Contextual Biasing in ASR Models Using Domain-Specific Sub-Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) models are inflexible and struggle to effectively utilize contextual information, leading to inaccurate results for specific user groups or domains, and are computationally expensive to update or retrain.
Innovation Solution
The implementation of a computer-implemented method and system that uses a base ASR model combined with sub-models trained on biased data representative of specific domains, allowing for contextual biasing of speech recognition results based on indicators such as one-hot vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single base ASR model is used for general speech recognition, then the model is simple and computationally efficient, but it lacks flexibility and accuracy for specific domains or user groups
Solution Approach 1:
The ASR system is segmented into a base model and multiple domain-specific sub-models. Each sub-model is trained on biased data representing a particular domain (e.g., medical, legal, technical), allowing the system to maintain a simple base structure while adding specialized capabilities when needed through selective activation of relevant sub-models.
2Measurement precision
If multiple separate ASR models are used for different domains, then domain-specific accuracy is improved, but computational cost and training redundancy increase
Solution Approach 1:
Multiple domain-specific models are merged into a single unified architecture where sub-models share the same base model structure and parameters. This allows the system to achieve domain-specific accuracy through parameter biasing while avoiding the computational redundancy of training and executing separate complete models for each domain.
Solution Approach 2:
Instead of training separate models with different architectures, the invention changes the parameters of the base model by training sub-models on biased data representing specific domains. This parameter adaptation allows domain-specific accuracy to be achieved through data-driven biasing rather than through separate model training, reducing computational costs.
3Adaptability or versatility
If ASR models are retrained to adapt to new domains, then domain-specific performance is improved, but training time and computational resources increase
Solution Approach 1:
The invention extracts the domain-specific knowledge from complete model retraining by training separate sub-models on biased data representing specific domains. These sub-models can be independently trained and then integrated into the base model, allowing rapid adaptation to new domains without requiring time-consuming retraining of the entire base model architecture.
Data Source
AI summary
A method for contextual biasing for speech recognition includes obtaining a base automatic speech recognition (ASR) model trained on non-biased data and a sub-model trained on biased data representative of a particular domain. The method includes receiving a speech recognition request including audio data characterizing an utterance captured in streaming audio. The method further includes determining whether the speech recognition request includes a contextual indicator indicating the particular domain. When the speech recognition request does not include the contextual indicator, the method includes generating, using the base ASR model, a first speech recognition result of the utterance by processing the audio data. When the speech recognition request includes the contextual indicator the method includes biasing, using the sub-model, the base ASR model toward the particular domain and generating, using the biased base ASR model, a second speech recognition result of the utterance by processing the audio data.