Speech Recognition Post-Processing Using Synthetic Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current grammar error correction (GEC) methods, particularly in machine translation and automatic speech recognition, face challenges such as high time and cost requirements, insufficient memory, and limitations in handling low-resource languages and postprocessing tasks.
Innovation Solution
A training data generation method that converts text to speech, adds noise to the speech data, and then converts it back to text, allowing for the generation of infinite training data. This method is used to develop a speech recognition postprocessing model based on a sequence-to-sequence (seq2seq) model, which corrects errors in speech recognition outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rule-based correction is used for grammar error correction, then correction accuracy can be improved, but time and cost investment increases significantly
Solution Approach 1:
The patent uses text-to-speech synthesis to generate synthetic speech data from text, creating artificial training samples without requiring actual human speech recording. This copying approach allows generating unlimited training data efficiently, resolving the contradiction between data quantity and time cost
Solution Approach 2:
The system uses automated text-to-speech conversion and speech-to-text recognition to self-generate training data without human intervention. The model trains on synthetically generated data and automatically performs post-processing correction, eliminating the need for manual rule establishment while maintaining correction effectiveness
2Reliability
If deep learning models with large size and parameters are used, then model performance is improved, but processing speed decreases and memory requirements increase
Solution Approach 1:
The patent extracts only the essential post-processing function from the complete speech recognition system. Instead of using a full deep learning speech recognition model, it applies a lightweight seq2seq model specifically for correcting recognition errors, thereby maintaining performance while reducing computational overhead and improving speed
Solution Approach 2:
The system segments the speech recognition task into two parts: initial recognition by a standard model, and error correction by a specialized post-processing model. This segmentation allows each component to be optimized independently, with the post-processing model being smaller and faster while handling only the correction function
3Measurement precision
If speech recognition models are used, then recognition accuracy is improved, but development and maintenance cost increases significantly
Solution Approach 1:
The patent makes the seq2seq post-processing model reusable across different domains and applications. The same model architecture can handle various speech recognition outputs and be applied to different speech recognition systems, reducing development cost and entry barrier while maintaining accuracy improvement
Solution Approach 2:
The system generates synthetic training data through text-to-speech copying, eliminating the need for expensive manual annotation and data collection. This approach significantly reduces development cost while providing sufficient training data for accurate post-processing
Data Source
AI summary
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.