Information Distance-Based Subvector Clustering for ASR Parameter Quantization

In this letter, we propose a new data-driven subvector clustering technique for parameter quantization in automatic speech recognition (ASR). Previous methods such as Greedy-n m-let and maximum clique partition have been proven to be effective. However, the former yields subvectors of equal sizes while the latter cannot determine the number of subvectors. In our method, we define a new object function based on information distance (ID) and optimize this using the cross-entropy (CE) method to overcome both of the aforementioned limitations. We compare the ASR performances using the Resource Management (RM) and Wall Street Journal (WSJ0) speech recognition tasks and show that the proposed technique performs better than previous heuristic techniques.