mehdiyds/tunisian-stt-benchmark

Benchmark and selection of Speech-to-Text models for Tunisian dialect (Derja) # Tunisian STT benchmark This repository documents an evaluation of open-source speech-to-text (STT) models for Tunisian Derja. Five model experiments were evaluated on an independent test set of 96 utterances. The benchmark records WER, CER, latency, real-time factor (RTF), and GPU-memory measurements when they are available. The selected baseline is `oddadmix/Whisperv3-tunisian-codeswitch`, a community model already fine-tuned before this internship and evaluated here as a baseline. The test set is strictly reserved for evaluation: it must not be added to training, validation, augmentation, LoRA, or fine-tuning data. Audio files are deliberately not included in this repository. See the benchmark documentation for the protocol, results, limitations, and reproducibility checks.