MarsadLab at NADI Shared Task: Arabic Dialect Identification and Speech Recognition using ECAPA-TDNN and Whisper
Year: 2025
Type: conference-paper
Abstract: We participated in NADI 2025 shared tasks on Arabic Dialect Identification (ADI) and Automatic Speech Recognition (ASR) across eight Arabic dialects.For ADI, we employ an enhanced ECAPA-TDNN with VoxLin-gua107 initialization, featuring self-attention classification head, progressive unfreezing, advanced augmentation, and test-time augmentation.This approach ranked third with 61.6% accuracy and 0.3068 macro cost.For ASR, we implement a zero-shot cascaded system using Whisper Large-v3 and MARBERT with extreme parameter efficiency (0.0004% trainable), ranking seventh with 104.895WER and 84.693 CER.Our results validate complementary paradigms: direct audio processing for competitive dialect classification versus foundation model robustness for cross-dialectal transcription.
Authors: Md. Rafiul Biswas, Kais Attia, Shimaa Ibrahim, Mabrouka Bessghaier, Wajdi Zaghouani
Institution: Northwestern University in Qatar
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037644