Update (June 2026): weights upgraded in place - real-phone word-error roughly halved (~20% → ~10%) with the same clean-studio accuracy, and the pronunciation/tajweed head retrained to match. Same file layout as before. Live low-latency streaming model: fastconformer-quran-streaming.

FastConformer-Quran

State-of-the-art automatic speech recognition for Quranic recitation, with a multi-signal mispronunciation detector built on top.

🚀 #1 on a leakage-free Quran benchmark (Hafs riwayah): 4.13% overall WER on held-out audio, ahead of nvidia FastConformer (8.14%, current public #1) and Tarteel Whisper (21.31%) on identical clips. On held-out reciters never seen in training: 0.93% WER. 🎯 ~80% sensitivity at ~9% FPR on a 39K-token held-out mispronunciation benchmark (verified on the current weights). 📱 iOS-ready via the companion CoreML repos (offline · streaming).

---

Headline numbers

Honesty note (June 2026). Earlier versions of this card quoted a 0.029% WER on the EveryAyah test split. That split overlaps our training data: the public EveryAyah test shard shares ~88% of its clips with reciters we trained on, so that number measured memorization, not generalization. The numbers below come from a leakage-free held-out benchmark - every clip is verified absent from training - and are the ones you should trust. The model is still SOTA on this stricter test; it just reports an honest ~0.9% on unseen reciters instead of a leaked 0.029%.

📊 Live, interactive board (this model vs nvidia, whisper, seamless, mms, omniASR, cohere, Tarteel): Quranic ASR Leaderboard.

Leakage-free Quran benchmark (600 clips, same decoding/normalization for every model)

Three held-out sources, 200 clips each: EveryAyah reciters with zero training samples (clean studio), a QUL reciter (Al-Nufais) we never trained on, and held-out real phone-mic recitation (tlog). Ranked by overall WER (lower = better); WER/CER are over diacritic-normalized text.

Model	Family	EveryAyah (unseen)	QUL (unseen)	Phone (tlog)	Overall WER	Overall CER
FastConformer-Quran (this model, offline)	ours	0.93	4.42	8.88	4.13	1.68
nvidia stt_ar_fastconformer_hybrid	baseline (public #1)	1.50	9.52	16.77	8.14	3.73
⭐ Tarteel (official, realtime)	Tarteel production (streaming)	5.97	12.91	16.17	10.99	7.14
whisper-large-v3	whisper	8.69	7.80	25.90	12.51	6.73
Tarteel whisper-base-ar-quran (old open model)	whisper	21.04	14.32	32.48	21.31	10.39
seamless-m4t-v2-large	seamless	18.97	29.67	29.59	25.48	15.35
mms-1b-all	mms	40.94	55.06	44.25	46.95	12.82

⭐ Tarteel (official, realtime) is Tarteel's current production ASR (a streaming FastConformer), not their old open Whisper model. We ran all 600 benchmark clips through it at voice-v2.tarteel.io and scored them with the same scorer. It is by far the strongest external system here (10.99 overall), far ahead of their old public whisper-base-ar-quran (21.31).

This model is #1 overall, ahead of the current public leaderboard #1 (nvidia FastConformer) on the same held-out clips. Against Tarteel's official production model (⭐, their realtime streaming FastConformer at voice-v2.tarteel.io), this model leads overall 4.13 vs 10.99.

One honest caveat on that head-to-head: this repo is the offline (full-utterance) model, and Tarteel's production model is streaming (realtime), so it is not a like-for-like comparison - full context is an advantage. The fair streaming-vs-streaming number is our separate streaming model at 11.96 overall, which is competitive with Tarteel's 10.99 (within the benchmark's noise on a 600-clip set). So: our offline model is clearly ahead of every external system including Tarteel official; our streaming model is roughly on par with Tarteel's streaming production system. General Arabic models (whisper-large-v3, seamless, mms) are strong on broadcast Arabic but degrade on diacritized Quranic recitation - which is the point of a Quran-specific model. Full interactive board: Quranic ASR Leaderboard Space.

The held-out phone-mic column is the hard, real-world case. tlog also carries substantial label noise (filename↔ayah mismatches, user-added basmala, partial recitations), so a portion of that 8.88% is the model correctly transcribing what was actually said against wrong metadata.

