Year: 2025
Type: conference-paper
Abstract: This paper presents Lahjati (ECAPA-WavLM Fusion with Multi-Stage Optimization) system for the spoken Arabic Dialect Identification (ADI) subtask at Nadi 2025 (Talafha et al., 2025), The task aims to automatically identify the dialect of spoken Arabic utterances, a challenging problem due to the rich linguistic diversity of Arabic and the scarcity of labeled speech resources.Our approach combines ECAPA-TDNN embeddings from SpeechBrain with WavLM-base representations.The proposed system achieved 94.08% accuracy on the validation set and ~51.0% on the test set.Challenges included differentiating acoustically similar dialect pairs and mitigating the effects of varied recording conditions, which likely contributed to performance degradation on unseen data.These findings highlight both the potential and limitations of fusing complementary speech representations for robust dialect identification.
Authors: Sanad Albawwab, Omar Qawasmeh
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037756