Multimodal Domain Generalization for Depression Detection: An Attention-Based BiLSTM Network with Domain-Adversarial Training
Depression detection models tend to pick up who is speaking, not whether they are depressed.
Research Highlights
Style-based Composer Identification and Attribution of Symbolic Music Scores: a Systematic Survey
A survey of composer-identification research with a simple point: before trusting a result, check whether the evaluation is strong enough to support it.
A Systematic Evaluation of Adversarial Attacks against Speech Emotion Recognition Models
A speech-emotion classifier looks fine on clean audio, then falls apart under small adversarial tweaks.
Optimizing Feature Extraction for Symbolic Music
A benchmark study of symbolic-music feature extraction showing where MIDI still holds up, where richer notation helps, and why harmony features can shift the result.
A perceptual measure for evaluating the resynthesis of automatic music transcriptions
A study on what gets lost when a performance is reduced to MIDI, and on how to evaluate resynthesis in a way that follows what listeners actually hear.