Applicability of Shannon's Theorem to the Processing of Weakly Structured Speech Information in Acoustic Noise Conditions

The development of modern technologies for processing weakly structured speech information is directed towards improving large language models by changing the structure of information processing systems and improving the quality of training samples. The presence of acoustic noise, on the one hand, is a factor limiting the applicability of methods, on the other hand, various acoustic noises are superimposed on good speech data to increase the volume of training samples. It is obvious that an auditor needs a signal cleared of noise, but this is not required for automatic processing. The limitations of existing approaches require establishing more complex dependencies taking into account the weak structuring of speech signal information and the formation of a meta-description of the speech stream. Problems at various levels of the speech system hierarchy are considered and classified, and conclusions are made on the applicability of various methods and algorithms for processing speech information.