Layer-wise Cross-Lingual Depression Detection from Speech: Analysis with Contrastive Alignment
Abstract
A supervised contrastive alignment framework maps WavLM embeddings from English and Mandarin into a shared clinical space for depression detection, addressing cross-lingual generalization challenges and revealing performance artifacts caused by speaker identity leakage.
Significant disparities exist in the diagnosis and clinical presentation of depression across different linguistic populations. Speech-based depression detection performs well monolingually, but cross-lingual generalization remains an open challenge. A key reason is that prior work uses segment-level random splits without speaker grouping, leading to identity leakage that inflates reported metrics. We propose CLeaD, a supervised contrastive alignment framework that maps WavLM embeddings from English and Mandarin into a shared clinical space, without parallel data or target-language fine-tuning. Evaluating 52 Mandarin speakers, contrastive alignment modestly outperforms the baseline (F1: 0.640 vs. 0.622) under leave-one-speaker-out evaluation. It also improves depressed-class recall at intermediate layers (7-8), though the small test set limits generalizability. Two findings remain robust: model scaling degrades cross-lingual performance while improving monolingual English, and speaker identity leakage artificially inflated previously reported Mandarin F1 scores to 0.954, an artifact we reproduce and quantify.
Community
Depression doesn't look or sound the same in every language, but most speech-based detection models are only tested in the language they were trained on. Worse, prior work split data without grouping by speaker, so models were quietly memorizing voices instead of detecting depression. We built CLeaD, a supervised contrastive framework that aligns English and Mandarin WavLM embeddings into a shared clinical space, no parallel data or target-language fine-tuning needed. On 52 Mandarin speakers with leave-one-speaker-out evaluation, it modestly beats the baseline (F1: 0.640 vs. 0.622). Two findings hold up strongly: bigger models get worse at cross-lingual transfer while getting better at English, and speaker leakage inflated prior Mandarin F1 scores to 0.954. We reproduce and quantify that artifact.
Get this paper in your agent:
hf papers read 2607.02920 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper