Persistence and Recoverability of Correct Information After Recognition Errors in Dysarthric Speech Recognition

Abstract

Automatic speech recognition (ASR) often produces incorrect 1-best transcriptions for dysarthric speech, but a recognition error does not necessarily imply that all information supporting the correct utterance has been lost. This study investigates the extent to which correct information remains within an ASR model after 1-best recognition errors, and whether that information can be recovered from internal hypotheses or frame-level acoustic evidence. We evaluated 30 predefined Japanese short utterances spoken by one speaker with dysarthria using a CTC-based Japanese ASR model. Among 22 substantive 1-best errors, the complete reference was found in deep-beam candidates for 5 utterances. Of the remaining 17 utterances, 8 were representable in a frame-level CTC candidate space under the main posterior threshold of 0.001. Thus, in 13 of 22 error cases (59.1%), either the complete reference or acoustic hypotheses sufficient to construct it remained internally under the evaluated conditions. Additional analysis showed that language-model reranking could substantially improve the rank of the reference in some cases, while in others the reference remained acoustically disfavored. Numerical utterances were difficult in this dataset, but the sample size was too small for generalization. These results suggest that post-recognition recovery strategies may be useful for dysarthric speech recognition when they explicitly distinguish between errors with recoverable internal evidence and those in which sufficient acoustic evidence cannot be confirmed.