Page  372 ï~~The Timbral Object - An Alternative Route to the Control of Timbre Space David P. Creasey, David M. Howard, Andrew M. Tyrrell Department of Electronics, University of York, Heslington, York YOI 5DD, United Kingdom. dpc]0,dh, amt@ohm.york. ac. uk Abstract This paper outlines a new approach to timbre control based on movement between known reference points in timbre space which is being investigated at the University of York. Such points are specified through analysis of sounds. Transitions between timbres are achieved through "timbral objects" which describe manipulable vectors in timbre space. These are used in instrument definition and performance. 1. Problem alter timbres dynamically. The ultimate synthesis control system is one which can bridge the divide between a timbral concept in a user's mind and the required sonic result. Controlling the parameters of a synthesis model directly is too complex and unnatural in most circumstances. At the University of York we are investigating some concepts which may produce a more appropriate control scheme by the use of timbral objects (manipulable and manipulating transition vectors) as the basis on which new sounds are constructed. That is, rather than specifying a timbre in terms of arbitrary engineering parameters, one might describe it in terms of manipulations (movements in timbre space) from an initial sound colour. Similarly, these timbral transitions might be used as performance dimensions which can be applied in real time, to add expressive qualities. An example of a transition is the effect of changing from blowing gently to blowing forcefully on a clarinet. Such sonic entities are hard to describe in words but represent important control changes which are desirable to capture and use. Concepts such as the application of a "harder blown" timbral object to a synthesiser pad sound are conceivable control changes in a system based on captured timbral change. These are more tangible entities than traditional engineering parameters, whose association with the sound medium is vague. It is possible to imagine choosing from a large collection of recognisable objects to create and To achieve this, it is necessary to be able to capture, scale, combine and translate timbral changes, while still maintaining the identity of the variation. Capture is a process of determining the difference between the underlying model states. Scaling is identifying how to interpolate between, and extrapolate beyond, those states in a realistic manner. Combination is establishing the way that the parts of the representation are blended realistically from two sources. Translation is the hardest problem, because a variation to the model parameters is unlikely to have an equal perceptual effect when starting from different sonic conditions. (a).. -.....(b) - - t ------.............." ': I: i ' Figure 1; Vectors in Timbre Space Creasey et al. 372 ICMC Proceedings 1996

Page  373 ï~~Figure 1 shows simple representations of vectors in timbre space. Point S represents silence, and the sides of the cube represent some of the dimensions in the space. 1(a) depicts the definition of a timbre. 1(b) shows how timbres might be related to each other, both within a small area (1--1b, which represents performance characteristics of a single instrument) or between very different sounds (1--2, 1-3). 1(c) shows how the vectors of the previous group can be translated to provide new performance vectors for a sound, which itself might be translated through timbre space to a different location. 2. Technique Firstly, the sounds which allow navigation from point to point in timbre space must be converted into a representation which supports timbral analysis and modification. Different models have different problems. Spectral models are established in this area, such as Serra and Smith's technique (Serra/Smith, 1990). The lack of association with the processes of sound creation in the spectral form might indicate the use of a physical model instead, but there are problems in modelling any desired sound as easily as with a spectral approach. Any generalised model description would be appropriate, assuming that it provides a specification technique that is not impossibly complex when one is attempting to achieve a particular timbre. For example, a set of time domain samples achieves generalised control. But, more practical methods, such as additive synthesis, granular synthesis, waveshaping and FM are more suitable. Our technique of spectral decomposition translates an arbitrary time waveform into a distributed partially-meaningful data form. We are using a multi-resolution DFT analysis technique. Through consideration of the stability (weighted periodicity metrics) of the resulting bin outputs it is possible to choose the most appropriate level of detail for a particular frame and position in frequency. As such, the time-frequency resolution is optimised for the timbre being considered. The intention being, to allow more effective extraction of the timbral features of the sound. The next step is to transform the representation into the dimensions of timbral control. The dimensions are intended to be a comprehensive set of non-orthogonal axes which are of direct relevance to the investigation of timbre. These include harmonics, formants, attack rate, noise content, spectral roll-off, and number of partials. An appropriate set for timbral control has been sought by many