omni-ctc-ipa-v7-attraux-aux0p2-20260928

Wav2Vec2-CTC Quranic phoneme recognizer. Fine-tunes omni-ctc-ipa-v7 with an attribute auxiliary loss (AttrAux) at weight 0.2, part of the attraux-v7 research roadmap (Stage 1 v7-recipe foundation, 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 0.2 relative to the primary CTC loss, with attribute-group class rebalancing enabled (--attr-aux-group-rebalance).
• Best training-time eval PER: 1.807% (hetchyy/everyayah-phonemes dev split).

Insights

Lowest pooled PER of all 17 attraux-v7/baseline checkpoints evaluated on a professional-reciter benchmark suite (quranmd + tadabur + mufti_malhan + quranlab, 187,417 pooled phoneme units):

Metric (pooled, 4 professional sets)	Value
PER	6.02% (lowest of all 17 evaluated models)
Substitution / Deletion / Insertion	2.39% / 1.17% / 2.46%
Ghunna F1	97.05%
Qalqalah F1	95.33%
Tafkheem F1	97.85%
Gemination F1	98.03%
Sifat-core F1	97.99%
Macro F1 (all families)	97.67%

For context, its Macro F1 (97.67%) sits just under the two higher-F1 siblings from the same sweep — omni-ctc-ipa-v7-attraux-aux0p1-20260928 (PER 6.09%, F1 97.72%) and omni-ctc-ipa-v7-attraux-aux0p4-20260928 (PER 6.07%, F1 97.74%, also the highest Macro-F1 of all 17). This checkpoint is the pick if minimizing raw transcription error rate is the priority; the aux0p4 sibling is the pick if Tajweed-family F1 matters more. The gap between all three is small (<0.1pp F1, <0.1pp PER) and likely within run-to-run noise for this recipe.

Caveat — read before use as evidence of AttrAux working: the parent research roadmap (docs/roadmaps/attraux-v7/ROADMAP.md) evaluated this exact aux-weight (w=0.2, Stage 1) 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 AttrAux arm across the full sweep (w=0.05-0.4) produced a statistically significant LCP gain over control (overall roadmap KILL on 2026-09-29 after Stages 1-5' exhausted every tested AttrAux configuration, including hard-negative frame weighting, density oversampling, and FP-mined replay). The pooled-PER/tajweed-F1 numbers above measure professional-recitation transcription quality — where this is the single best-PER checkpoint tried — but should not be read as evidence that AttrAux training improved mispronunciation-detection precision; that hypothesis was tested directly and rejected.

In short: the best general-purpose Quranic phoneme transcriber by raw PER found in this sweep, ahead of its own base model (omni-ctc-ipa-v7: PER 6.37%, F1 97.50%) and every other attraux-v7/legacy checkpoint benchmarked, 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: 0.2
attr_aux_warmup_steps: 240
attr_aux_group_rebalance: true
seed: 42
fp16: true