Efficient Recognition of Vulgar Words in End-to-End ASR

The accurate identification of infrequent profane expressions presents a persistent challenge for end-to-end automatic speech recognition (ASR) systems. While existing techniques for rare term processing have enhanced general rareword detection, their efficacy diminishes significantly for infrequent vulgar terms, particularly those that are lengthy and exhibit low occurrence rates. Furthermore, research frequently neglects the mismatch rate associated with these terms during search algorithm execution. This work introduces an efficient methodology leveraging the Aho-Corasick algorithm. Our approach elevates ASR efficiency and decoding precision concerning vulgar vocabulary by minimizing matching failures and applying length-dependent weighting coefficients to infrequent profane words, thereby mitigating erroneous insertions. Integrating this approach with internal language model estimation yields substantial gains in the detection rate of rare vulgar expressions within specialized domains, as well as their recognition accuracy in cross-domain speech utterances. Empirical evaluations validate the effectiveness of the proposed technique in augmenting the recognition capabilities for rare profane words within fully neural ASR systems.