"Bonjour, n7eb nkomandi deux pizzas, w zid coca." An ordinary sentence in Tunis, and a real headache for most speech recognition models. Applied research subject for a pair of students.
Context
Ordy takes phone orders in French, English and Tunisian Darja. Off-the-shelf models handle Darja poorly, and struggle with Arabic/French code-switching within a single sentence, especially over a noisy phone line. Every misrecognised word can become a wrong order. This project aims to close that gap, measurably.
What you will do
- Build an annotated Darja audio corpus, collected with speakers' consent, including code-switching
- Define consistent transcription conventions (Arabic script, Latin script, "Arabizi" digits)
- Fine-tune Whisper with LoRA, using data augmentation: telephone noise, regional accents, varied speaking rates
- Extract intents and entities (order, appointment, HR request) from transcripts
- Deploy a low-latency, real-time streaming pipeline
- Evaluate each stage separately, then end to end
Technical challenges
- Scarce data: every hour of annotated audio counts, and quality matters more than volume
- Code-switching: several languages and scripts within one sentence
- Real-time constraints: a phone conversation does not tolerate waiting
- Fine-tuning without degrading French and English performance
Deliverables
- An annotated, documented corpus, with its collection and consent protocol
- An adapted Whisper model and its evaluation results
- An understanding module (intents, entities) integrated into the B2H platform
- A streaming demo and technical documentation
- The final-year thesis; a scientific publication is possible if the results warrant it
Success criteria
- WER measured before and after adaptation, on a held-out test set
- End-to-end latency measured under streaming conditions
- Accuracy of extracted orders on test calls
- No significant regression in French and English
Tech stack
- Whisper, LoRA / PEFT
- PyTorch
- Streaming ASR
- NLU (intent and entity extraction)
- Python
Who we are looking for
- Final-year engineering or master's students with an AI or data science background
- Hands-on deep learning experience; audio or NLP experience is a plus
- Arabic speakers, comfortable with Tunisian Darja
- An interest in research and experimental rigour
- French and technical English
Practical details
- Duration: 4 to 6 months (February – June 2027, indicative)
- Supervised by the B2H engineering team, with a weekly check-in
- Integrated into the internal B2H platform (single identity, event bus, shared data)
- How to apply: CV + cover letter, at most 2 subjects per candidate
- Deadline: 15/12/2026