Audio Codec Index Compression Using Reduced Multidimensional Dictionaries

The patent segments the compression process into two independent stages: first, vector quantization converts audio vectors to indices; second, a separate neural network model compresses these indices. This segmentation allows each component to be optimized independently, avoiding the exponential complexity growth that occurs when trying to model long sequences in a single entropy coding stage.

The patent introduces a neural network model as an intermediary between the vector quantizer and the final encoded output. This neural network learns the statistical dependencies in the index sequences and performs the compression function, replacing the traditional entropy coding approach that requires large databases of statistics. The neural network acts as a mediator that captures sequence patterns without requiring exponential database growth.

The patent replaces the mechanical/statistical approach of building large databases from long sequences with a neural network-based approach. Instead of collecting and processing exponentially growing amounts of data to build statistics databases, the neural network learns compression patterns directly from training data, achieving similar or better compression efficiency without the exponential increase in data requirements.

The patent changes the fundamental parameter of how compression statistics are obtained. Rather than using long sequences to generate statistics databases through traditional entropy coding, the system uses a neural network model trained on sequences to learn compression patterns. This parameter change from statistical database building to neural network learning breaks the exponential relationship between sequence length and data requirements.