TriVector@DravidianLangTech 2026: Depression Detection from Tamil and Malayalam Speech with Speaker-Independent Evaluation using MFCC and Wav2Vec2

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing & Speech Technology · Depth: Advanced, medium

Summary

The "TriVector@DravidianLangTech 2026" system addresses depression detection from Tamil and Malayalam speech, a significant mental health concern often reflected in subtle speech changes. This work tackles the particular challenges of low-resource and multilingual environments. The system integrates both handcrafted acoustic features, specifically MFCC, and pretrained speech representations from Wav2Vec2, employing a straightforward fusion strategy to combine their strengths. Observations indicated that Wav2Vec2 generalized more effectively for Malayalam speech, while for Tamil, a validation-tuned probability fusion yielded superior results. The system achieved impressive macro-F1 scores of 99.5% for Malayalam and 88.6% for Tamil, earning 3rd place in both categories of the Shared Task on Depression Detection from Malayalam and Tamil.

Key takeaway

NLP Engineers developing speech-based mental health detection in low-resource, multilingual settings should combine MFCC features with Wav2Vec2 models. This hybrid strategy, especially with validation-tuned fusion, significantly boosts macro-F1 scores. It achieved 99.5% for Malayalam and 88.6% for Tamil. Tailor fusion strategies to each language for optimal performance.

Key insights

Combining MFCC and Wav2Vec2 features effectively detects depression in low-resource Tamil and Malayalam speech, achieving high F1 scores.

Principles

• Speech patterns reveal mental health states.
• Feature fusion enhances detection accuracy.
• Pretrained models generalize well for low-resource languages.

Method

The system uses MFCC and Wav2Vec2 features, applying a simple fusion strategy. For Tamil, a validation-tuned probability fusion was optimal, while Wav2Vec2 alone performed best for Malayalam.

In practice

• Explore MFCC and Wav2Vec2 for speech tasks.
• Implement feature fusion for improved robustness.
• Tune fusion strategies per language.

Topics

• Depression Detection
• Speech Processing
• Dravidian Languages
• MFCC Features
• Wav2Vec2
• Feature Fusion

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Speech-Driven End-to-End Language Discrimination towards Chinese Dialects — Speech-driven features, integrating MFCC and word embeddings through a CNN, significantly improve Chinese dialect discrimination.
• Recurrence-Based Nonlinear Vocal Dynamics as Digital Biomarkers for Depression Detection from Conversational Speech — Depression alters vocal recurrence structure, detectable via nonlinear state-space analysis of conversational speech.
• Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages — Data scarcity, not phonological complexity, primarily drives ASR errors in low-resource, typologically complex languages.
• TMASC: Transmasculine Attitude and Speech Corpus — TMASC provides a multimodal corpus for research into transmasculine vocal health and characteristics.
• Hybrid CNN-Transformer Architecture for Arabic Speech Emotion Recognition — A hybrid CNN–Transformer model significantly improves Arabic Speech Emotion Recognition by combining local spectral and global temporal feature learning.
• Towards Personalized Federated Learning for Dysarthric Speech Recognition — Personalized federated learning significantly improves dysarthric speech recognition by addressing speaker heterogeneity while maintaining privacy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.