Dhor recitation model (Core ML)
Core ML build of the IqraEval wav2vec2 CTC baseline, used by the Dhor app to score Quran recitation against the canonical phoneme sequence on-device.
Two packages, because the halves need different precision:
| File | Contents | Precision |
|---|---|---|
encoder.mlpackage.zip |
frozen wav2vec2-base + learned 13-layer featurizer | int8 weights, must run CPU_ONLY |
head.mlpackage.zip |
projector -> 2-layer BiLSTM -> linear (71 CTC classes) | fp32 |
Input is 10 s of 16 kHz mono float32; output logits [1, 499, 71]. Quantised
weights with fp32 compute crash MPSGraph on the GPU path, hence CPU_ONLY for
the encoder.
Measured against the PyTorch reference on real recitation (speech frames only): per-frame log-prob error on the top-1 path has median 0.0000, p90 0.0036, p99 0.3976; 0.98% of frames exceed 0.5 nat.
Attribution
- Model weights: IqraEval/Iqra_wav2vec2_base, Apache-2.0
- Upstream encoder:
facebook/wav2vec2-base(Meta AI), Apache-2.0
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for kahk-group/dhor-recitation
Base model
IqraEval/Iqra_wav2vec2_base