oddadmix/cohere-transcribe-arabic-07-2026-dialectal-v2

cohere-transcribe-arabic-07-2026-dialectal-v2

Fine-tune of [`CohereLabs/cohere-transcribe-arabic-07-2026`](https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026) (2.07B) for multi-dialect Arabic speech recognition (undiacritized output), on the dialect-balanced dataset oddadmix/lahgtna-v3-small (52k train / 2.6k test, clean_text, seed 42).

Open weights. Part of the v2 dialectal-ASR family. The base model is already the #1 model on the Open Universal Arabic ASR Leaderboard, so this is a specialization experiment: it improves the core dialects with the most balanced training data (Gulf, Egyptian, Iraqi, Moroccan, Palestinian, Lebanese) but regresses on others (notably a Tunisian failure mode), landing slightly below the base overall. Use the base if you want the best all-round accuracy; this checkpoint if your traffic is concentrated in the core dialects above.

Results (balanced 13-dialect test, ~2,600 clips, clean_text WER/CER)

Per-dialect — fine-tuned vs. base

Fine-tuned wins on 6/13 dialects (large gains on Gulf/Iraqi/Moroccan) but a Tunisian hallucination failure (WER > 1) and Libyan/Yemeni/Syrian regressions pull the average above the base.

v2 family — all models on the same balanced test (clean_text, seed 42)

Examples — base vs. fine-tuned (one clip per dialect)

13 held-out test clips, one per dialect. Text is clean_text-normalized (undiacritized), matching the WER scoring. Press play to listen.

Usage

import torch
from transformers import AutoProcessor, CohereAsrForConditionalGeneration  # transformers 5.14-dev

proc = AutoProcessor.from_pretrained("oddadmix/cohere-transcribe-arabic-07-2026-dialectal-v2")
model = CohereAsrForConditionalGeneration.from_pretrained(
    "oddadmix/cohere-transcribe-arabic-07-2026-dialectal-v2", torch_dtype=torch.bfloat16).to("cuda").eval()

# Process ONE bare array at a time — batching+padding degrades this model.
inp = proc(audio_16k_float32, sampling_rate=16000, language="ar", return_tensors="pt").to(model.device)
inp["input_features"] = inp["input_features"].to(model.dtype)
g = model.generate(**inp, max_new_tokens=256)
print(proc.tokenizer.batch_decode(g, skip_special_tokens=True)[0])

Limitations

• —Below the base overall (0.514 vs 0.470) — the base is a leaderboard-#1 generalist.
• —Tunisian failure mode (WER > 1): avoid for Tunisian; Libyan/Yemeni/Syrian also regress.
• —Undiacritized output; evaluated with clean_text (strip tashkil/tags, keep dialect spelling).