Exploiting properties of the human auditory system and compressive sensing methods to increase noise robustness in ASR

摘要

Human speech comprehension is much more resilient against background noise than the most powerful automatic speech recognizers. That is due in no small part to the fact that conventional systems discard potentially relevant information upfront: the analysis of the audio signals. This is because classical theories do not know how to harness this information. We tried to develop techniques that approximate the way in which the human auditory system analyzes speech, and the way in which the brain processes the analysis results. as closely as possible. The results of the project are positive and not so positive. In the presence of loud background noise our approach is superior; it makes fewer errors, and the errors are more similar to human errors. In the absence of background noise our approach is somewhat less accurate, for reasons that we do not yet fully understand. However, the blame surely is not with our way of analyzing acoustic signals.