LiCA: Text-Independent Linguistic Content Alignment for End-to-End Mandarin Dysarthric Voice Conversion

Dysarthric voice conversion (DVC) aims to transform dysarthric speech into a more natural and intelligible form. Existing DVC methods typically rely on complex auxiliary strategies or text supervision to recover linguistic content, and their performance on Mandarin remains limited. The tonal nature of Mandarin, along with its rich prosody, homophones, and polyphones, further increases the difficulty of modeling dysarthric speech. To address these challenges, we propose the Text-independent Linguistic Content Alignment Model (LiCA) for Mandarin DVC. Specifically, LiCA aligns dysarthric and healthy Whisper encoder representations through a Linguistic Content Alignment Module, and synthesizes natural speech using a Speech Generation Module. A Duration Predictor is further incorporated to model duration patterns from the posterior distribution and guide natural prosody generation. Moreover, a Syllable-rate Consistency Loss is introduced to enhance temporal stability and prosodic coherence. Extensive experiments on the CDSD and MDSC Mandarin corpora and the UASpeech English corpus demonstrate that LiCA substantially improves intelligibility. In speaker-dependent settings, LiCA reduces the relative CER by 54.7% and increases MOS from 2.75 to 4.27 on CDSD, while reducing the relative WER by 30.0% and increasing MOS from 2.24 to 4.17 on UASpeech. In addition, LiCA also achieves the best performance on 29 out of 32 objective acoustic metrics on CDSD. For speaker-independent evaluations using a character-length-based protocol, LiCA reduces the relative CER by 55.51%, increases MOS from 2.18 to 3.38 on MDSC, and shows particular effectiveness for short- and medium-length Mandarin dysarthric utterances.