Audio classification by search of primary components

This work addresses the soundtrack indexing of multimedia documents. Our purpose is to detect and locate sound unity to structure the audio dataflow in program broadcasts. We present three different audio classification tools that we have developed. The first one, a speech/music classification tool, is based on three original features: entropy modulation, stationary segment duration and number of stationary segments. It provides about 90 % of accuracy for speech and music detection. Another system, a jingle identification tool, uses an Euclidean distance in the spectral domain to index the audio data flow. Results show its efficiency: among 132 jingles to recognize, we have detected 130. The last one, a key sound classification tool, permits to extract applause and laughter. Results are perfect for applause (only few shift) and quite good for laughter (some missing). Systems are tested on TV and radio corpora (more than 12 hours)