Speech Model Pre-Training Using Deterministic Feature-Vector Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech processing technologies require large amounts of transcribed speech data for supervised learning, which is costly, and semi-supervised learning methods using speech without transcription can lead to inaccurate training due to reliance on statistical distributions and language information, hindering efficient execution of speech processing tasks.
Innovation Solution
A pre-trained model is generated using speech data and self-supervised labels derived from feature vectors through a deterministic conversion rule, independent of data distribution, allowing efficient execution of speech processing tasks with a small amount of transcribed text.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used with transcribed speech data, then speech processing accuracy is improved, but data acquisition cost and time increase significantly
Solution Approach 1:
The system performs preliminary action by automatically generating transcribed text data through speech-to-text conversion before the actual speech processing task. This pre-generated text data is then used as training material, eliminating the need for manual transcription and reducing both time and cost while maintaining accuracy requirements.
2Measurement precision
If supervised learning is used with transcribed speech data, then speech processing accuracy is improved, but data acquisition cost increases
Solution Approach 1:
The system implements self-service by automatically generating its own training data through speech-to-text conversion without requiring external manual transcription services. The system uses its own computational resources to convert speech to text, creating a self-sufficient data generation pipeline that reduces external costs while maintaining data quality for accurate processing.
3Loss of energy
If semi-supervised learning is used with speech without transcription, then data acquisition cost is reduced, but training accuracy deteriorates due to reliance on statistical distributions
Solution Approach 1:
The system introduces an intermediary element - automatically generated transcribed text - that bridges the gap between untranscribed speech and accurate training requirements. This intermediary text data serves as a mediator that provides the necessary linguistic information for accurate training while keeping data acquisition costs low, avoiding the need for expensive manual transcription while maintaining training quality.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A speech processing apparatus (20) includes a task execution unit (220). The task execution unit (220) executes a task related to speech processing based on a trained model. The trained model is trained using speech data and one or more labels obtained by converting, according to a predetermined rule, one or more feature vectors extracted from the speech data.