N-best List
An output form in which an ASR system retains the N most probable candidate transcriptions during decoding (typically N between 5 and 10). It preserves the uncertainty information from acoustic and language model decoding and is used for rescoring or generative post-editing.
Timeline (last 90 days)
The N-best list is the standard output form of ASR systems, where the recognizer retains the N candidate transcriptions with the highest probabilities during decoding, with N typically ranging from 5 to 10.
The N-best list is not only an input at inference time but also a source of supervision during training, further amplifying its value.
All Facts (2)
The N-best list is the standard output form of ASR systems, where the recognizer retains the N candidate transcriptions with the highest probabilities during decoding, with N typically ranging from 5 to 10.
50%UnverifiedThe N-best list is not only an input at inference time but also a source of supervision during training, further amplifying its value.
50%