A pattern classification proposal for object-oriented audio coding in MPEG-4

The future MPEG-4 standard will adopt an object-oriented encoding strategy whereby an audio source is encoded at a very low bit-rate by adapting a suitable coding scheme to the local characteristics of the signal. One of the most delicate issues in this approach is that the overall performance of the audio encoder greatly depends on the accuracy with which the input signal is classified. This paper shows that the difficult problem of audio classification for object-oriented coding can be effectively solved by selecting a salient set of acoustic parameters and adopting a fuzzy model for each audio object, obtained by a soft computing-hybrid learning tool. The audio classifier proposed operates at two levels: recognition of the class to which the input signal belongs (talkspurt, music, noise, signaling tones) and then recognition of the subclass to which it belongs. The results obtained show that fuzzy logic is a valid alternative to the matching techniques of a traditional pattern recognition approach.