COALA framework enhances ASR contextual biasing via contrastive regularization

Researchers propose COALA, a robust framework designed to enhance speech-augmented language models (SLMs) for automatic speech recognition in complex multi-entity scenarios. The system addresses context-window limitations by mapping SLM latent representations into a discriminative space to quantify matching intensity between audio segments and candidate entities.

• Maps SLM latent representations into a specialized discriminative space to identify relevant target entities from large-scale biasing lists.
• Addresses training collapse in prior studies when handling multi-target utterances where multiple rare words co-occur.
• Utilizes contrastive regularizer and biasing score estimation techniques.
• Demonstrates superior contextual biasing performance across various biasing list scales on the LibriSpeech benchmark.