Exploring Dominant Paths in CTC-Like ASR Models: Unraveling the Effectiveness of Viterbi Decoding

Connectionist Temporal Classification (CTC) has emerged as a fundamental technique in Automatic Speech Recognition (ASR), renowned for its ability to marginalise all possible alignments between input and target sequences. This study reevaluates the traditional dependency on prefix beam search, which typically considers multiple alignments for each hypothesis, in CTC-based models. Our findings initially indicate the absence of a single dominant path in the lattices. However, we discover that the Viterbi decoder can effectively emulate the performance of a prefix beam search, as the hypotheses it identifies tend to be dominant in the lattices. This leads us to propose that an effective CTC-like model should not only aim for high accuracy but also align the optimal hypothesis with the most probable path, thereby enhancing the robustness of Viterbi decoding. Furthermore, our insights extend to a variety of topologies, demonstrating their applicability within a more comprehensive ASR framework.