Method for generating training data and method for post-processing of speech recognition using the same
AssigneeKorea University Research and Business Foundation
InventorsHeuiseok Lim, Chanjun Park
Life of this patent
Estimating the term…
Abstract
Disclosed is a training data construction method and a speech recognition method using the same. The training data construction method is performed by a computing apparatus including at least one processor and includes converting first text data including a plurality of sentences to first speech data; acquiring second speech data by adding noise to the first speech data; and converting the second speech data to second text data. The second text data includes a sentence corresponding to each of the plurality of sentences included in the first text data.
Claims6
A training data construction method performed by a computing apparatus comprising at least one processor, the training data construction method comprising:
- converting first text data, which is an initial text of mono corpus including a plurality of sentences, to first speech data using a predetermined text-to-speech scheme;
- generating second speech data based on the first speech data, the generating comprising:
- adding noise to the first speech data;
- performing a frequency transformation on the first speech data;
- removing a portion of a predetermined frequency band of the first speech data;
- and
- performing a retransformation to a time domain to generate the second speech data;
- and
- converting the second speech data to second text data, which is a final text,
- wherein the second text data includes a plurality of sentences corresponding to each of the plurality of sentences included in the first text data, thereby constituting a parallel corpus with the first text data.
A speech recognition method performed by a computing apparatus comprising at least one processor, the speech recognition method comprising:
- generating training data;
- generating a speech recognition postprocessing model by training a sequence-to-sequence (seq2seq) model using the training data;
- generating a text sentence corresponding to input speech data using a predetermined speech recognition model;
- and
- correcting the text sentence using the speech recognition postprocessing model and outputting the corrected text sentence,
- wherein the generating of the training data comprises:
- converting first text data, which is an initial text of mono corpus including a plurality of sentences, to first speech data using a predetermined text-to-speech scheme;
- generating second speech data based on the first speech data, the generating comprising:
- adding noise to the first speech data;
- performing a frequency transformation on the first speech data;
- removing a portion of a predetermined frequency band of the first speech data;
- and
- performing a retransformation to a time domain to generate the second speech data;
- and
- converting the second speech data to second text data, which is a final text,
- wherein the second text data includes a plurality of sentences corresponding to each of the plurality of sentences included in the first text data, thereby constituting a parallel corpus with the first text data.
Dependencies are inferred from the claim text.