Low-rank‐adaptation (LoRA) fine-tune of tarteel-ai/whisper-base-ar-quran for Arabic Qurʾān recitation (tilâwah).
Provides diacritic-sensitive ASR with a test WER ≈ 5.98 %, beating:

model	WER ↓	Δ vs ours
`KheemP/whisper-base-quran-lora`	0.0598	—
tarteel-ai/whisper-base-ar-quran	0.073	-1.3 ×
tarteel-ai/whisper-tiny-ar-quran	0.096	-1.6 ×
NVIDIA FastConformer large (NeMo)	≈ 0.069	-1.2 ×

(All scores measured on the same 610-ayah hold-out set, with no text normalisation – tashkīl included).

Quick start

python

from transformers import WhisperProcessor, WhisperForConditionalGeneration
from peft import PeftModel
import torch, soundfile as sf

base_id  = "tarteel-ai/whisper-base-ar-quran"
lora_id  = "KheemP/whisper-base-quran-lora"

# load model+processor
model  = WhisperForConditionalGeneration.from_pretrained(base_id, torch_dtype=torch.float16)
model  = PeftModel.from_pretrained(model, lora_id)
proc   = WhisperProcessor.from_pretrained(base_id)

# transcribe an mp3 -> text
audio, _ = sf.read("my_recitation.mp3")
inputs   = proc(audio, sampling_rate=16_000, return_tensors="pt").to(model.device)
pred_ids = model.generate(**inputs)
print(proc.decode(pred_ids[0]))

⚠️ This repo only stores the LoRA adapter (~2 MB). The code above automatically downloads the original Whisper base model and injects the adapter.

Model details


Back-bone	Whisper Base (77 M params)
LoRA rank / α / drop	8 / 16 / 0.05
Trainable params	0.59 M (0.8 %)
Epochs	5
Batch / grad-accum	2×4 (effective = 8)
LR / sched	5 · 10⁻⁴, constant
Mixed-precision	fp16
Hardware	single NVIDIA A100 40 GB

Target modules

q_proj, k_proj, v_proj, out_proj in both encoder & decoder self-attn and encoder-cross-attn blocks.

Training data

Dataset: 446 k MP3 ayāt scraped from https://quran.ksu.edu.sa, resampled to 16 kHz and paired with canonical text from all_ayat.json.
Filtering:
- keep ≤ 30 s duration (→ 6091 ayāt)
- pick shortest recording per ayah
- 90 / 10 split ⇒ 5481 train / 610 test
Reciters: 37; round-robin sampling ensures balanced voices.

Evaluation

Metric: jiwer WER with no normalisation (diacritics matter).
Result: 0.0598 on the 610-ayah test split (95 % CI ± 0.003).

Intended use & limitations

Designed for speech-to-text of Qurʾān recitations in Modern Standard Arabic. Not expected to work for:

conversational Arabic, dialects or non-Qurʾānic liturgy
noisy, low-quality microphones
verses longer than 30 seconds

Citation

bibtex

@software{quran_whisper_lora_2024,
  author       = {Kheem Dharmani},
  title        = {Whisper-Base Qurʾān LoRA Adapter},
  year         = 2024,
  url          = {https://huggingface.co/KheemP/whisper-base-quran-lora}
}

Licence

Back-bone weights under MIT (same as Whisper). Dataset sourced from the public domain. Adapter itself released under MIT.