Cepstral features for classification of an impulse response with varying sample size dataset

Cepstrum-based features have proved useful in audio and speech characterisation. In this paper a feature vector of cepstral polynomial regression is introduced for the detection and classification of impulse responses. A recursive algorithm is proposed to compute the feature vector. This recursive formulation is appealing when used in a sequential learning framework. The discriminative power of these features to detect and isolate racket hits from the audio stream of a tennis video clip is discussed and compared with standard cepstrum-based features. Finally, a new formulation of the Average Normalised Modified Retrieval Rank (ANMRR) is proposed that exhibits relevant statistical properties for assessing the performance of a retrieval system.