Improvements to the switched parametric and transform audio coder

Abstract

We introduce improvements to previous sines+transients+noise audio modeling systems, including new sinusoidal trajectory selection and quantization procedures. In a previous work by Levine and Smith (see Proc. Int. Conf. Acoustics, Speech, and Signal Processing, Phoenix, 1999), the audio is first segmented into transient and non-transient regions. The transient region is modeled using traditional transform coding techniques, while the non-transient regions are modeled using parametric sines plus noise modeling. Because such a system contains a mix of parametric and non-parametric techniques, compressed-domain processing such as time-scale modifications are possible.