Projects

BiCrossMamba-ST

One of the first successful applications of Mamba state-space models to speech anti-spoofing. A bidirectional Mamba architecture with spectro-temporal cross-attention that achieves SOTA on ASVSpoof LA21 and DF21 with 28% fewer parameters than Transformer baselines.

• Shows Mamba's sequence modeling translates directly to anti-spoofing: captures temporal artifact patterns more efficiently than attention-based models.
• Architecture later extended in the DFKI-SLT system for the ESDD 2026 challenge ranked 1st on the track 2, and 2nd on the track 1.

SSL Interpretability, Spectral Fusion & Efficient Adaptation

A connected line of work on understanding and improving SSL-based deepfake detectors:

• Layer-wise SSL Analysis (NAACL Findings 2025): first large-scale study showing lower SSL layers suffice for artifact detection — enabling ~3× faster inference with no accuracy loss. Challenges a decade of practice.
• Two Views, One Truth (WASPAA 2025): hybrid fusion of SSL embeddings with handcrafted spectral descriptors (MFCC, LFCC, CQCC) for cross-domain robustness.
• Parameter-Efficient Multi-Scale Adapter (ICASSP 2026): adapter-style fine-tuning of large SSL backbones with a tiny parameter footprint, keeping cross-domain generalization.

NII Japan · Audio LLMs for Speech Understanding

Visiting research stay working with Xin Wang and Junichi Yamagishi — co-founders of the ASVspoof challenge — on applying large language models to speech understanding and deepfake detection. Extended the detection problem beyond classifiers toward LLM-based reasoning over audio.

• Explored explicit textual grounding for audio deepfakes: teaching LLMs to reason about speech authenticity in natural language rather than binary classifier scores.
• Directly informed the design of Bridging the Modality Gap (see below), submitted to SLT 2026.

Bridging the Modality Gap · SLT 2026 (under review)

Applies explicit textual grounding to LLM-based deepfake voice detection — moving beyond classification heads toward models that reason about speech authenticity in natural language. Bridges acoustic and semantic modalities by grounding audio representations in textual descriptions of synthesis artifacts.

• Part of a broader direction on Audio LLMs for trustworthy speech understanding.
• Built on the foundation of the NII Japan residency with Xin Wang and Junichi Yamagishi.

DeepFense, Modular Deepfake Audio Detection Framework

Open-source, configuration-driven framework that makes deepfake audio detection research reproducible and composable. Mix and match SSL frontends, classifier backends, losses, and augmentations entirely via YAML — without touching code.

• Plug-and-play registry: Wav2Vec 2.0, WavLM, HuBERT, MERT, EAT frontends; AASIST, ECAPA-TDNN, Nes2Net, RawNet2 backends.
• 455+ pretrained models and 12 benchmark datasets (ASVSpoof19, CompSpoof, DECRO, SONICS) on HuggingFace.
• Built-in augmentations (RawBoost, RIR, codec, noise) and standardized metrics (EER, minDCF, actDCF, F1).

News-Polygraph

A collaborative German research initiative building a multimodal platform for detecting and analyzing disinformation across speech, image, and text. My contribution is on the speech and audio side: robust generalizable deepfake detection, anti-spoofing, and interpretability — which feed into the DeepFense framework above.

Gretchen AI, Research & Engineering Lead

Co-leading R&D for generative-media detection at a Berlin-based AI startup. Built a proprietary image deepfake detector that ranked #1 on the Deepfake-Eval 2024 benchmark (~83% accuracy), and designed the production inference pipeline. Direct productionization of PhD research.

Research Consultant · Production Speech Systems

Research consulting on speech AI for production-scale multilingual TTS and content localization. Work spans model robustness, speech quality assessment, and evaluation across 100+ languages in live production environments.

AtlasIA, Open-Source AI for Morocco

Leading ASR and TTS development for Moroccan Darija in an open-source community building AI grounded in Moroccan linguistic identity:

• MoulSot (2026): curated 80 hours of high-quality Moroccan Darija speech distilled from 1,500 hours of YouTube content through a 9-stage pipeline (Silero VAD, SQUIM, Audiobox Aesthetics, DNS64 denoising, pyannote diarization, Argilla human-in-the-loop, Gemini 2.5 Pro transcription), then fine-tuned Qwen3-ASR-1.7B on top. Released as atlasia/MoulSot-Full and atlasia/moulsot.v0.3.
• Building the first natural Text-to-Speech system and tokenizer for Moroccan Darija.
• Mentoring HackAI Morocco teams on forced alignment and low-resource ASR.

Iqra'Eval, Quranic Pronunciation Assessment

A multi-institutional initiative to benchmark Arabic pronunciation assessment via Quranic recitation. Accepted as an Interspeech 2026 Challenge — establishing a lasting community benchmark:

• Founded and ran the Iqra'Eval Shared Task at ArabicNLP 2025, the first standard benchmark for Quranic recitation analysis.
• Co-built QuranMB, the first expert-annotated dataset for Quranic mispronunciation (~2 h), leading a 15-researcher team at the SDAIA Winter School.
• Second iteration accepted as IQRA 2026 at Interspeech, expanding scope to general MSA pronunciation.

QVoice and the Arabic Speech Stack

Led QVoice, the first end-to-end mispronunciation detection system for Modern Standard Arabic, and contributed a series of SOTA results on L2 speech assessment and multilingual ASR:

• SpeechBlender (SLaTE 2023), fine-grained augmentation for mispronunciation data scarcity.
• L1-aware Multilingual MDD (ICASSP 2024), incorporating native-language priors into multilingual mispronunciation detection.
• Beyond Orthography (ACL 2024), recovering short vowels and dialectal sounds in Arabic with limited data.
• AraVoiceL2, non-native Arabic speech dataset for phoneme-level evaluation.

Fanar & MorphBPE

Contributed to Fanar, Qatar's Arabic-centric multimodal LLM, and designed MorphBPE, a morphology-aware tokenizer that integrates linguistic structure into subword segmentation. Shows how foundation-model design choices can be tuned to morphologically rich languages without sacrificing efficiency.

• MorphBPE improves fertility rates and downstream generation quality for morphologically rich languages, with negligible overhead vs. plain BPE.
• Helped run large-scale data filtering and training pipelines (300M to 3B parameters) with MosaicML's LLM-Foundry.