ROBUST PARAMETERS FOR NOISY SPEECH RECOGNITION

摘要

The invention concerns a method for automatically processing noisy speech comprising the following steps: sensing and digitising speech (1); retrieving several frames (15) corresponding to said signal (10); breaking down each frame (15) using an analysing system (20, 40) into at least two different frequency bands to obtain two vectors of representative parameters (45); converting (50) the vectors (45) into second vectors relatively insensitive to noise (55), each converter system (50) being associated with a frequency band. The knowledge acquisition of said converter systems (50) being produced on a noise-contaminated speech corpus (102).