Page  126 ï~~Durham Music Technology: Activity Report Takebumi ITAGAKI #+ Simon JOHNSON + Peter D. MANNING # Douglas J.E. NUNN + Desmond K. PHILLIPS + Â~ Alan PURVIS + Jonathan R. SPANIER + Durham Music Technology VVWW homepage URL=http: // # Department of Music, University of Durham Palace Green, DURHAM DHL 3RL, UK + School of Engineering, University of Durham South Road, DURHAM DH1 3LE, UK Abstract Durham Music Technology is an ongoing collaboration between the Concurrent Digital Signal Processing group in the School of Engineering and the Electroacoustic Music Studio in the Department of Music at Durham University. In this activity report, we present current research interests in multi-processor architectures for real-time synthesis, VLSI implementation of synthesis algorithms, sound analysis-resynthesis, and works in the electroacoustic studio. 1. Optimised Synthesis Algorithm for VLSI Implementation This part of the group is concerned with optimising synthesis algorithms for VLSI implementation. This work is performed in collaboration with the VLSI Group in the School of Engineering at Durham University. We have recently built an oscillator chip and are currently researching multirate additive synthesis and Forme d'Onde Formantique (FOF) synthesis engines for CMOS implementation. Multirate Additive Synthesis (MAS) reduces the computation of additive synthesis via sinusoidal oscillator banks by exploiting the frequency stationarities of partials in note-based music. Near-optimal sample rates allocated to oscillators are much lower than industry standard rates (e.g. 44.1 kHz) thus reducing control and oscillator update bandwidth. An overhead is that interpolation via multirate filter banks is required [Phillips et al. 1994]. Its partial implementation onto a multi-processor network has been reported [Itagaki et al. 1995] and advanced issues on a MAS algorithm and coprocessor architecture within practical constraints of implementation (e.g. functional transparency, net cost of computation, latency) will be presented [Phillips et al. 1996]. A sinusoidal generator was implemented in a 0.7 -im double metal CMOS process. This chip generates a sinusoidal waveform by a phase accumulator linked to a sine/cosine generator based on the COrdinate Rotation DIgital Computer (CORDIC) algorithm. The table lookup approach to sine generation has large storage requirements; the CORDIC approach uses just 21 words. It should be noted that the algorithm is iterative (requiring 21 iterations), but it is possible to un-fold the algorithm in space to improve performance through pipelining. Figure 1: CORDIC based Digital Sine Generator In the figure, the chip layout is shown and the die size is 2.5 mm x 2.5 mm and, uses 7000 transistors. The chip has been tested at 20 MHz, is capable of running at higher speeds and would outperform programmable hardware. Â~ Currently working at Loughborough University (UK). Itagaki et al. 126 ICMC Proceedings 1996

Page  127 ï~~The table below describes some of the pins, which allow the chip to be programmed via a microprocessor. Input Signals Output Signals 8-bit interface Sinusoid (in 16-bit) 2 select lines diagnostic signals load input clock Table 1:1/0 signals for Digital Sine Generator In this ongoing research, we are investigating an approach to implementing the FOF algorithm in silicon. The method provides a simple control mechanism to provide spectral morphing via the octavation parameter. The scheme proposed in the paper, to be presented in this year, uses an impulse generator running at a high sample rate followed by a decimator which feeds the impulse into a parallel filter bank. The filter bank creates the formant structure in the base-band and the result is then heterodyned at the formant frequency. The outputs are then summed and this constitutes one "voice". The optimisation provides a means of time-division multiplexing the arithmetic components to provide a polyphonic synthesis engine [Spanier et al. 1996]. The current status of this research is mapping the architecture to a TMS320C40 DSP chip to test the overall scheme, and then onto VLSI design. 2. Multi-processor Network Since 1988, this research group has reported on issues concerning multi-processor based system [Purvis et al. 1988]. A prototype architecture for a transputer network was demonstrated in the 1990 ICMC [Bailey et al. 1990a]. This subsequently was developed into an audio processor using 160 T800 transputers as a ternary tree, as shown in Figure 2. The network has been used as a test-bed for network architecture for real-time additive synthesis [Itagaki et al. 1994]. Meanwhile, we have been implementing real-time granular synthesis using a part of the network [Itagaki et al. 1996]. The development of even faster processors has made at first glance the overall performance of the 160 transputer system seem outdated. However we would emphasise again that though the computing power represented by our network is broadly equivalent to about 10 Texas Instrument C40 processors the main issue that has concerned us is the optimal use of processors and the effective use of the high communication bandwidth. Advances in the way microprocessors communicate with each other have, however, also seen significant improvements and it is with this in mind that we have begun to construct a multiprocessor SGS T9000 system. The T9000 architecture represents the next generation transputer and by the use of 6 serial communication links having 10 times the performance of the T800 links and onboard processing to make a physical link equivalent to multiple virtual links to communicating processes we have the aim of extending our research on optimising the communication architecture within a network of real-time audio generators. The processing performance will also be scaled by an order of magnitude though we are giving thought in our design as to how an individual processor's computational performance might be assisted by a local P6 network. Again our concern is to develop the communication strategy by increasing demand upon the real-time output until the processing power is communication bound despite our best attempts to develop improved configurations. Work is also continuing on real time polyphonic pitch recognition with the aim of producing practical audio to score converters. 3. Music Analysis and Resynthesis This research has focused on two methods that offer the potential of allowing analysis and resynthesis. The first system was designed for polyphonic transcription and uses additive synthesis as the resynthesis engine [Nunn et al. 1994]. The initial analysis, implemented on the Texas Instruments TMS320C40, uses multirate STFTs in order to satisfy the trade-off between time and frequency resolution. Several further processing stages on a standalone PC attempt to estimate the notes and timbral envelopes. The results are displayed graphically, and real-time animations of the sound Figure 2: 160 Transputer Network System ICMC Proceedings 1996 127 Itagaki et al.

