Discrete Speech Representation for Target-Domain ASR Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) models struggle with generalization to unseen data due to overfitting when training data is insufficient, leading to reduced accuracy in underrepresented domains.
Innovation Solution
A method involving a contrastive data selection model to process unlabeled training data, selecting a subset of utterances relevant to a target domain using discrete speech representations and self-supervised training techniques to pre-train the ASR model, followed by fine-tuning with labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ASR models are trained on larger training datasets to improve accuracy, then the accuracy of the ASR model improves, but the computational resources and time required for training increase
Solution Approach 1:
The patent extracts and selects only the most relevant subset of training data for the target domain using a contrastive data selection model, rather than training on the entire large dataset. This selective extraction maintains training efficiency while improving accuracy for the specific target domain by focusing computational resources on the most useful data portions.
Solution Approach 2:
The patent performs preliminary data selection and preprocessing before the actual ASR model training. The contrastive data selection model pre-identifies and prepares the optimal subset of training data, which then accelerates the subsequent training process while ensuring high accuracy for the target domain.
2Reliability
If ASR models are trained on larger training datasets to improve generalization, then the model's ability to handle unseen data improves, but the device complexity and computational resources required increase
Solution Approach 1:
The patent applies local quality by tailoring the training data selection specifically to the target domain characteristics. The contrastive data selection model identifies data with local relevance to the target domain, ensuring the ASR model develops specialized generalization ability for that domain without requiring computational resources proportional to the entire dataset size.
Solution Approach 2:
The patent changes the parameter of training data composition by using a contrastive selection approach that transforms the training set from a generic large dataset into a domain-specific optimized subset. This parameter change in data selection strategy improves generalization ability while reducing computational complexity.
3Loss of energy
If contrastive data selection is used to select a subset of spoken utterances from unlabeled training data, then the computational resources required for training are reduced, but the complexity of the data selection process increases
Solution Approach 1:
The contrastive data selection model operates in a self-supervised manner, automatically selecting relevant training data without requiring manual annotation or intervention. The model uses contrastive learning to self-identify the most useful data portions, reducing the need for complex manual data curation processes while conserving computational resources.
Solution Approach 2:
The patent introduces a contrastive data selection model as an intermediary component between the raw unlabeled training data and the ASR model training process. This intermediary automatically performs the complex task of data selection, simplifying the overall workflow while reducing computational resources needed for the main training task.
Data Source
AI summary
A method includes obtaining a corpus of unlabeled training data including a plurality of spoken utterances, each corresponding spoken utterance of the plurality of spoken utterances includes audio data characterizing the corresponding spoken utterance. The method also includes receiving a target domain. The method also includes selecting, using a contrastive data selection model, a subset of the utterances from the corpus of unlabeled training data that correspond to the target domain. The method includes training an automatic speech recognition (ASR) model on the subset of utterances.