Modelpublic

baristiran/whisper-large-v3-quran-lora

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes36downloads

Model Card

Whisper Large-v3 Quran LoRA

Fine-tuned openai/whisper-large-v3 for Quranic Arabic speech recognition using LoRA.

Performance

Metric	Score
WER (normalized)	1.58%
WER (diacritized)	3.92%
Tarteel AI baseline	5.75%

Training Details

Parameter	Value
Base model	openai/whisper-large-v3 (1.5B params)
Method	LoRA (r=32, alpha=64)
Target modules	qproj, kproj, vproj, outproj, fc1, fc2
Trainable params	~13M (0.87%)
Dataset	tarteel-ai/everyayah (~127K samples)
Epochs	5 (~18K steps)
Batch size	32 x 2 = 64 effective
Learning rate	3e-4 (cosine scheduler)
Precision	fp16
GPU	NVIDIA A100 80GB
Training time	~12 hours

Usage

python

from transformers import WhisperProcessor, WhisperForConditionalGeneration
from peft import PeftModel

processor = WhisperProcessor.from_pretrained("baristiran/whisper-large-v3-quran-lora")
base_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v3")
model = PeftModel.from_pretrained(base_model, "baristiran/whisper-large-v3-quran-lora")

# Transcribe
inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")
predicted_ids = model.generate(**inputs, max_length=225)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

Project

Part of Tilavet — open-source Quran STT + translation platform.

Author

Tarik Ismet ALKAN