Reducing the search complexity for low bit rate vector quantization based on shells of Golay codes

We study in this paper ways to reduce the complexity required by a vector quantization scheme, which uses the shells of Golay codes for coding sub-vectors of the input vector. Such a scheme was recently shown to achieve top performance in terms of segmental SNR in the generic situation of data affected by outliers, which is relevant for the case of audio data in various sub-layers of transform based speech and audio codecs. The optimization of bit allocation and shell selection process can be simplified into a greedy algorithm with a very slight loss of performance over the exact optimal solution. We study here various algorithmic solutions, able to keep a good performance, while drastically reducing the complexity of the previously proposed full search algorithm.