omni-ctc-ipa-v7-attraux-aux0p1-20260928

Wav2Vec2-CTC Quranic phoneme recognizer. Fine-tunes omni-ctc-ipa-v7 with an attribute auxiliary loss (AttrAux) at weight 0.1, part of the attraux-v7 research roadmap (Stage 2 aux-weight dose sweep, seed 42, trained 2026-09-28).

Model description

• Architecture: Wav2Vec2ForCTC, 24 transformer layers, hidden size 1024, 71-token IPA-style phoneme vocabulary (Tajweed-aware: ghunna/tafkheem/gemination/qalqalah symbols included).
• Base checkpoint: omni-ctc-ipa-v7 (frozen feature encoder, transformer layers 0-11 frozen during this fine-tune, layers 12-23 trainable).
• Training data: hetchyy/everyayah-phonemes (207,114 max train rows), 2,400 steps, batch size 1 x grad-accum 16, LR 1e-5, 240 warmup steps, fp16.
• AttrAux mechanism: an auxiliary BCE loss over four pooled attribute channels (ghunna/tafkheem/gemination/qalqalah), derived by pooling the existing CTC logits (not a separate model head), weighted at 0.1 relative to the primary CTC loss, with attribute-group class rebalancing enabled (--attr-aux-group-rebalance).
• Best training-time eval PER: 1.970% (hetchyy/everyayah-phonemes dev split).

Insights

This checkpoint is one of two survivors (out of 17 evaluated attraux-v7 variants plus baselines) that beat the joint bar pooled PER < 6.10% AND pooled tajweed macro-F1 > 97.70% on a professional-reciter benchmark suite (quranmd + tadabur + mufti_malhan + quranlab, 187,417 pooled phoneme units):

Metric (pooled, 4 professional sets)	Value
PER	6.09%
Substitution / Deletion / Insertion	2.45% / 1.05% / 2.59%
Ghunna F1	97.16%
Qalqalah F1	95.64% (best qalqalah score across all 17 evaluated models)
Tafkheem F1	97.85%
Gemination F1	98.01%
Sifat-core F1	98.01%
Macro F1 (all families)	97.72%

Caveat — read before use as evidence of AttrAux working: the parent research roadmap (docs/roadmaps/attraux-v7/ROADMAP.md) evaluated this exact aux-weight sweep (w=0.05/0.1/0.4) against a matched no-aux control on the roadmap's primary metric — learner-correction precision (LCP) on a speaker-split QuranMB.v2 dev set with a paired bootstrap CI — and found no arm, including this one, produced a statistically significant LCP gain over control (Stage 2 verdict: REGRESSION/NO EFFECT for all three weights; overall roadmap KILL on 2026-09-29 after Stages 1-5' exhausted every tested AttrAux configuration). The pooled-PER/tajweed-F1 numbers above measure professional-recitation transcription quality, which this checkpoint is good at, but should not be read as evidence that its AttrAux training signal improved mispronunciation-detection precision — that hypothesis was tested and rejected on a different, dedicated benchmark.

In short: a solid general-purpose Quranic phoneme transcriber, marginally ahead of its own base model (omni-ctc-ipa-v7: PER 6.37%, F1 97.50%) and of every other attraux-v7 variant tried, but its AttrAux loss did not measurably improve mispronunciation-detection quality over a plain no-aux fine-tune at the same recipe.

Intended use

Quranic recitation phoneme transcription (near-offline/offline ASR) and Tajweed rule classification (ghunna/tafkheem/gemination/qalqalah) from audio. Not validated as a mispronunciation-detection/correction system — use omni-ctc-ipa-v7 or a control checkpoint from the same sweep for that use case pending further evidence.

Training hyperparameters

dataset: hetchyy/everyayah-phonemes
base_model: omni-ctc-ipa-v7
max_steps: 2400
batch_size: 1, grad_accum: 16
learning_rate: 1e-5, warmup_steps: 240
freeze_transformer_layers: 12 (of 24)
attr_aux_weight: 0.1
attr_aux_warmup_steps: 240
attr_aux_group_rebalance: true
seed: 42
fp16: true