Two further leaderboard families were attempted but not scored here: cohere-transcribe-03-2026 (gated, requires license access) and omniASR-LLM-7B (fairseq2 7B - loads but its CUDA inference path segfaults on our hardware). They can be folded into the benchmark on a machine where they run.

---

Demo audio

Alafasy reciting Q 1:4:

Predicted: مَالِكِ يَوْمِ الدِّينِ ✓

Abdullah Basfar reciting Q 112:1:

Predicted: قُلْ هُوَ اللَّهُ أَحَدٌ ✓

Alafasy reciting Q 78:2:

Predicted: قُلْ هُوَ نَبَأٌ عَظِيمٌ ✓

Try the live demo: Space.

---

Mispronunciation detection

We also ship a multi-signal pronunciation scorer combining three orthogonal signals on the same CTC architecture:

Signal	What it measures	Style-invariant
Pronunciation head v7	Learned P(token correctly pronounced) on 1.33 M-parameter MLP over encoder features	✓
Reference-anchor distance	Cosine distance to master qari centroid bank (multi-ayah aware)	partial
CTC GOP	log P(expected token) minus max log P(non-blank token), averaged over CTC interval	✓

A token is flagged by the consensus rule when at least 2 of 3 signals agree (default thresholds: head < 0.5, anchor > 0.20, GOP < -3.0).

Held-out evaluation

On 39,173 tokens from 996 tlog clips that were never seen by the pronunciation head during training, with per-token consensus labels from our ASR and ElevenLabs Scribe v2:

Detector	TPR @ 1% FPR	TPR @ 5% FPR	AUC
GOP (style-invariant)	73.7%	78.9%	0.907
Pronunciation head	73.0%	77.6%	0.940
Anchor distance (where bank covers)	2.6%	13.8%	0.722
Consensus (2 of 3)	n/a	n/a	80.3% TPR / 9.3% FPR

The combined detector catches ~80% of real mispronunciations at a ~9% false-positive rate. This is the deployable operating point. (Re-verified June 2026 on the current shipped weights; the head and encoder are the same vintage.)

---

What this model is and isn't

Is:

• The best published ASR for Quranic recitation in Hafs riwayah.
• A frame-level CTC model with 512-dim encoder features exposed for downstream scoring.
• Production-ready ONNX (fp32 437 MB, fp16 219 MB), running at RTF ~0.001 on an RTX 4090.
• Diacritic-aware: outputs fully harakat-marked Arabic text.

Isn't:

• A general Arabic ASR. Trained only on Quranic audio with a 1,024-token BPE tokenizer. Performance on dialectal, news, or conversational Arabic will be poor by design.
• A streaming model itself. This repo is the offline (full-utterance) model. For true token-by-token low-latency streaming there is now a separately-trained cache-aware variant: fastconformer-quran-streaming (+ CoreML/ANE). On the leakage-free benchmark the streaming model trades some accuracy for latency (11.96% overall WER vs 4.13% offline), as expected.
• Trained for other qira'at. Hafs riwayah only.

---

How it was built

Stage 1: base training on EveryAyah. Fine-tuned NVIDIA's stt_ar_fastconformer_hybrid_large_pcd_v1.0 (Arabic-pretrained, ~1100 h) on 22 K EveryAyah clips. Reached 0.0757% WER on the held-out test split.

Stage 2: pronunciation scoring stack. Built a per-token head on top of frozen encoder features (512-dim pooled + 64-dim token embedding + 16-dim Quran-phonology features into an MLP). Initial training on weak labels (CTC-vs-expected disagreement + GOP scores) plus master qari anchors (Husary, Abdul Basit, Alafasy clean recitations).

Stage 3: phone-audio fine-tune. Three rounds of low-LR continuation on EveryAyah and tlog: 22,585 clean clips, 6,869 high-quality tlog clips (full weight), 20,589 borderline tlog clips (half weight). LR schedule 1e-5, 5e-6, 2.5e-6 over six epochs total. Trend was monotonic improvement on a held-out tlog slice through round three, then saturated.

Stage 4: dual-ASR consensus labels. Ran ElevenLabs Scribe v2 over the 6,168 highest-quality tlog clips. Aligned Scribe transcripts vs. expected ayah text at the character level (diacritic-insensitive Levenshtein), then aligned vs. our ASR output. Asymmetric-trust consensus rule:

A token is labeled CORRECT if EITHER ASR or Scribe says correct. It is labeled WRONG only when BOTH agree wrong.

