01Yassine/cohere-transcribe-darija

Cohere Transcribe — Moroccan Darija (hybrid)

Public adapter on Cohere Transcribe Arabic. Hybrid: MultiConv on encoder layers 15–47 + LoRA on the decoder. Trained on 3h YouTube Darija (01Yassine/darija-asr-3h).

Eval: atlasia/darija-asr-benchmark (114 clips, human).

	CER	WER
base	20.2	49.1
hybrid (this repo)	14.4	38.3

You need the Cohere base weights (and its license). This repo is only the adapter.

Inference (from the Hub, no training clone)

pip install "transformers>=5.4" peft torch torchaudio soundfile huggingface_hub

from huggingface_hub import snapshot_download
import sys
sys.path.insert(0, snapshot_download("01Yassine/cohere-transcribe-darija"))
from infer import transcribe
print(transcribe("clip.wav"))

Or, if you already have infer.py from this repo:

python infer.py clip.wav --model hybrid
python infer.py clip.wav --model 01Yassine/cohere-transcribe-darija

Other open checkpoints

Same data and seed, different trainable slice:

recipe	Hub	AtlasIA CER
hybrid (MultiConv + LoRA)	01Yassine/cohere-transcribe-darija	14.4
full LoRA	01Yassine/cohere-transcribe-darija-full-lora	16.5
encoder LoRA	01Yassine/cohere-transcribe-darija-encoder-lora	17.4
decoder LoRA	01Yassine/cohere-transcribe-darija-decoder-lora	20.2

print(transcribe("clip.wav", model_id="full_lora"))
print(transcribe("clip.wav", model_id="01Yassine/cohere-transcribe-darija-encoder-lora"))