Algorithms and VLSI architectures for parametric additive synthesis

Abstract

A parametric additive synthesis approach to sound synthesis is advantageous as it can model sounds in a large scale manner, unlike the classical sinusoidal additive based synthesis paradigms. It is known that a large body of naturally occurring sounds are resonant in character and thus fit the concept well. This thesis is concerned with the computational optimisation of a super class of form ant synthesis which extends the sinusoidal parameters with a spread parameter known as band width. Here a modified formant algorithm is introduced which can be traced back to work done at IRCAM, Paris. When impulse driven, a filter based approach to modelling a formant limits the computational work-load. It is assumed that the filter's coefficients are fixed at initialisation, thus avoiding interpolation which can cause the filter to become chaotic. A filter which is more complex than a second order section is required. Temporal resolution of an impulse generator is achieved by using a two stage polyphase decimator which drives many filterbanks. Each filterbank describes one formant and is composed of sub-elements which allow variation of the formant’s parameters. A resource manager is discussed to overcome the possibility of all sub- banks operating in unison. All filterbanks for one voice are connected in series to the impulse generator and their outputs are summed and scaled accordingly. An explorative study of number systems for DSP algorithms and their architectures is investigated. I invented a new theoretical mechanism for multi-level logic based DSP. Its aims are to reduce the number of transistors and to increase their functionality. A review of synthesis algorithms and VLSI architectures are discussed in a case study between a filter based bit-serial and a CORDIC based sinusoidal generator. They are both of similar size, but the latter is always guaranteed to be stable.