Error Propagation
In streaming speech recognition, a phenomenon in which already emitted tokens are treated as conditional context by subsequent decoding, causing early errors to be reinforced by the language model and to accumulate and expand.
Timeline (last 90 days)
In streaming ASR, already-emitted tokens are treated as conditioning context by subsequent decoding steps, and errors may be solidified and propagated by the language model.
Error propagation is a typical problem in cascaded systems, where errors from preceding modules propagate to and are amplified in subsequent modules, whereas detecting code-switching directly from speech signals can fundamentally cut off this propagation chain.
All Facts (2)
In streaming ASR, already-emitted tokens are treated as conditioning context by subsequent decoding steps, and errors may be solidified and propagated by the language model.
50%UnverifiedError propagation is a typical problem in cascaded systems, where errors from preceding modules propagate to and are amplified in subsequent modules, whereas detecting code-switching directly from speech signals can fundamentally cut off this propagation chain.
50%