Page  128 ï~~can also be produced [Nunn et al. 1995]. The accuracy of the transcription is strongly dependent on the polyphony, timbres, and musical structure of the source material. Later research has suggested that Gabor wavelets [Nunn et al. 1996] are wellsuited to analysis, transformation, and synthesis, and this has developed into both a procedural language and a graphical user interface for music composition. 4. Composition of Electroacoustic Music The electroacoustic music studios, which complement our audio signal processing research laboratories, were extensively refurbished and upgraded in 1993-4 to support a comprehensive range of compositional activities. Although these studios are now predominantly based on digital equipment we have retained a significant number of analogue devices in one of the studios to ensure that composers can experience a truly comprehensive range of electroacoustic resources and techniques, many of which in our view are not yet satisfactorily represented in terms of digital equivalents. On the purely digital side our parallel signal processing platform based on just three T800 transputers and a PC host, first demonstrated at the 1990 ICMC using our in-house implementation of parallel-CSOUND [Bailey et al. 1990b], continues to service a number of compositional requirements, although this particular resource is now largely outpaced by our new Macintosh Power PC platforms which now support the bulk of our synthesis and signal processing work specifically in the field of music production and performance. We have developed a particular interest in granular synthesis (an aspect of our work which is reported upon elsewhere in this Proceedings [Itagaki et al. 1996]), building upon research originally carried out at Simon Fraser University in 1994 using the GSAMX facility developed by Barry Truax. A number of works have been completed using granular methods including The Ghost of Eriboll (Peter D. Manning) which is being performed at the 1996 ICMC. References [Bailey et al. 1990a] Bailey, N., Bowler, I., Purvis, A. and Manning, P.D. 1990. "An Highly Parallel Architecture for Realtime Music Synthesis and Digital Signal Processing" In Proceedings of ICMC 1990, Glasgow, UK, pp. 169-171 [Bailey et al. 1990b] Bailey, N., Bowler, I., Purvis, A. and Manning, P.D. "Concurrent CSOUND Parallel Execution for High Speed Direct Synthesis." In Proceedings ofICMC 1990, Glasgow, UK, pp. 46-49 [Itagaki et al. 1994] Itagaki, T., Purvis, A. and Manning, P.D. "Real-time Synthesis on a Multi-processor Network." In Proceedings of ICMC 1994, Arhus, DENMARK, pp. 382-385 [Itagaki et al. 1995] Itagaki, T., Phillips, D.K., Manning P.D. and Purvis, A. "An Implementation of Optimised Methods for Real-time Sound Synthesis on a Multiprocessor Network." In Book of Abstracts Parallel Computing 1995, Gent, BELGIUM, p. 100 [Itagaki et al. 1996] Itagaki, T., Manning P.D. and Purvis, A. "Real-time Granular Synthesis on a Distributed Multi-processor Platform." In Proceedings of ICMC 1996, HONG KONG [Nunn et al. 19941 Nunn, D.J.E., Purvis, A. and Manning P.D., "Source separation and transcription of polyphonic music." In Proceedings of the International Colloquium on New Music Research, Gent, BELGIUM, 1994 [Nunn et al. 1995] Nunn, D.J.E., Purvis A. and Manning P.D. "Graphical display of musical information." In Proceedings of XI Colloquio di Informatica Musicale, Bologna, ITALY [Nunn et al. 1996] Nunn, D.J.E., Purvis A. and Manning P.D., "Acoustic quanta." In Proceedings of ICMC 1996, HONG KONG [Phillips et al. 1994] Phillips, D.K., Purvis, A. and Johnson, S. "A Multirate Optimisation for Real-Time Additive Synthesis." In Proceedings of ICMC 1994, Arhus, DENMARK, pp. 364-367 [Phillips et al. 1996] Phillips, D.K., Purvis, A. Johnson, S. "Multirate Additive Synthesis." Proceedings of ICMC 1996, HONG KONG and In [Purvis et al. 1988] Purvis, A., Berry, R. and Manning P.D. "A Multi-transputer Based Audio Computer with MIDI and Analogue Interfaces." presented at Euromicro 1988, Zurich, SWITZERLAND, published in Microprocessing and Microcomputing 25 (1989): pp. 271-276 [Spanier et al. 1996] Spanier, J.R., Johnson, S. and Purvis, A. "Optimisations of the FOF Algorithm for VLSI Implementation." In Proceedings of ICMC 1996, HONG KONG Itagaki et al. 128 ICMC Proceedings 1996