high performance computing on graphics processing units: hgpu.org

We describe a new approach to the processing of audio by way of transformations to and from the frequency domain. In previous papers we described the Sliding Discrete Fourier Transform (SDFT), comprising an extension to the classic phase vocoder algorithm to perform a frame update every sample. We proposed this as offering musical advantages over the common STDFT, including lower latency and the potential for new classes of effect. Its major disadvantage has been the very high computational cost, which makes it intractable for real-time use even on high-specification consumer workstations. We report on a version of the SDFT that exploits the intrinsic parallelism of the scheme on a commodity GPU, to implement the transform and its inverse in real time for multiple audio channels. This implementation is used to provide previously uncomputable real-time effects. We describe a key new effect, enabled by the method, which we have called Transformational FM (TFM).