Researchers have developed a method to identify rooms using audio recordings that were captured in them. Nils Peters, a DAAD-funded postdoctoral fellow, and Speech Group researchers Gerald Friedland and Howard Lei used audio recordings from seven different rooms – a bedroom, library, studio, two churches, great hall, and classroom – to develop an acoustic profile for each. These profiles were based on audio features that are frequently used by speech systems to automatically recognize words.

The audio features of new unknown audio recordings, containing either speech or music, were then compared against these acoustic profiles. When there was no overlap between the training and test data, the system accurately identified the type of room 61 percent of the time when music was used and 85 percent when speech signals were used. By contrast, random chance would identify rooms with about 14 percent accuracy.

The work may help improve location estimation, which tries to automatically estimate the origins of videos, in situations where other techniques fail. GPS, for example, often fails indoors. Other applications include smart hearing aids able to automatically adapt to the acoustic environment.

The paper is related to ICSI’s ongoing research that attempts to use multiple modalities, such as audio, text, and visual cues, to understand multimedia content automatically.

Related Paper: “Name That Room: Room Identification Using Acoustic Features in a Recording.” Proceedings of ACM Multimedia 2012, Nara, Japan, October 2012. Available online at http://www.icsi.berkeley.edu/icsi/publication_details?n=3318 .