I am a Speech and AI researcher/engineer working on automatic speech recognition (ASR), Arabic speech technologies, and robust speech modeling beyond adult read speech.

My current work focuses on building and evaluating ASR systems for challenging real-world conditions, including Arabic dialects, code-switching, children’s speech, long-form speech, and streaming ASR. I am especially interested in the gap between research prototypes and reliable production-ready speech systems.

• 🎙️ Research Assistant working on speech and AI at QCRI
• 🔭 Currently working on:
  • Arabic and multilingual ASR
  • Algerian dialect and code-switching ASR
  • Children’s speech recognition
  • Streaming ASR systems
  • Robustness and adaptation under distribution shift
• 🧠 Interested in both research and engineering: model training, evaluation, deployment, and reproducibility

• End-to-end ASR systems
• Arabic dialect ASR
• Code-switching ASR
• Children’s ASR
• Streaming and low-latency ASR
• Long-form ASR
• ASR evaluation, normalization, and benchmarking

• Adult-to-child ASR adaptation
• Robustness under domain shift
• Fine-tuning and LoRA-based adaptation
• Weight-space model merging
• Retention-aware adaptation

• Arabic text normalization
• Punctuation restoration
• Speech-to-text post-processing
• Multilingual and code-switched language modeling

• PyTorch and Hugging Face workflows
• ESPnet, k2/icefall, Sherpa-ONNX, and Whisper-based pipelines
• Dataset preparation and large-scale evaluation
• Reproducible experiments and benchmarking
• Deployment-oriented ASR pipelines

AlgerianSpeech is a platform dedicated to advancing speech recognition for Algerian Arabic, especially in realistic multilingual and code-switched settings involving Arabic, French, and English.

The project includes:

• A speech annotation platform for Algerian dialect and code-switching speech
• Real-world spontaneous speech collected from online recordings
• Transcription and annotation workflows for multilingual speech
• ASR evaluation pipelines using metrics such as WER, CER, and MER
• Resources for building more robust ASR systems for underrepresented Arabic dialects

This work supports the broader goal of improving speech technology for Arabic dialects and low-resource multilingual communities.

• Arabic ASR benchmarking
• Streaming ASR for Arabic
• Code-switching recognition and analysis
• Children’s speech recognition
• ASR robustness and domain adaptation
• Punctuation restoration for Arabic ASR output
• Dataset curation, normalization, and evaluation design

• Languages: Python, Bash, LaTeX
• ML/DL: PyTorch, Hugging Face Transformers, NumPy, pandas
• ASR: ESPnet, k2/icefall, Whisper, Sherpa-ONNX, NeMo
• Evaluation: jiwer, custom WER/CER pipelines, Arabic normalization tools
• Experimentation: Slurm, Conda, Git, Linux, GPU-based training/inference
• Deployment/Inference: ONNX, streaming ASR pipelines, model serving workflows

I am open to collaboration on projects related to:

• Arabic ASR and dialectal speech technologies
• Code-switching speech recognition
• Children’s speech recognition
• ASR benchmarking and evaluation
• Open-source speech tools and datasets
• Robust and deployable speech AI systems