Dillier Büchler
Fakultäten » Medizinische Fakultät » Ohren-, Nasen-, Hals- und Gesichtschirurgie, Klinik für » Prof. Dr. Norbert Dillier » Dillier Büchler
Completed research project
| Title / Titel | Algorithms for Sound Classification in Hearing Instruments | ||||
|---|---|---|---|---|---|
| Abstract (PDF, 14 KB) | |||||
| Summary / Zusammenfassung | In this project, an automatic sound classification system for application in hearing instruments is developed. Our goal was on the one hand to develop a robust classification algorithm for at least the four classes 'speech', 'speech in noise', 'noise', and 'music', on the other hand to collect fundamental knowledge as a basis for a more detailed classification. The refined classes may for example be different noise types and music styles, or clean and reverberated speech. In order to study how the human auditory system classifies sound, the mechanisms of Auditory Scene Analysis were investigated. The extraction of auditory features was shown to be an important step in the process of sound segmentation performed by the human auditory system. Thus, one of the main goals in this thesis was to find appropriate features. A number of adequate auditory features have been modeled, including amplitude modulations, harmonicity, spectral profile, amplitude onsets, and rhythm. These auditory features were evaluated together with different pattern classifiers. Considering the application in hearing instruments, where computing time and memory are limited, simple classifiers (rule-based and minimum-distance classifiers) have been compared with more complex ones (Bayes classifier, neural network, hidden Markov model, and a multistage approach). A hit rate of about 80 % was achieved with the simpler classifiers, which could be increased up to some 90 % when a more complex classifier was used. However, both the computing time and memory requirements are about four times larger with the more complex than with the simpler approaches. Weitere Informationen |
||||
| Publications / Publikationen | Büchler, M. (2002). “Algorithms for Sound Classification in Hearing Instruments,” PhD Thesis at Swiss Federal Institute of Technology, Zurich, no 14498.Büchler, M., Dillier, N., Allegro, S., Launer, S. (2002). „Algorithmen für die Geräuschklassifizierung in Hörgeräten (Algorithms for sound classification in hearing instruments),“ Jahrestagung der Deutschen Gesellschaft für Audiologie, Zürich, SwitzerlandAllegro, S, Büchler, M., Launer, S. (2001). „Automatic Sound Classification Inspired by Auditory Scene Analysis,“ Eurospeech, Aalborg, Denmark.Weitere Informationen | ||||
| Keywords / Suchbegriffe | Neurosciences, sound classification, neural networks, hearing instruments, auditory scene analysis, hidden Markov models, speech and noise | ||||
| Project leadership and contacts / Projektleitung und Kontakte |
|
||||
| Funding source(s) / Unterstützt durch |
Private Sector (e.g. Industry) Phonak AG |
||||
| In collaboration with / In Zusammenarbeit mit |
|
||||
| Duration of Project / Projektdauer | Jul 1998 to Dec 2004 |