Search

This paper analysis the behavior of forward and backward-based decoders used for speech transcription. Experiments have showed that backward-based decoding leads to similar recognition performance as forward-based decoding, which is consistent with the fact that both systems handle similar information through the acoustic, lexical and language models. However, because of heuristics, search algorithms used in decoding explore only a limited portion of the search space. As forward-based and backward-based approaches do not process the speech signal in the same temporal way, they explore different portions of the search space; leading to complementary systems that can be efficiently combined using the ROVER approach. The speech transcription results achieved by combining forward-based and backward-based systems are significantly better than the results obtained by combining the same amount of forward-only or backward-only systems. This confirms the complementary of the forward and backward approaches and thus the usefulness of their combination.