Yassine El Kheir

Yassine El Kheir

Speech and audio · DFKI and TU Berlin · PhD expected April 2027

Yassine El Kheir

I am a PhD researcher at the German Research Center for AI (DFKI) and TU Berlin, advised by Sebastian Möller and Tim Polzehl. I build speech and audio systems that have to hold up outside a single benchmark: self-supervised representations, small detectors, and speech LLMs that can say why a voice is synthetic. I expect to defend in April 2027.

I am looking for research and applied scientist roles from November 2026. Write to elkheiryassine0@gmail.com. A two-page CV is here.

How the work connects

Arabic speech came first. At QCRI I built pronunciation and recognition systems, including QVoice, and contributed to Fanar, Qatar’s Arabic-centric multimodal model, including the MorphBPE tokenizer.

That turned into a question about representations. A layer-wise study of Wav2Vec 2.0, WavLM, and HuBERT showed that early layers carry the deepfake signal, which makes inference about three times faster. BiCrossMamba-ST then applied a Mamba model to the same problem, with 28% fewer parameters than a Transformer baseline. The DFKI-SLT system built on it placed 1st of 97 teams on Track 2 of the ESDD 2026 challenge at ICASSP, and 2nd on Track 1 (results).

The next step is a model that can explain the decision. From December 2025 to April 2026 I visited NII in Tokyo, working with Xin Wang and Junichi Yamagishi on speech LLMs that ground a detection in text. The paper is Textual Acoustic Grounding, under review at IEEE SLT 2026. Models from that work are on Hugging Face.

The version other people can run is DeepFense: an open, YAML-driven detection framework with 455+ pretrained models, which I presented at Interspeech 2026. The same line ships at Gretchen AI, whose detector ranked first on Deepfake-Eval 2024, and in production speech systems at CAMB.AI across 150+ languages.

I lead Fake-O-Meter and coordinate it day to day. DFKI runs the consortium. The partners are TU Berlin, Otto von Guericke University Magdeburg, Ernst-Abbe-Hochschule Jena, Deutsche Welle, and dpa. The project builds a multimodal assistant for image, audio, and video deepfakes: detection, explanations grounded in professional fact-checks, a dialogue assistant, and prebunking around elections. Fact-checkers stay in the loop. It is funded by the German Federal Ministry of Research, Technology and Space, from May 2026 to February 2029. Project site. DFKI page.

I also lead VERIDIC, and a Software Campus project with Holtzbrinck and Springer Nature: €115k in federal funding, from July 2026, on multimodal deepfake detection and fact-checking, with a team of students and research assistants.

Recent

  • October 2026. Invited talk at the GrokBot Casablanca Meetup: MoulSot, the first Darija ASR model, with Abdeljalil El Majjodi. MoulSot.
  • September 2026. Interspeech 2026, Sydney. I presented DeepFense and the IQRA 2026 challenge, and co-authored Evidence Subspace Projection with the University of Stuttgart.
  • July 2026. Co-organized ArA-DF 2026, the Arabic speech deepfake detection shared task at ArabicNLP, co-located with EMNLP 2026. Sixteen teams, two tracks.
  • July 2026. I lead a Software Campus project with Holtzbrinck and Springer Nature (€115k, federal funding) on multimodal deepfake detection and fact-checking, and I lead VERIDIC.
  • May 2026. Fake-O-Meter starts. I lead the project and coordinate the partners day to day. DFKI page.
  • May 2026. Speech ML mentor and PhD panelist at HackAI Morocco.
  • April 2026. Finished the visiting stay at NII Tokyo. Paper.

Talks

GrokBot and Interspeech 2026 are in the list above. Slides, video, and the earlier talks are on the talks page.

Selected work

The projects page has the longer notes.

Selected papers

Full list on the publications page and on Google Scholar.

Positions

  • PhD researcher, DFKI and TU Berlin, 2024 to present. Advisors: Sebastian Möller and Tim Polzehl.
  • Project lead, Fake-O-Meter, May 2026 to February 2029. I lead the project and coordinate the consortium day to day: TU Berlin, OVGU Magdeburg, Ernst-Abbe-Hochschule Jena, Deutsche Welle, and dpa. DFKI is the coordinator. Funded by the BMFTR. Project site. DFKI record.
  • Project lead, Software Campus, July 2026 to present. I lead a €115k federally funded project with Holtzbrinck and Springer Nature on multimodal deepfake detection and fact-checking, with a team of students and research assistants.
  • Project lead, VERIDIC.
  • Visiting researcher, NII, Tokyo, December 2025 to April 2026. Xin Wang and Junichi Yamagishi.
  • Founding research scientist (part-time), Gretchen AI, Berlin, January 2025 to present.
  • Research consultant, CAMB.AI, April 2025 to present. Multilingual ASR, TTS, and dubbing.
  • Research associate, QCRI, Doha, 2022 to 2024. Fanar and QVoice.

Education. PhD, Computer Science, TU Berlin / DFKI, 2024 to 2027, defense expected April 2027. MSc, Machine Learning, KTH Royal Institute of Technology, 2022. Diplôme d’Ingénieur, Télécom Paris, 2022, top 5%.

Collaboration

I work with students and researchers anywhere. Master’s theses, research visits, and joint papers are all welcome: speech foundation models, audio deepfake detection, speech LLMs, ASR and TTS, pronunciation, and low-resource speech, including Arabic and Darija.

Current students and how to write are on the talks and collaboration page.

Contact

elkheiryassine0@gmail.com · yassine.el_kheir@dfki.de · Google Scholar · GitHub · LinkedIn · CV