ENRICH

ICASSP2020 Demo: Neural Based Speech Enrichment

Mr. Muhammed Shifas PV
Speech Signal Processing Lab (SSPL)
University of Crete (UoC), Greece

Email: shifaspv@csd.uoc.gr


Abstract :- When listening in noise, speech intelligibility fades time-to-time. To ensure the effortless listening, one must modify the speech acoustic cues to be stronger under noise masking. Conventional signal processing approach (SigPro) has been widely used for the task, But it fails to mitigate the background noise effect at the recording end (far-end). To address this, We proposed the neural network approach (NeuralNet) that can boost perform the intelligibility modification of speech recorded even in inferior acoustic condition.


Few samples from the two models in quite listening condition

Noisy recording SigPro NeuralNet


Listen the same samples under speech shaped noise (SSN) at -7dB level

Noisy recording SigPro NeuralNet

Acknowledgment: This work was funded by the E.U. Horizon2020 Grant Agreement 675324, Marie Sklodowska-Curie Innovative Training Network, ENRICH.