BranchShine-CR: Compact Multilingual IPA Transcription with Self-Conditioned CTC and Consistency Regularization

Summary

The paper presents BranchShine-CR, a 25.4M-parameter model trained from scratch for multilingual transcription into the International Phonetic Alphabet. It combines 80-bin log-mel features, a 12-block rotary-position E-Branchformer encoder, intermediate self-conditioned CTC supervision at blocks 6 and 10, and a consistency objective that penalizes symmetric KL divergence between two independently augmented views of each speed-perturbed utterance. Trained for 350k updates in 4.7 days on a single GPU using IPApack++ partition #3 (~1.63M utterances), it achieves 4.47% IPA character error rate on a matched test set of 16,646 utterances — a 22.3% relative reduction from ZIPA-CTC-NS (300M parameters) — and outperforms a similarly sized NeMo Conformer on all 41 language labels. A two-seed, nine-configuration ablation attributes the largest gains to consistency regularization (+2.05pp when removed) and two-view training; notably, removing prediction feedback slightly improves accuracy and RoPE has little effect. Exact match stays below ZIPA-CTC-NS (41.37% vs 45.15%). Limitations: no statistical significance testing or confidence intervals, single-seed main results, an ablation setup differing from the main training configuration, a capacity confound in the attention-only variant, no stated code or checkpoint release, and baseline comparisons built on saved predictions from potentially overlapping training data — a skew the authors acknowledge favors their model.