The time–frequency representation (TFR) is the initial stage of analysis in sound/music analysis–resynthesis (A–R) systems. Given a time-domain waveform, the TFR makes temporal and spectral detail available to the remainder of the analysis, so that the component features may be extracted. The resulting ‘feature set’ must represent the sound as completely as the original time-domain signal, if the A–R system is to be capable of effective transformation and good synthesis sound quality. Therefore the system as a whole is reliant upon the TFR to make the sound components detectable, separable and measurable. Yet the standard TFR to-date is the short-time Fourier transform (STFT), of which the shortcomings, in terms of resolution, are well recognised. The purpose of this paper is to demonstrate the importance of the TFR to system function and system design. Poor feature extraction is shown to result from the use of inappropriate TFRs, whose underlying assumptions and expectations do not match those of the system. Existing models are used as case studies, with examples of performance for different sound types. A philosophy for A–R system design that includes TFR design is presented and a methodology for implementing it is proposed.