing on these parameter data, and synthesizing model unit data obtained by synthesizing them at the synthesis ratio. Generating as model data under the specific environment,
A model data generation method characterized by the following. An acoustic model output device that generates acoustic model unit data under a specific environment,
For the acquired plurality of acoustic model unit data, the acoustic model unit data is synthesized at a certain ratio in the same acoustic model unit to generate synthetic acoustic model unit data, and these are connected in the order of the notation character string. A fitness calculating unit that generates acoustic model unit data and calculates a fitness between the connected acoustic model unit data and the test data for determination;
A ratio determining unit that changes the ratio based on a predetermined ratio determination rule and determines a more suitable ratio,
Based on the synthesis ratio, receives all the acoustic model unit data stored in the selected candidate data storage device or basic data for generating the same, and synthesizes acoustic model unit data synthesized at the determined synthesis ratio. Output means for generating and outputting as an acoustic model under the specific environment,
An acoustic model output device comprising:

Images