Speech Model Pre-Training With Rule-Based Feature Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies face challenges in achieving high accuracy with limited transcription data, as semi-supervised learning methods using erroneous transcriptions hinder performance, and self-supervised learning methods struggle with data distribution imbalances when applying pre-trained models to new speech data.
Innovation Solution
A pre-trained model is developed using self-supervised labels derived from feature vectors of speech data through a deterministic conversion rule, excluding statistical inference, and further trained with a small amount of transcribed text to enhance speech processing tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If self-supervised learning methods are used to pre-train models, then the model can be trained without transcription data, but the model struggles with data distribution imbalances when applying to new speech data
Solution Approach 1:
The patent applies preliminary action by pre-training the model using self-supervised learning on feature vectors extracted from speech data before fine-tuning with limited transcription data. This preliminary pre-training phase prepares the model to handle various speech patterns and distributions, improving its reliability when applied to new speech data while requiring minimal transcription data.
2Quantity of substance
If semi-supervised learning methods using erroneous transcriptions are used, then the model can be trained with more data, but the erroneous transcriptions hinder performance
Solution Approach 1:
The patent extracts and removes erroneous transcription data from the training set, using only high-quality transcribed text data for fine-tuning. By taking out the harmful erroneous transcriptions while retaining useful clean data, the model achieves high speech processing accuracy without being degraded by poor quality training examples.
3Manufacturing precision
If more transcribed text data is used for training, then speech processing accuracy improves, but the cost and time for data preparation increases
Solution Approach 1:
The patent performs preliminary self-supervised pre-training on large amounts of unlabeled speech data, extracting feature vectors and training the model without requiring transcription. This preliminary action reduces the need for extensive transcribed text data later, significantly reducing data preparation time and cost while still achieving high speech processing accuracy through subsequent fine-tuning with minimal clean data.
Data Source
AI summary
A speech processing apparatus processing circuitry. The processing circuitry 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.