Google, IBM & Meta Certificates – 40% Off
One plan covers every Professional Certificate on Coursera.
Unlock All Certificates
Explore the challenges and solutions in large vocabulary speech recognition through this lecture on bipartite graph factorization in static decoding graphs with long-span acoustic context. Delve into two standard approaches for searching the highest likelihood word sequence: on-demand construction and ahead-of-time full representation of the search space. Examine the problem arising from long-span acoustic context in full representation, focusing on bipartite graphs with O(V^2) edges in large vocabulary systems. Learn about the edge-wise minimal representation technique, involving the identification of complete bipartite sub-graphs and their replacement with extra vertices and connecting edges. Understand the NP-hard nature of finding the smallest representation and explore a heuristic approach for practical solutions. Gain insights from experimental results on a large-vocabulary speech recognition system and discuss related problems in the field.
