FastConformer-Quran CoreML — Offline
iOS/macOS offline (full-utterance) Quranic ASR. Two variants — pick by your need:
| Variant | File | Precision | Device | WER* | When to use |
|---|---|---|---|---|---|
| Max accuracy | fastconformer-quran-offline.mlpackage |
fp32 | GPU/CPU | ~3% | best transcript, latency not critical |
| ANE (real-time) | fastconformer-quran-offline-ane.mlpackage |
fp16 | Neural Engine | ~6% | on-device, low-latency, battery-friendly |
*WER on a leakage-free held-out Quran set (EveryAyah reciters never trained on + QUL + real-phone tlog). For reference, the public Arabic-ASR-leaderboard #1 (nvidia FastConformer) scores ~5.7% on the same clips — both our variants are competitive, the fp32 one better.
Why two variants (the ANE precision story)
The Neural Engine is fp16-only, and the full-attention encoder blanks on ANE in fp16 (rounding accumulates over the whole utterance and erodes thin CTC margins → empty output on inputs like the Basmala). That can't be fixed by precision/quantization tricks. The fix is architectural: the ANE variant uses limited windowed attention att_context_size=[32,32] (each frame attends ±32 frames ≈ 6.5 s), which bounds the accumulation so it survives fp16 on ANE. It's fine-tuned at that window, costs a few WER points vs full attention, but runs correctly and real-time on ANE (and beats our cache-aware streaming model: ~6% vs ~10% WER on the same held-out set).
So: want maximum accuracy → fp32 (GPU). Want ANE acceleration → windowed fp16. Full attention on ANE is physically impossible (fp16 + whole-utterance accumulation → blank).
Both are multi-function (7 ANE entry points)
predict_T80 … predict_T4800 (0.8 s … 48 s), pick by padded audio length. Input (1,80,T) → logprobs (1,T/8,1025) + encoder_output. pos_enc clamp baked in.
Streaming variant
For live/low-latency recitation use the cache-aware Muno459/fastconformer-quran-coreml-streaming (validated on ANE, 5–8 ms/chunk).
License
This model is licensed under the Quran-Lab No-Profit License v1.0 (NPL-1.0). See LICENSE for the full terms. In short:
- No profit. You may not use this model, any derivative (fine-tune, quantization, distillation, re-export), or any service built on it to make money.
- Local use is free. Running it yourself must cost users nothing.
- Hosted services: cost recovery only. If you serve it over a network, you may recover only the strictly necessary direct running costs (compute, storage, bandwidth), never a profit.
- Share-alike. Every derivative must be released under this same license.
- Attribution. Credit
Muno459 / Quran-Lab, fastconformer-quran.
This replaces the previous Apache-2.0 license. The change is not retroactive: copies obtained earlier under Apache-2.0 keep those terms, but all use going forward is governed by NPL-1.0.
Upstream components this model builds on (NVIDIA FastConformer and the upstream Muno459/fastconformer-quran) were released under Apache-2.0 and retain their own Apache-2.0 terms for those original portions. The NPL-1.0 terms above apply to this model as distributed here.
Benchmark
Leakage-free held-out WER vs nvidia / whisper / seamless / mms / omniASR / Tarteel: Quranic ASR Leaderboard.
- Downloads last month
- 2