Abstract:

Because of the centrality of pitch to music, most music signal processing has been focused on pitch-related timescales, e.g., short-time analysis over 30--50-ms windows. But there is, of course, important information at other scales as well. By looking at ways to model and extract very fine time-scale information one can extract attributes that lead to the perception of ``texture'' in unpitched instruments such as maracas. At the other extreme, interesting effects can be revealed by investigating the effect of ``modulation-domain'' processing, such as filtering subband energy envelopes in the 0.1--10-Hz range. A single representation, frequency-domain linear prediction, gives convenient access to both these scales: its use in music analysis/synthesis will be illustrated by examples.