Mathani-Ayat/rattil-qaloon-v3

Rattil Qaloon v3

Whisper-base for Quran recitation in Qaloon an Nafi (al-Fatiha and surahs 78-114). Recognises recited words; does not judge tajweed.

Results

Test voice: Sheikh Waleed, never used in training. Labels: normalizer qaloon-asr-v2-vocative-consonantal-yaa.

	No training	v2	v3
WER, normal speed	3.9%	2.0%	2.2%
WER, 1.25x speed	14.0%	16.9%	13.4%
WER, 1.5x speed	23.2%	21.5%	21.2%
Qaloon words (ملك, شا, كفؤا...)	64%	100%	100%
Invents a skipped ayah	0%	0%	0%

History

• Team model -> v1: base changed to deepdml/whisper-base-ar-quran-mix-norm; approved sheikhs only (Huthaify, Husary, Dokali); fixed corrupted labels (القيامة, يا أيها); removed mismatched clips.
• v1 -> v2: 6 more Qaloon sheikhs from mp3quran (forced-aligned, 3-model verified); pitch-preserving tempo augmentation (0.85x-1.45x).
• v2 -> v3: added Badr's Garu clips (544 after audit; 3 shifted Al-Asr clips removed); labels from the team's v2 normalizer (adds ءامنوا -> امنوا).

Training: 10 sheikhs, 5,205 clips, full fine-tune, lr 1e-5, 2 epochs, 72 min on an RTX 4060 Laptop GPU.

Use

from transformers import WhisperForConditionalGeneration, WhisperProcessor

processor = WhisperProcessor.from_pretrained("Mathani-Ayat/rattil-qaloon-v3")
model = WhisperForConditionalGeneration.from_pretrained("Mathani-Ayat/rattil-qaloon-v3")
features = processor(audio_16khz_mono, sampling_rate=16000, return_tensors="pt").input_features
print(processor.batch_decode(model.generate(input_features=features, language="arabic", task="transcribe"),
                             skip_special_tokens=True)[0])

In the Rattil studio: python -m src.streaming.serve --model gpu-full-base --model-path <local copy>.

Limits

• No tajweed assessment; tested on professional reciters, not learners.
• Qaloon word list and ayah merges still need a Qaloon teacher's review.
• Source recordings' licenses are not established; keep private.