Dynamic adaptation of speech comprehension systems to acoustic environments

摘要

Techniques are provided for dynamically adapting speech understanding systems to acoustic environments. One methodology implementing the techniques according to one embodiment includes generating a trigger in response to detecting a "Wa-ke-on-voice" key phrase in or in front of an audio stream. The trigger is for switching the processing modes from an adaptation mode to an interrogation detection mode. The method further includes performing automatic speech recognition on the audio stream during the interrogation mode to detect a domain-internal query. The method further comprises applying both a static speech understanding classifier and a dynamic speech understanding classifier to the recognized intra-domain query. The static speech understanding classifier uses a static semantic model, and the dynamic speech understanding classifier uses a dynamic semantic model. The method further includes determining an intention associated with the detected domain internal query based on results of the static speech understanding classifier and results of the dynamic speech understanding classifier.