BranchShine: Compact Raw-Audio-to-IPA Transcription with a RoPE E-Branchformer Encoder

Nikhil Navas; Saeed Afshar; Sergio Chevtchenko; Talisson Damiao

REVIEW 2 major objections 2 minor 28 references

A 33M-parameter raw-audio model reaches 9.19% IPA error rate, slightly ahead of a 575M-parameter baseline under matched conditions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-26 09:14 UTC pith:DQG5O7QE

load-bearing objection BranchShine shows a 33M model reaching 9.19% IPA CER versus 9.78% for a 575M baseline, but the comparison depends on unshown normalization details and the paper supplies almost no training or variance information. the 2 major comments →

arxiv 2606.22824 v1 pith:DQG5O7QE submitted 2026-06-22 cs.LG

BranchShine: Compact Raw-Audio-to-IPA Transcription with a RoPE E-Branchformer Encoder

Nikhil Navas , Sergio Chevtchenko , Talisson Damiao , Saeed Afshar This is my paper

classification cs.LG

keywords raw audioIPA transcriptionCTCE-Branchformermultilingual speechcompact modelphonetic recognitionRoPE

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces BranchShine as a compact CTC system that converts raw audio directly into IPA symbols. It pairs a lightweight convolutional front end with a 19-block RoPE E-Branchformer encoder to reach only 33 million parameters. On a 16,660-utterance test set spanning 41 language labels, the model records 9.19 percent whitespace-insensitive character error rate. This edges the 9.78 percent of the much larger PhoneticXEUS baseline when the same normalization and scoring rules are used for both. The results indicate that small raw-audio models can deliver competitive phonetic transcription performance.

Core claim

BranchShine demonstrates that a 33M-parameter raw-audio CTC recognizer built from a lightweight convolutional front end and a 19-block RoPE E-Branchformer encoder obtains 9.19% whitespace-insensitive IPA character error rate on a 16,660-utterance multilingual test set covering 41 language labels, compared with 9.78% for the 575M-parameter PhoneticXEUS baseline under matched normalization and scoring.

What carries the argument

The RoPE E-Branchformer encoder with 19 blocks, which processes convolutional features to produce CTC alignments for IPA symbol sequences.

Load-bearing premise

The normalization choices and scoring rules applied to BranchShine and the PhoneticXEUS baseline are matched closely enough to permit a fair head-to-head comparison of their IPA transcription accuracy.

What would settle it

Re-evaluate both models on the same test set after changing the normalization rules so that the whitespace-insensitive character error rate ordering between the 33M and 575M models reverses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

A compact raw-audio model can approach the IPA accuracy of systems that use an order of magnitude more parameters.
Direct waveform input supports competitive phonetic transcription without separate acoustic feature pipelines.
The model produces a more conservative profile on child speech, rejecting more incorrect readings than Whisper-Medium.
Multilingual coverage across 41 language labels remains feasible inside a 33M-parameter footprint.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The small size could allow deployment of IPA transcription on edge devices with limited memory.
The conservative error behavior might suit applications where false positives in pronunciation feedback are costly.
Hybrid systems pairing this model with larger ones could trade off precision and recall on different reading tasks.
The same encoder structure might transfer to other sequence labeling tasks that benefit from raw waveform input.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

BranchShine shows a 33M model reaching 9.19% IPA CER versus 9.78% for a 575M baseline, but the comparison depends on unshown normalization details and the paper supplies almost no training or variance information.

read the letter

BranchShine reports a 33M parameter raw-audio CTC model using a RoPE E-Branchformer encoder that gets 9.19% whitespace-insensitive IPA CER on a 16,660 utterance multilingual test set, compared to 9.78% for the 575M PhoneticXEUS. That's the headline result.

The paper applies an existing E-Branchformer architecture with rotary position embeddings to the task of direct raw audio to IPA transcription. It adds a lightweight convolutional front-end and uses CTC loss. The comparison to the larger baseline is direct, and there's an additional look at how it behaves on child speech reading tasks versus Whisper. This gives a sense of different operating profiles.

What stands out is the size reduction while staying competitive on the error metric. For applications where model size matters, like on-device tools, this operating point could be practical.

