Training data generating method for training filled pause detecting model and device therefor
Disclosed is a method for generating training data for training a filled pause detecting model and a device therefor, which execute mounted artificial intelligence (AI) and/or machine learning algorithms in a 5G communication environment. The method includes acquiring acoustic data including first speech data including a filled pause, second speech data not including a filled pause, and noise, generating a plurality of noise data based on the acoustic data, and generating first training data including a plurality of filled pauses and second training data not including a plurality of filled pauses by synthesizing the plurality of noise data with the first speech data and the second speech data. According to the present disclosure, training data for training a filled pause detecting model in a simulation noise environment can be generated, and filled pause detection performance for speech data generated in an actual noise environment can be enhanced.
1. A method for generating training data for training a filled pause detecting model, the method comprising:
acquiring acoustic data comprising first speech data including a filled pause, second speech data that does not include a filled pause, and noise, wherein the first speech data is labeled as speech data including a filled pause and the second speech data is labeled as speech data that does not include a filled pause;
generating a plurality of noise data based on the acoustic data by:
setting one or more types of noise parameters for establishing a simulation noise environment,
setting a range of the one or more types of the noise parameters,
performing random sampling on an arbitrary noise parameter value of the set range of the noise parameters, and
generating the plurality of noise data by synthesizing the random-sampled noise parameter value; and
generating first training data including a plurality of filled pauses and second training data that does not include a plurality of filled pauses by synthesizing the plurality of noise data with the first speech data and the second speech data,
wherein the first training data is labeled as speech data including a filled pause and the second training data is labeled as speech data that does not include a filled pause, and
wherein the first training data and the second training data are generated using a domain randomization algorithm.
2. The method of claim 1 , wherein the acquiring the acoustic data comprises acquiring clean data from which noise is not synthesized with respect to the first speech data and the second speech data.
3. The method of claim 1 , wherein the setting the range of the one or more types of the noise parameters comprises setting the range based on an analysis result of an actual noise sample extracted in an actual noise environment.
4. The method of claim 1 , wherein the performing the random sampling comprises performing the random sampling on a noise parameter value according to uniform distribution having a preset interval between an upper value and a lower value in the range of the noise parameters.
5. The method of claim 4 , wherein the noise parameter value of an actual noise environment is included in the preset interval.
6. A non-transitory computer-readable recording medium on which a program for executing a method of generating training data for training a filled pause detecting model, the method comprising:
acquiring acoustic data comprising first speech data comprising a filled pause, second speech data that does not comprise a filled pause, and noise, wherein the first speech data is labeled as speech data comprising a filled pause and the second speech data is labeled as speech data that does not comprise a filled pause;
generating a plurality of noise data based on the acoustic data by:
setting one or more types of noise parameters for establishing a simulation noise environment,
setting a range of the one or more types of the noise parameters,
performing random sampling on an arbitrary noise parameter value of the set range of the noise parameters, and
generating the plurality of noise data by synthesizing the random-sampled noise parameter value; and
generating first training data comprising a plurality of filled pauses and second training data that does not comprise a plurality of filled pauses by synthesizing the plurality of noise data with the first speech data and the second speech data, wherein the first training data is labeled as speech data comprising a filled pause and the second training data is labeled as speech data that does not comprise a filled pause, wherein the first training data and the second training data are generated using a domain randomization algorithm.
7. A training data generating device for training a filled pause detecting model, comprising:
an acquirer configured to acquire acoustic data comprising first speech data including a filled pause, second speech data that does not include a filled pause, and noise, wherein the first speech data is labeled as speech data including a filled pause and the second speech data is labeled as speech data that does not include a filled pause;
a first generator configured to generate a plurality of noise data based on the acoustic data by:
setting one or more types of noise parameters required to establish a simulation noise environment,
setting a range of the one or more types of the noise parameters,
performing random sampling on an arbitrary noise parameter value of the set range of the noise parameters, and
generating the plurality of noise data by synthesizing the random-sampled noise parameter value; and
a second generator configured to generate first training data including a plurality of filled pauses and second training data that does not include a plurality of filled pauses by synthesizing the plurality of noise data with the first speech data and the second speech data, wherein the first training data is labeled as speech data including a filled pause and the second training data is labeled as speech data that does not include a filled pause, wherein the first training data and the second training data are generated using a domain randomization algorithm.
8. The training data generating device of claim 7 , wherein the acquirer is configured to acquire clean data from which noise is not synthesized with respect to the first speech data and the second speech data.
9. The training data generating device of claim 7 , wherein the first generator is configured to set the range based on an analysis result of an actual noise sample extracted in an actual noise environment.
10. The training data generating device of claim 7 , wherein the first generator is configured to perform the random sampling on a noise parameter value according to uniform distribution having a preset interval between an upper value and a lower value in the range of the noise parameters during the random sampling of the noise parameter value.
11. The training data generating device of claim 10 , wherein the first generator is configured to include the noise parameter value of an actual noise environment in the preset interval.