DCASE 2025 Workshop · Published
A state-space model detects anomalous machine sounds without anomaly labels; a pixel-unshuffle input cuts pre-training time by about 87.5%, and the system beats the DCASE 2025 autoencoder baseline with a harmonic-mean score of 0.722.
ASRU 2025 · Accepted
WhisQ predicts mean opinion scores for text-to-music systems from Whisper audio features and Qwen3 text features, linked by sequence co-attention and optimal-transport alignment, and improves Spearman correlation over the MusicEval baseline by 7% for musical quality and 14% for text alignment.
MLSP 2025 · Accepted
WSI turns the multilingual Whisper encoder into a speaker-embedding model, trained with online hard triplet mining and an NT-Xent loss, and reaches lower equal error rates than Pyannote, ECAPA-TDNN and x-vector baselines on English and non-English corpora.