Tomoki Toda

UTMOS is vulnerable to attacks that can manipulate perceived speech quality while preserving or degrading its predicted scores, revealing critical flaws in its robustness.

Fine-tuning for phoneme addition may enhance naturalness but demands as much data as training from scratch to achieve comparable performance on new phonemes.

Achieving a notable increase in zero-shot cross-lingual SER performance by aligning emotions while eliminating speaker biases could redefine how we approach multilingual emotion recognition.

Integrating text and speech representations can dramatically enhance the intelligibility of electrolaryngeal speech, outperforming traditional methods by leveraging auxiliary information.