Cocktail party data, a neural network, a subject test
For the study setup, ARTORG Center engineers used the “Bern Cocktail Party Dataset”, a collection of a variety of noise scenarios with multiple sound sources from multi-microphone recordings of hearing aid or cochlear implant users. Using 65 hours of audio recordings (more than 78,000 audio files), they trained a neural network to refine a commonly used directionality algorithm (beamformer). For improved speech understanding the deep learning approach calculated additional virtual microphone signals from the audio data mixture. 20 subjects tested the AI-enhanced hearing in a subjective hearing test accompanied by objective measurements. Particularly in cocktail party settings, the virtually sampled microphone signals significantly improved the speech quality. Hearing aid and cochlear implant users could therefore benefit from the presented approach, especially in noisy environments.
"I think that artificial intelligence represents an important contribution to the next generation of hearing prostheses, as it has great potential for improving speech understanding, especially in difficult listening situations," says Marco Caversaccio, Chief Physician and ENT Department Head. As auditory assistive technologies and implants are a major focus of research at the Inselspital, important data-based foundations are being laid here for further development that should bring the natural hearing experience closer. The novel approaches will directly benefit patients within the framework of translational studies.