Result: 144,664 per-token labels at ~98.0% positive rate, with 6,789 ASR-vs-Scribe disagreements flagged as high-information tokens. Cost: ~$11 in ElevenLabs Creator-plan credits.

Stage 5: final pronunciation head. Retrained on encoder features extracted with the final ASR (consistent features for both consensus labels and master qari anchors).

Stage 6: tajweed rule engine. Pure-Python rule engine that takes the expected diacritized text, audio, and per-token alignment, producing per-letter tajweed feedback across 27 dispatched rules (noon sakinah, meem sakinah, madd typology, qalqalah sughra/kubra, ra tafkheem/tarqeeq, Allah lafdh, lam shamsiyyah, hamzat wasl, idgham types, leen letters, more).

---

Architecture

• Backbone: NVIDIA FastConformer Large encoder + CTC head (114.6 M params)
• Tokenizer: SentencePiece BPE, 1,024 vocab + 1 blank = 1,025 output classes
• Audio: 16 kHz, log-mel features (80 channels)
• Decoder: CTC greedy_batched (frame-independent, noise-robust, deterministic)
• Output: token log-probabilities and 512-dim encoder features per output frame

CTC is the right decoder for this task: frame-level alignment for the pronunciation pipeline, no auto-correction that would mask user mispronunciations, faster inference, and (per the decoder ablation) sub-1% WER on held-out studio reciters - so there's no quality reason to add decoder complexity that would also hide mispronunciations.

---

Quick start (CTC, ONNX)

import numpy as np, soundfile as sf, onnxruntime as ort, sentencepiece as spm

session = ort.InferenceSession("onnx/model_with_encoder.onnx",
                                providers=["CUDAExecutionProvider", "CPUExecutionProvider"])
sp = spm.SentencePieceProcessor(model_file="tokenizer.model")

wav, sr = sf.read("clip.wav")           # 16 kHz mono float32
features = log_mel(wav)                  # see tajweed/aligner.py for the pipeline
features = features[None, ...]           # (B=1, 80, T_in)
length = np.array([features.shape[2]], dtype=np.int64)

logprobs, encoder_features = session.run(
    ["logprobs", "encoder_output"],
    {"audio_signal": features, "length": length},
)
# encoder_features: (1, 512, T_out). Feed to the pronunciation head.
# logprobs:         (1, T_out, 1025). Argmax + CTC collapse to get tokens.

The full multi-signal scorer (CTC + head + anchor + GOP + tajweed) is in tajweed/full_scorer.py.

---

Files

Path	Description
nemo/fastconformer-quran.nemo	NeMo checkpoint, current (June 2026) weights (459 MB)
onnx/model.onnx	CTC-only ONNX, fp32 (437 MB)
onnx/model.fp16.onnx	CTC-only ONNX, fp16 (219 MB)
onnx/model_with_encoder.onnx	CTC + encoder features, fp32 (437 MB)
head/pronunciation_head.pt	Pronunciation head v7 (5.4 MB)
tajweed/	Python module: text analyzer, 27 rules, full scorer
tokenizer.model	SentencePiece tokenizer
model_config.yaml	NeMo model config
demo/	Three sample clips with known transcriptions

For iOS / CoreML deployment, see Muno459/fastconformer-quran-coreml.

---

Datasets

The current (June 2026) weights were trained on a fully canonical-imlaei labeled mix of:

• Tarteel EveryAyah (tarteel-ai/everyayah): CC-BY 4.0. Multi-reciter Hafs studio recitation, the clean backbone.
• Tarteel tlog (tarteel-ai/tlog): gated. Real user phone recordings (the real-world / robustness signal).
• Muaalem (obadx/muaalem-annotated-v3): additional clean Hafs recitation (~12 K clips), relabeled from its diacritized text to our consistent imlaei orthography.

All audio is Hafs riwayah. Labels are each clip's canonical ayah text in one imlaei orthography.

---

License

CC-BY-4.0, matching the upstream nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0 license this model is fine-tuned from (attribution to NVIDIA required). Previously listed as Apache-2.0 in error.

Citation

@misc{fastconformer-quran-2026,
  title  = {FastConformer-Quran: Quranic ASR and unsupervised mispronunciation scoring},
  author = {Anon},
  year   = {2026},
  url    = {https://huggingface.co/Muno459/fastconformer-quran},
}