On an efficient VLSI architecture for the multirate additive synthesis of musical tones

摘要

Additive Synthesis (AS) is a general method for the synthesis of pitched musical tones which are modelled in frequency-domain as a sum of Q sinusoidal partials, with time-varying amplitude and frequency envelopes for each partial denoted as Aq[n] and Fj[n] where 1 < q < Q and n is the sample index at an industry-standard rate of e.g. /s = 44.1 kHz. A disadvantage which is often cited is the high bandwidth of computation. Fortunately, piecewise linear approximation of Ae/[n] and Fq[n] causes a dramatic reduction in the amount of data required to represent a spectral model of the tone. However, Atl[n] and Fq[n] still require uncompression for real-time AS at a rate of /s. Many proposals have been advanced to optimise computation and one such is Multirate Additive Synthesis (MAS); for the rationale see [1]. Using expected a priori bounds of [rnin(Fq[n]), max(/r,/[n])}, each partial q is allocated into the "best-fitting" sub-band of one of n QMF filterbank instantiations of depth K which results in near-critical sampling from a selection of K + 1 dyadically-related sample rates of fs2k where 0 < k < K. Our work indicates mat K = 3 is a good working value.