The soft spots are in the experimental reporting. The abstract mentions matched normalization and scoring but does not detail the exact steps or confirm they were applied identically to the baseline outputs. The stress-test concern about whether preprocessing like character inventory and diacritic handling is the same is reasonable given the information available. Without that, the small difference in error rate is difficult to attribute solely to the model. There are also no mentions of training hyperparameters, number of runs, or variance in the results, and no data split information.

This paper is aimed at researchers building efficient speech recognition systems for phonetic output, particularly in multilingual or low-resource settings. A reader looking for concrete numbers on compact IPA models would find the reported point useful.

The work shows clear thinking in focusing on a narrow but practical task and providing a baseline comparison. It deserves peer review to see if the methods section resolves the evaluation questions and to check the full experimental setup. I would recommend sending it for review rather than desk rejecting.

Referee Report

2 major / 2 minor

Summary. The paper introduces BranchShine, a 33M-parameter CTC model for raw-audio-to-IPA transcription consisting of a lightweight convolutional front-end and a 19-block RoPE E-Branchformer encoder. It claims that this compact architecture achieves a whitespace-insensitive IPA character error rate of 9.19% on a 16,660-utterance multilingual test set spanning 41 language labels, outperforming the 575M-parameter PhoneticXEUS baseline at 9.78% under matched normalization and scoring; a secondary analysis on child speech reading is also presented.

Significance. If the empirical comparison holds, the result establishes a competitive operating point for small-footprint IPA transcription, highlighting the viability of RoPE-augmented E-Branchformer encoders for efficiency-critical applications. The explicit parameter count and direct baseline comparison on a defined test set provide a concrete, falsifiable data point in the space of multilingual phonetic recognizers.

major comments (2)

[§4] §4 (Evaluation Protocol): the central claim that BranchShine is 'competitive' rests on the 9.19% vs. 9.78% whitespace-insensitive CER gap, yet the manuscript provides no explicit, reproducible enumeration of the normalization steps (character inventory, diacritic handling, whitespace removal, language-label filtering) applied identically to both BranchShine outputs and the PhoneticXEUS baseline; without this protocol the numerical difference is not diagnostic of model quality.
[§3.2, §5] §3.2 and §5 (Training and Results): no training hyperparameters, data-split details, or variance estimates across runs are reported for the 9.19% figure, leaving the soundness of the headline comparison only moderately supported as noted in the abstract's error-rate claims.

minor comments (2)

[Table 1] Table 1: the parameter count for PhoneticXEUS is listed as 575.00M; confirm whether this includes all components or only the encoder, for consistency with the 33M count given for BranchShine.
[Abstract, §5] Abstract and §5: the secondary child-speech analysis is mentioned but lacks quantitative metrics or a table; either expand or move to supplementary material to avoid dangling claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on reproducibility. We address each major comment below and will incorporate clarifications into the revised manuscript.

read point-by-point responses

Referee: [§4] §4 (Evaluation Protocol): the central claim that BranchShine is 'competitive' rests on the 9.19% vs. 9.78% whitespace-insensitive CER gap, yet the manuscript provides no explicit, reproducible enumeration of the normalization steps (character inventory, diacritic handling, whitespace removal, language-label filtering) applied identically to both BranchShine outputs and the PhoneticXEUS baseline; without this protocol the numerical difference is not diagnostic of model quality.

Authors: We agree that an explicit enumeration is required for full reproducibility. In the revised manuscript we will add a new subsection to §4 that lists the complete normalization protocol applied identically to both systems: the exact IPA character inventory, diacritic handling rules, whitespace removal procedure, and language-label filtering criteria. This will make the 9.19% vs. 9.78% comparison fully diagnostic. revision: yes
Referee: [§3.2, §5] §3.2 and §5 (Training and Results): no training hyperparameters, data-split details, or variance estimates across runs are reported for the 9.19% figure, leaving the soundness of the headline comparison only moderately supported as noted in the abstract's error-rate claims.

