OthmaneW/tachelhit-asr

Tachelhit ASR ⵜⴰⵛⵍⵃⵉⵜ

🚧 Work in progress — no trained weights yet.

Speech recognition and speech translation for Tachelhit (Souss Amazigh, shi).

• Phase A: Tachelhit speech → Arabic text (Whisper fine-tuned on existing open data).
• Phase B: first open Tachelhit transcription benchmark (IRCAM Latin, Tifinagh derived automatically).

Code, configs and roadmap: github.com/OthmaneW37/tachelhit-asr

Training data

Dataset	Licence
SoufianeDahimi/Tamazight-ASR-Dataset-v2	Apache-2.0
Ma-OpenHub/Tutlait-v1 (Souss accent)	CC-BY-4.0

Results

Coming soon.

Hardware

Trained on a single AMD Radeon RX 9060 XT (16 GB) with ROCm.