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

• Direct IPA transcription offers a shared representation across languages for language documentation, pronunciation analysis, and cross-lingual speech research.Source: §1 · p. 1
• BranchShine-CR replaces the original model’s learned raw-waveform front end with log-mel features, a smaller encoder, intermediate CTC conditioning, and consistency-regularized training.The original BranchShine is retained as a reference [4].Source: §1 · p. 1
• The study compares six systems on identical IPApack++ test utterances and references, including a separate ablation study of individual components.BranchShine-CR combines a smaller parameter count with lower IPA-CER, while ZIPA-CTC-NS retains higher exact match.Source: §1 · p. 1

2.1. Multilingual phone recognition

• AlloVera and Allosaurus established shared multilingual phone-recognition resources and models.Later work explored compositional phone representations, language-specific inventories, and articulatory supervision.Source: §2.1 · p. 1
• ZIPA, POWSM, and PhoneticXEUS provide multilingual IPA, joint phonetic–orthographic, and self-conditioned-CTC reference systems for compact-recognizer comparisons [3] [12].MultiIPA investigated direct IPA transcription and cleaner multilingual training data; Wav2Vec2Phoneme combined multilingual representations with articulatory mappings.Source: §2.1 · p. 1

2.2. Encoder design and supervision

• The paper integrates established E-Branchformer, self-conditioned CTC, and consistency-regularization mechanisms in a compact, from-scratch multilingual IPA recognizer [15] [16] [17] [21].Its two-view objective uses independently sampled masking and dropout and penalizes final-posterior disagreement with symmetric KL, stopping gradients through each target distribution.Source: §2.2 · p. 1

3.1. Acoustic encoder and conditioning

• The model converts mono 16 kHz audio into normalized 80-bin log-mel features, then uses two stride-2 convolutions to reduce resolution and project features to 256 dimensions.Features use a 25 ms Hann window, 512-point FFT, and 10 ms hop.Source: §3.1 · p. 1
• Twelve RoPE E-Branchformer blocks combine parallel attention and convolutional-gating branches, with depthwise residual merging and projections back to 256 dimensions.Intermediate states at blocks 6 and 10 receive CTC projections whose posteriors are mapped back into the hidden representation; final logits share the same CTC projection, and inference uses greedy CTC decoding without a language model or beam search.Source: §3.1 · pp. 1–2, Figure 1 · p. 2

3.2. Two-view training objective

• Two independently masked views of the same speed-perturbed waveform pass through a shared network, and the training objective combines their CTC supervision with consistency regularization.The consistency term uses symmetric KL between final posteriors, stops gradients through the target distribution, and excludes padding frames.Source: §3.2 · p. 2
• Training applies independent SpecAugment and dropout to the two views, ramps the consistency weight to 0.2 over 2,000 updates, and trains from random initialization for 350,000 updates.Speed factors are 0.9, 1.0, and 1.1 when CTC-feasible; the run uses BF16, activation checkpointing, and an RTX PRO 5000 Blackwell GPU.Source: §3.2 · p. 2

4.1. Data and comparison systems

• All six systems share 16,646 test utterances (25.6 hours; 818,485 normalized reference characters) across 41 stored language labels.BranchShine-CR, original BranchShine, and NeMo have slightly different prepared training and development counts because of model-specific CTC-feasibility requirements.Source: §4.1 · p. 2
• ZIPA and PhoneticXEUS use saved predictions from training on IPApack++ splits 1–4, while the compact systems are trained from scratch on the canonical split.NeMo Conformer-CTC Medium has 30.53M parameters; training regimes differ across systems.Source: §4.1 · p. 3, Table 2 · p. 3
• ZIPA and PhoneticXEUS may overlap the test set, potentially skewing comparisons against the proposed models.The paper does not rule out overlap because these baselines were trained on the entire IPApack corpus, including splits 1–4.Source: §4.1 · p. 3

4.2. Transcription metrics

• IPA-CER and exact match evaluate normalized character strings, while PFER measures PanPhon feature-edit cost per reference phone as a supplementary diagnostic.PFER uses PanPhon 0.22.2 with additional NFD normalization for segmentation; unrecognized material is skipped by the parser, so parse-warning coverage is retained.Source: §4.2 · p. 3

5.1. Matched multilingual recognition

• 4.47% IPA-CER is a 22.3% relative reduction from ZIPA-CTC-NS with approximately one-twelfth as many parameters.CR makes 36,607 character edits.Source: §5.1 · p. 3

5.2. Language labels and edit types

• CR has lower IPA-CER than original BranchShine and NeMo on all 41 language labels, versus 31 for ZIPA-CTC-NS and 38 for PhoneticXEUS.Its unweighted mean of label-level error rates is 10.18%, compared with 14.36% for original BranchShine and 17.61% for NeMo.Source: §5.2 · p. 3
• Excluding both Tamil labels preserves the ordering: corpus IPA-CER is 4.59% for CR, 5.90% for ZIPA-CTC-NS, 7.39% for original BranchShine, and 8.94% for NeMo.The remaining test set contains 15,499 utterances, so the improvement is not confined to the two Tamil labels or the largest labels.Source: §5.2 · p. 3
• CR makes 17,805 substitutions, 7,253 insertions, and 11,549 deletions, each lower than original BranchShine, NeMo, and ZIPA-CTC-NS under identical deterministic alignment.The Fig. 3 bars sum to corpus IPA-CER.Source: §5.2 · p. 3, Figure 3 · p. 4

5.3. Ablation study

• Removing consistency increases mean CER by 2.05 percentage points, while removing the second view adds 1.82 points.Removing auxiliary supervision or enhanced merging costs 0.66 or 0.87 points; removing feedback improves CER by 0.21 and 0.34 points across seeds, and removing RoPE has little effect. Removing the local branch costs 1.80 points while reducing parameters from 25.40M to 17.09M.Source: §5.3 · p. 4

6. CONCLUSION

• 4.47% IPA-CER with 25.39M parameters represents a 22.3% relative reduction from ZIPA-CTC-NS and 37.2% from original BranchShine.The model is trained from scratch on the matched IPApack++ test-set task.Source: §6 · p. 4
• Reduced-budget ablations favor consistency regularization, while feedback removal slightly improves accuracy and RoPE has little effect.Table 3 reports final-step development CER at 18,000 updates for seeds 17 and 29.Source: §6 · p. 4, Table 3 · p. 4
• Component findings remain conditional on the ablation data and budget.Source: §6 · p. 4

8. COMPLIANCE WITH ETHICAL STANDARDS

• The study used open-access human-subject data retrospectively; the authors state that the attached license confirmed ethical approval was not required.Source: §8 · p. 4

9. CONFLICT OF INTEREST DISCLOSURE

• The authors report no relevant financial or nonfinancial interests to disclose.Source: §9 · p. 5