Adapting entropy constrained coding of spectral envelope for fixed-rate coding in AMR speech codec

The Adaptive Multirate (AMR) speech codec operates in 8 different fixed-rate modes. In every mode, it uses a specified number of bits to quantize and encode the current speech frame. It encodes the spectral envelope by means of Line Spectral Frequencies (LSF) parameters, by using the constrained resolution (CR) fixed-rate Split Matrix Quantization (SMQ) and Split Vector Quantization (SVQ) methods. However, by using the Gaussian Mixture Model (GMM) based transform coding technique, the quantization of the spectral envelope can be significantly improved in the spectral distortion (SD) sense. This technique involves adaptive decorrelation of the LSF vectors by applying an orthogonal linear transformation combined with an ordinary scalar quantization of the decorrelated vector´s components. We apply uniform scalar quantization followed by entropy constrained (EC) coding as it appears to be generally more efficient than non-uniform scalar quantizers. This paper describes the techniques used to adapt the entropy coded variable bit-rate output bit strings to the AMR codec modes using fixed-rate output bit strings. In order to constrain the length of the code, we use variable quantization step size and vector truncation techniques. Their application aspects are thoroughly investigated and described in this paper.