Reconstructing Correct Utterance Candidates from Frame-Level CTC Evidence in Dysarthric Speech Recognition

Dysarthric speech recognition may produce an incorrect final transcription even when the recognition model retains information from which the correct utterance can be constructed. In the preceding Study 02, we examined 22 substantive recognition errors among 30 utterances and identified eight cases in which the reference token sequence was constructible from frame-level Connectionist Temporal Classification (CTC) evidence. In the present study, we conducted an exploratory evaluation of candidate reconstruction for these eight utterances, distinguishing the representability of the reference sequence from its actual generation by a reference-free search procedure. First, a reference-guided CTC path diagnostic confirmed a valid path to the reference token sequence in all eight cases. We then applied CTC Prefix Beam Search without providing the reference utterance during candidate generation. With a beam width of 10,000, a frame-level top-K of 20, and a probability threshold of 0.001, the reference token sequence was generated for two of the eight utterances (25.0%). Analysis of the search process identified cases in which prefixes leading to the reference sequence were generated but subsequently removed by beam pruning. An exploratory length-based adjustment to prefix retention did not improve the reference generation rate. These findings demonstrate a distinction between the constructibility of the correct sequence from internal CTC evidence and its successful generation by a constrained search procedure. They highlight the importance of candidate construction strategies that preserve relevant acoustic evidence during post-recognition error recovery.