A Preliminary Study on Constructing Mandarin Personalized Speech Recognition Systems for the Speech Impaired

This study explores personalized Mandarin ASR for the speech impaired. Sentence and short phrase tasks selected from the VoiceBank-2023 corpus are examined for amyotrophic lat-eral sclerosis and esophageal-speech speakers. Unadapted speaker-independent ASR model is trained with the NER-Pro corpus to serve as a pretrained model for the follow-up speaker-independent and dependent models fine-tuned with the CDSD and VoiceBank-2023 corpora. Several fine-tuning strategies are investigated. Experimental results show that the adapted speaker-dependent models attain average character error rates (CERs) ranging from 6.8% (normal) to 11.9% (mod-erate dysarthria) and 5.9% (normal) to 12.9% (moderate) for sentence and short phrase tasks, respectively. Last, we have implemented the fine-tuning strategies on a large corpus of a hereditary spastic paraplegia patient to evaluate frequently used sentence tasks under open-set and close-set conditions, resulting in CERs of 10.8% and 3.8%, respectively.