Contextual Biasing
A technique in automatic speech recognition that injects prior information into a model via an external candidate word list to improve recognition accuracy for specific vocabulary. Primary implementation approaches include attention-based methods and prefix-tree decoding intervention.
Timeline (last 90 days)
Context biasing engineering implementations mainly follow two approaches: attention mechanism-based methods and prefix tree (Trie)-based decoding-time intervention.
Attention mechanism-based biasing methods are flexible but have computational costs that grow linearly with vocabulary size.
All Facts (2)
Context biasing engineering implementations mainly follow two approaches: attention mechanism-based methods and prefix tree (Trie)-based decoding-time intervention.
50%UnverifiedAttention mechanism-based biasing methods are flexible but have computational costs that grow linearly with vocabulary size.
50%