Publications

A unified, configuration-driven open-source framework for deepfake audio detection research: 455+ pretrained models, 12 benchmark datasets, plug-and-play SSL frontends, classifier backends, losses, and augmentations, all via YAML.

The second iteration of the Iqra'Eval shared task, accepted as an Interspeech 2026 Challenge, expanding scope to general MSA pronunciation assessment beyond Quranic recitation.

A lightweight multi-scale convolutional adapter that adapts large SSL backbones to synthetic-speech detection with a tiny parameter footprint, while keeping cross-domain robustness.

Our DFKI-SLT submission to the ESDD 2026 Challenge: BiCrossMamba-ST combined with an attentive SSL fusion module for end-to-end speech deepfake detection.

Our DFKI-Speech entry to the WildSpoof Challenge: a robust framework for spoofing-aware automatic speaker verification (SASV) in unconstrained real-world conditions.

Shows that integrating para-linguistic cues (prosody, voice quality) alongside phonetic content improves robustness of pronunciation assessment systems in noisy / mismatched conditions.

A hybrid fusion framework integrating SSL-based representations with handcrafted spectral descriptors (MFCC, LFCC, CQCC) for robust speech deepfake detection across unseen domains.

A dual-branch spectro-temporal architecture with bidirectional Mamba blocks and mutual cross-attention. +67.74% / +26.3% relative gain over AASIST on ASVSpoof LA21 / DF21 with 28% fewer parameters.

First unified benchmark for Arabic pronunciation assessment using Quranic recitation.

First large-scale layer-wise study of Wav2Vec 2.0, HuBERT, WavLM for deepfake detection across multilingual and partial / song / scene-based scenarios. Lower layers consistently provide the most discriminative features → ~3× faster inference at comparable EER.

The first shared task benchmarking Qur'anic pronunciation assessment, with task design, baselines, and lessons learned.

Arabic-centric multimodal generative AI platform supporting language, speech, and image generation with Islamic RAG.

A morphology-aware extension of BPE that integrates linguistic structure into subword tokenization while preserving statistical efficiency.

A framework for dialectal sound and vowelization recovery in Arabic, tackling borrowed and dialectal sounds in phonologically diverse languages with limited data.

L1-MultiMDD: incorporates L1-aware speech representation via an L1-L2 embedding and multi-task learning to improve multilingual MDD.

A comprehensive review of recent advances in automatic pronunciation assessment for both phonemic and prosodic aspects, covering methods, challenges, resources, and future directions.

First end-to-end mispronunciation detection system for Modern Standard Arabic.

Multiple input views with auxiliary tasks yield more distinctive phonetic representations in low-resource MDD settings.

Fine-grained data augmentation pipeline that generates mispronunciation errors via masked, mix-factor blending of phonetic units. SOTA on Speechocean762 (+2.0 PCC) and +4.6 F1 on AraVoiceL2.

Supervisors: Dr. Ahmed Ali and Dr. Shammur Chowdhury.