With ever-increasing number of car-mounted electric devices and theircomplexity, audio classification is increasingly important for the automotiveindustry as a fundamental tool for human-device interactions. Existingapproaches for audio classification, however, fall short as the unique anddynamic audio characteristics of in-vehicle environments are not appropriatelytaken into account. In this paper, we develop an audio classification systemthat classifies an audio stream into music, speech, speech+music, and noise,adaptably depending on driving environments including highway, local road,crowded city, and stopped vehicle. More than 420 minutes of audio dataincluding various genres of music, speech, speech+music, and noise arecollected from diverse driving environments. The results demonstrate that theproposed approach improves the average classification accuracy up to 166%, and64% for speech, and speech+music, respectively, compared with a non-adaptiveapproach in our experimental settings.
展开▼
