MaddoggProduction/whisper-l-v3-turbo-quran-lora-dataset-mix-ct2

whisper-l-v3-turbo-quran-lora-dataset-mix-ct2

CTranslate2 (float16) conversion of MaddoggProduction/whisper-l-v3-turbo-quran-lora-dataset-mix, packaged for fast inference with faster-whisper.

Same weights as the original (a whisper-large-v3-turbo fine-tune for diacritized Arabic Quran recitation ASR) — just the CTranslate2 runtime format, which is markedly faster and lighter than the transformers pipeline. For training details, evaluation (WER), and the dataset, see the original model card.

Usage

from faster_whisper import WhisperModel

model = WhisperModel(
    "MaddoggProduction/whisper-l-v3-turbo-quran-lora-dataset-mix-ct2",
    device="cuda",
    compute_type="float16",
)

segments, info = model.transcribe("recitation.wav", language="ar", beam_size=5)
print("duration:", info.duration)
print(" ".join(s.text for s in segments).strip())

Notes:

• —Long audio is handled automatically, no manual chunking needed.
• —Converted with ctranslate2==4.7.1 / faster-whisper==1.2.1. Ships tokenizer.json (large-v3 vocab) and preprocessor_config.json (128 mel bins).