Answers are grounded in this paper and cite their sources.
Southern Min automatic speech recognition (ASR) is challenged by non-verbatim Mandarin subtitles, unreliable pseudo-transcripts, tone-related acoustic variation, and programme-dependent domain shift. We propose the Southern Min Reliability-Bridging Automatic Speech Recognition framework (MinBridge-ASR), organized around two core mechanisms: cross-scale reliability bridging and branch-selective tone–segment learning. The former combines speech-conditioned character-level refinement of weak subtitles with multi-cue utterance-level weighting of pseudo-transcripts while maintaining teacher–student consistency independently. The latter uses controlled pitch interventions to encourage complementary tone-related and segmental representations while restricting programme- and pseudo-speaker-adversarial learning to the segmental pathway. On MinSpeech, MinBridge-ASR achieved character error rates (CERs) of 17.21%, 22.96%, and 20.10% on the official, primary programme-disjoint, and human-verified evaluations, respectively. Across five programme-disjoint splits, it achieved a mean CER of 23.11 ± 0.44%. Direct lexical-tone evaluation showed that the tonal representation outperformed explicit fundamental-frequency (F0) fusion in tone-category and tonal-minimal-pair discrimination despite lower raw pitch recoverability. Progressive ablations and reliability calibration further supported the two core mechanisms. These results demonstrate improved weakly supervised Southern Min ASR under subtitle noise and programme variation.
No takes yet. Share an insight, caveat, or question.
Qiu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: