The decomposition by ISA is shown to give physiologically intuitive features for instruments identification. The inclusion of the lesser energy components in classification can be beneficial for isolated tones but not so for the two-tonal mixtures. In real song, their roles are diminished. Only the sustain components will be learned and used in identification. A better learning algorithm is called for, for example, the non-negative matrix factorization which has been used for piano transcription in [10] might be investigated. Future works include more experiments on a larger database, pitch ranges and instruments, as well as its robustness to noise and percussive interference.
Table 2:
k-NN score as percentage of 7-nearest neighbors at 4 levels for instruments Piano (P), Flute (F), Bb-Clarinet (Bb), Cello (C), Alto-Saxophone (AS), French-Horn (FH), Oboe (O) and Violin (V).
