Answers are grounded in this paper and cite their sources.
Claim Detection involves the task of recognizing claims within an utterance, differentiating between claims and non-claims. Although significant progress has been achieved in text-based claim detection for high-resource languages, audio-based claim detection remains largely unexplored, particularly in low-resource languages such as Bangla. To the best of our knowledge, our study presents the first systematic approach to audio-based claim detection for Bangla and introduces the first annotated Bangla audio claim detection dataset comprising 4,000 manually annotated speech samples, equally divided between claim and non-claim instances. We evaluate acoustic, ASR-based text, and combined multi-modal baselines under a leakage-safe, speaker- and video-disjoint evaluation protocol. We adapt a Transformer encoder with Learned pos. encoding (projected one-hot) for frame-level MFCC sequences, demonstrating its effectiveness for temporal acoustic modeling in Bangla, a low-resource language. Unlike conventional classifiers that rely on aggregated acoustic statistics, the proposed model preserves frame-level temporal structure and employs multi-head self-attention to capture long-range dependencies across speech segments. A two-phase training strategy is further adopted. Its contribution lies in demonstrating that this adaptation is effective for Bangla audio claim detection, not in architectural novelty. To provide a comprehensive benchmark, we evaluate traditional machine learning models (Logistic Regression, Random Forest, SVM, XGBoost), deep learning architectures (CNN, LSTM, CNN-LSTM hybrids, DNN, CNN+BiLSTM and CRNN+Attention), transformer-based models (XLS-R, Vision Transformer, AST, XLM-R), large self-supervised speech models (WavLM-Large, HuBERT-Large, XLS-R-300M), text SSL models (BanglaBERT, XLM-RoBERTa, mDeBERTa-V3) with parameter-efficient fine-tuning and an audio-capable large language model (Qwen2-Audio-7B). The proposed Time-Series Transformer achieves the strongest overall performance at 84.4% accuracy, performing comparably to the best text-based baseline (XLM-RoBERTa Large, 83.0%) while significantly outperforming the strongest audio-based deep learning and self-supervised baselines, and eliminating the need for an intermediate ASR transcription pipeline. The audio-capable LLM Qwen2-Audio-7B achieves only 53% accuracy despite its 7B parameter scale, confirming that model size alone does not compensate for language and domain gap in low-resource settings.
No takes yet. Share an insight, caveat, or question.
Roy et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: