Donnerstag, 12.11.2026 16:00-17:30 Uhr

öffentliche Veranstaltung in englischer Sprache

Everybody over the age of 18 has less than perfect hearing. We wish to model these imperfections or impairments, from the cochlea to the cortex, to form what can be called a digital twin of the auditory brain. Then given this personalized auditory model we design a deep neural network (DNN) to enhance incoming audio so that the output of the model matches a normal auditory system, or maximizes performance on metrics like the recognition of speech in noise.

Historically, auditory models have endeavored to explain the average or normal ear. But now with big-data approaches and machine-learning techniques we can do better. The key to this process is to represent the auditory model, the enhancer, and the optimization criteria (or loss) as differentiable computational graphs. In the first step we drive the parameters of the auditory model to match the results of psychoacoustic tests on the human subject. In the second phase we design a speech enhancer to modify the dynamics of audio, both in time and frequency, so that when played through the impaired auditory model we minimize the loss. Modern AI has taught us that these high parameter count, highly non-linear models can be fruitfully optimized to minimize an arbitrary (but differentiable) loss function.

Dr. Slaney will describe progress in all three areas. An important criteria is balancing explainability and biophysical realism versus black-box solutions.

Veranstalter: Universität zu Lübeck

Art der Veranstaltung: Vortrag in englischer Sprache

Referent: Malcolm Slaney, Stanford University (CCRMA), International Computer Science Institute (Berkeley), University of Washington (EE)

Zielgruppe: Fachpublikum

Ort: Audimax, AM4

Anmeldung: nicht notwendig