Instrument Recognition Based Wavelet Packet Trees in Audio Feature Extraction

Feature extraction from audio data is a major concern in computer assisted music applications and content based audio retrieval. For general non-stationary signals, wavelet packet decomposition is used with entropy functions for best basis search. Musical instruments have well defined frequency ranges. Thus when audio data containing a solo instrument is concerned, wavelet packet decomposition may be adapted to that instrument’s individual characteristics. The method discussed in this paper uses a number of wavelet packet coefficients’ averages as feature vectors in a multi-layer perceptron (MLP) network to classify the type of the instrument and form an instrument specific wavelet packet tree for extraction of further features (e.g. pitch, duration, brightness etc.) The results presented for flute give accurate estimates of pitch and duration.