shallow fusion
A technique that fuses scores from an external language model with those of the base model during ASR decoding, commonly used for contextual biasing.
Timeline (last 90 days)
Contextual biasing is a technique that introduces an external vocabulary or phrase list at the decoding stage to increase the recognition probability of specific word sequences by adjusting language model scores or using methods such as shallow fusion.
In speech recognition research, contextual biasing and prompt-based shallow fusion are existing concepts, the former injecting hotwords into the decoding graph and the latter encoding preceding text into vectors that participate in acoustic model scoring.
All Facts (2)
Contextual biasing is a technique that introduces an external vocabulary or phrase list at the decoding stage to increase the recognition probability of specific word sequences by adjusting language model scores or using methods such as shallow fusion.
50%UnverifiedIn speech recognition research, contextual biasing and prompt-based shallow fusion are existing concepts, the former injecting hotwords into the decoding graph and the latter encoding preceding text into vectors that participate in acoustic model scoring.
50%