R&D - Page 2 of 2 - What’s next

Delivering non-intrusive signal intelligence with deep learning

Teaching machines to perceive the overall quality of speech and estimate the underlying acoustic characteristics of the environment in which the signal was captured without a clean reference signal (i.e. non intrusively) is a challenging task that has many real-world applications. These include audio forensics, hearing aids, speaker diarization and automatic speech recognition (ASR). In the context of ASR, an important application is for the purpose of data augmentation and selection, where an algorithm like Non-Intrusive Speech Analysis (NISA) can be used for analyzing field data to estimate the distributions of various acoustic parameters present and then to use those distributions to select appropriate data for ASR model training as well as for the purpose of configuring and automating data augmentation strategies.

Dushyant Sharma

Posted January 13, 2021