Optimized Finite State ASR Model for the Arabic Language

Abstract Since its introduction, the Weighted Finite-State Transducers (WFSTs) networks, with their precompiled search space in a single finite state, have provided an exceptional performance with an order of magnitude faster decoders for Automatic Speech Recognition systems compared to the conventional systems that construct the search space progressively. Using WFSTs for morphology-rich languages such as Arabic faces the challenge of the large vocabulary that results in very large networks that are hard to fit in the memory of common CPUs. In this work, we introduce several approaches that have been used to reduce the size of large Vocabulary Arabic WFSTs with a very minor impact on the system accuracy. We adopted the star architecture for the network topology and that reduced the network size and increased the decoding speed. Also, we adopted a two-pass decoding approach that utilizes a short history language model, and consequently smaller network size, in the first pass then in the second pass rescore the produced lattice with the longer history language model. We investigated several tuning parameters to reach the best compromise between network size and system accuracy. Applying these results on two test sets has shown consistent results of a 45% overall reduction in network size with only less than 1% degradation in accuracy. On the standard MGB3 benchmark, we managed to get 40x real-time Arabic ASR with an accuracy of 83.67% compared to 85.82% of state-of-the-art systems that have only 8x real-time performance on common commodity CPU processors.