A simplified computational technique for an LSF-based speech decoder

Line Spectral Frequencies (LSF) have been widely used to represent Linear Prediction Coefficients (LPC) due to their excellent quantization characteristics. Among others, MPEG4-compliant audio decoders include a module for converting the LSF to LPC coefficients. This paper presents a new simplified computational technique for carrying out the LSF-to-LPC conversion process. The proposed technique cuts down on several arithmetic operations by establishing direct links between the LPC coefficients and the two polynomials representing the two extreme glottis conditions. The technique is a good candidate for hardware implementation on mobile devices designed for low bit-rate wireless communication. Results of two testing scenarios are reported to show the computational distortion of the hardware conversion module relative to that resulting from quantizing the LSF vectors.