Mengzhe Geng

Experiments conducted on the UASpeech and TORGO corpora suggest that Conformer models trained using the DSI tokens outperform the comparable baseline HuBERT discrete/continuous features by statistically significant WER reductions.

An auditable decision record that carries four aligned cues into a late calibration step: a passive detector score, a conditional keyed-probe score on a marked derivative, retrieval support, and a speaker-profile margin, together with explicit disagreement coordinates is answered.

A 7B model can boost source-grounded speech planning accuracy from 68.4% to 91.9% by addressing citation legality before synthesis.

Watermarking TTS outputs without retraining or quality loss is now feasible, thanks to a novel method that exploits noise correlations.

Zero-shot adaptation can reduce word error rates by over 0.6% while achieving nearly 10 times faster real-time processing for elderly speech recognition.

Speaker adaptive training combined with confidence-guided pseudo-labeling leads to substantial gains in elderly speech recognition accuracy, outperforming conventional methods.

Personalized federated learning can reduce word error rates in dysarthric speech recognition by up to 0.99%, highlighting the potential for tailored ASR solutions.

Achieving up to 27.73% absolute WER reduction without any training or additional data challenges the conventional reliance on fine-tuning for model compression.