mehdiyds/tunisian-speech-ai

Experimental STT/TTS pipeline for evaluation, benchmarking and fine-tuning of speech models for Tunisian Arabic (Derja). # Tunisian Derja STT — LoRA fine-tuning preparation Technical preparation for adapting the community baseline `oddadmix/Whisperv3-tunisian-codeswitch` to complementary Tunisian speech data. ## Status | Item | Status | | --- | --- | | Environment | prepared (requires Python 3.11) | | LoRA pipeline | prepared | | Dataset collection | in progress | | Training | **NOT STARTED** | | Final evaluation | pending | ## Baseline and evaluation boundary The baseline is a third-party fine-tuned community model. This project does not claim that its original fine-tuning was performed during the internship. The independent 96-utterance benchmark test set is strictly **external**. It must never be copied into `data/raw`, `data/processed`, `train`, or `validation`; it is reserved solely for the final baseline-versus-LoRA comparison. Decoding remains `language=ar`, `task=transcribe`, with the baseline processor and tokenizer unchanged. ## Setup Use Python **3.11** (not 3.14) and create an isolated virtual environment: ```powershell py -3.11 -m venv .venv .\.venv\Scripts\Activate.ps1 python -m pip install --upgrade pip pip install -r requirements.txt python scripts/check_environment.py ``` For exact reference versions, use `requirements-lock.txt`. Install a PyTorch CUDA build appropriate for the machine first via the official PyTorch selector. ## Dataset and smoke test Place only future training-corpus manifests under `data/raw/` and follow `docs/dataset_convention.md`. Validate a manifest: ```powershell python -m src.data.validate_dataset data/raw/manifest.jsonl ``` Once dependencies and model access are available, run the short technical test: ```powershell python scripts/smoke_test_lora.py --steps 2 --device auto ``` This uses synthetic audio only, attaches LoRA, performs a few optimisation steps, and writes an adapter in a temporary directory. It is not a scientific training run and does not demonstrate an accuracy improvement. See `docs/fine_tuning_plan.md` for t …