Fixed point vocoder implementation
L. Liu · 2002
Speech parameters of an analysis-by-synthesis vocoder are generated by complex mathematics such as Levinson-Durbin recursion, root solving method, etc. Fixed point mathematics cause loss in precision due to truncation errors. This paper presents a systematic approach to implementing a fixed point vocoder. The techniques used are speech parameter scaling, creating tables for special mathematical functions and normalization. Validation techniques include bit exactness test and perception test. The accuracy of the encoder output was found critical for synthesizing the speech at the decoder side. Finally, a key contribution to a highly compressed vocoder was to perform code optimization.