authors (Grey, 1975 and Langmead, 1995 being good examples). Part of this research is to test previous authors' results and the universality of a composite set. It is vital to use this transformed set of dimensions. Otherwise, the ability to specify and analyse timbral changes resulting from a modification is reduced. For example, changes to and results affecting formants are more meaningful in investigation, compared to changes to the underlying individual partials whos amplitudes compose the formant. The control set can therefore be reduced from a large number of linearly-spaced time-varying amplitude and frequency partials, to a set with timbral control significance. If the list of features is comprehensive enough, then one should be able to classify any timbral change. Attempts to find timbrally important classifiable changes within the spectral model have often concerned particular verbal descriptions of timbre. These include roughness, hardness, openness clarity and warmth. These provide an excellent starting point for testing the viability of the composite set, but are also restrictive in that they do not cover aspects which cannot be described so simply. It is important to use a range of sounds which effectively covers the timbre sub-space of interest, as well. For example, Grey concentrated on short orchestral instrument notes. It is then possible to investigate the interactions within the representation. Determining the relationship between the balance of timbral aspects and the alteration of the timbral features is necessary. For example, how attack rate affects the perception of change to harmonic amplitude roll-off. This is achieved by applying a large group of perceptually significant timbral changes in the dimension set to a representative set of sounds, and judging the amount of the expected (and other) effects in each case. Then the perceived changes due to a similar physical change ICMC Proceedings 1996 373 Creasey et al.

Page  374 ï~~can be correlated with the timbral attributes of the set of sounds. It is then possible to compensate for the cross-coupling effects, such that any object can be applied to any sound (that is, translated through timbre space) with equal perceptual effect. This means that timbral entities are no longer restricted to the immediate timbral sub-space. An object can be applied to any sound, rather than requiring the synthesis of an effect from raw synthesis parameters. For example, the timbral change from a smooth trumpet to a rasping one might be captured, and then applied as a performance dimension to a piano sound. Alternatively, such a change might be extrapolated further to achieve a new instrument type. These changes require the use of scaling and combination to enable navigation of the chosen timbral domain, hopefully in a more intuitive and expressive fashion than has previously been achieved. 3. Usage The process of creating this powerful control form will result in a greater understanding of the way in which timbre is perceived. This may lead to better models of timbre and of the processes of the ear. On another level, capturing performance dimensions which may be applied in any situation means the ability to examine such concepts as the existence of fundamental orthogonal dimensions of timbre. These are difficult to prove without the ability to translate, scale and combine timbral changes. That is, to navigate any postulated timbre space. The most interesting musical application is in creating instruments. The process of timbral change can be viewed as a translation within a performance region (a timbral sub-space) where the fundamental identity of an instrument is maintained, or between instruments beyond that sub-space. In electroacoustic music the difference between the two is not as clear as in older musical disciplines. Applying timbral objects is not, however, restricted to a particular musical style. The ultimate challenge is in integrating the ability to apply performance dimensions equally to any timbre with the real-time physical control to maximise it. The final part of this research will consider how to apply timbral objects in real-time. 4. Conclusion This paper has discussed a fundamental problem in creating timbres; that of mapping a mental sonic image, through a synthesis model, to the desired sound. One of our solutions concerns the specification of timbral changes with manipulable and manipulating entities called timbral objects. These are based upon translations between known reference points (analysed sounds). The aim is to build a representation of timbre space control through combination, translation and scaling of the objects with a view to learning more about the control of timbre. 5. References Grey, J.M. (1975), An Exploration of Musical Timbre using Computer-Based Techniques for Analysis, Synthesis and Perceptual Scaling: PhD Thesis, Stanford University, California. Serra, X, Smith, J. (1990), Spectral Modeling Synthesis: A Sound Analysis/Synthesis System Based on a Deterministic plus Stochastic Decomposition, Computer Music Journal, Vol.14, No.4, p12-24. Langmead, C.J. (1995), A Theoretical Model of Timbre Perception Based on Morphological Representations of Time-Varying Spectra: MA Thesis, Dartmouth College, New Hampshire. 6. Acknowledgements Many thanks to Andy Hunt for fruitful discussion of these topics. This work is supported by EPSRC grant 94315368. Creasey et al. 374 ICMC Proceedings 1996