Authors: We will expand §3.2 to report the full training hyperparameters (optimizer, learning-rate schedule, batch size, epochs) and data-split statistics (utterance counts and language distribution per split). Variance estimates across independent runs were not computed in the original experiments owing to the substantial compute required for the 41-language training set; we will state this limitation explicitly so readers understand the reported figure is a single-run point estimate. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical model comparison against external baseline

full rationale

The paper reports an empirical result: training a 33M-parameter CTC model on raw audio and measuring whitespace-insensitive IPA CER (9.19%) against an external 575M-parameter PhoneticXEUS baseline (9.78%) on a fixed 16,660-utterance test set. No derivation, equation, or claim reduces a reported quantity to a fitted parameter or self-citation by construction. The comparison is presented as a direct head-to-head measurement under asserted matched normalization; any concern about protocol identity is a reproducibility question, not a circularity reduction. The work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The performance claim rests on standard CTC training assumptions and the authors' choice of a 19-block encoder plus convolutional front-end configuration; these architectural decisions function as free parameters tuned to reach the reported operating point.

free parameters (2)

encoder block count = 19
The choice of exactly 19 blocks is part of the lightweight design that yields the 33M parameter total.
convolutional front-end hyperparameters
Lightweight convolutional parameters are selected to process raw audio before the encoder.

axioms (1)

domain assumption CTC loss supports end-to-end sequence transcription without explicit alignments
Invoked for the raw-audio-to-IPA recognizer as the standard loss for unaligned ASR.

pith-pipeline@v0.9.1-grok · 5725 in / 1337 out tokens · 41601 ms · 2026-06-26T09:14:11.391619+00:00 · methodology

0 comments

read the original abstract

Speech-to-IPA transcription is useful when the desired output is pronunciation rather than orthographic text, but competitive multilingual systems are often large and evaluation is sensitive to normalization choices. This paper presents BranchShine, a 33M-parameter raw-audio CTC recognizer with a lightweight convolutional front end and a 19-block RoPE E-Branchformer encoder. We find that BranchShine provides a compact and competitive operating point for IPA transcription under matched normalization and scoring. On a 16,660-utterance multilingual test set covering 41 language labels, BranchShine obtains 9.19% whitespace-insensitive IPA character error rate, compared with 9.78% for the 575.00M-parameter PhoneticXEUS baseline. A secondary child speech reading analysis shows a complementary operating profile: BranchShine is more conservative on incorrect readings, while Whisper-Medium is stronger on exact acceptance of correct readings. Overall, the results indicate that a compact raw-audio-to-IPA model can approach much larger baselines on character-level IPA transcription.

Figures

Figures reproduced from arXiv: 2606.22824 by Nikhil Navas, Saeed Afshar, Sergio Chevtchenko, Talisson Damiao.

**Figure 2.** Figure 2: Parameter efficiency in the main multilingual comparison. BranchShine is much smaller than PhoneticXEUS and lower on IPA-CER, while ZIPA-CTC-NS remains the best overall system. evaluation has not been ruled out. The central size-aware comparison is with PhoneticXEUS. BranchShine obtains 9.19% IPA-CER versus 9.78% for PhoneticXEUS while using 5.8% as many parameters. This result is meaningful but metric-de… view at source ↗

**Figure 3.** Figure 3: Edit-operation decomposition for BranchShine and PhoneticXEUS. BranchShine has fewer insertions and deletions but more substitutions, explaining the difference between IPA-CER and exact-match behavior. robustness claim across labels or durations. 5.3 Ablation evidence [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Edit savings for BranchShine relative to PhoneticXEUS, decomposed by language label and duration bin. Positive values favor BranchShine; negative values favor PhoneticXEUS [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Observed evaluation-error trajectories over the available training window for each compact variant. The dashed line marks 150k steps, where BranchShine is already lower than the absolute-position E-Branchformer and RoPE Transformer are after 250k and 350k steps, respectively. especially effective at avoiding insertions and deletions, but it makes more substitutions than PhoneticXEUS. On the child speech re… view at source ↗

**Figure 6.** Figure 6: Correct-reading exact match versus incorrect-reading mismatch on the child speech reading benchmark. BranchShine is the most conservative row shown, while Whisper-Medium has the highest correct-reading exact match. Michael Auli. 2022. XLS-R: Self-supervised cross-lingual speech representation learning at scale. Interspeech. Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 7 canonical work pages · 3 internal anchors

[1]

Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2022. XLS-R: Self-supervised cross-lingual speech representation learning at scale. Interspeech

2022
[2]

Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems

2020
[3]

Mortensen

Shikhar Bharadwaj, Chin-Jou Li, Kwanghee Choi, Eunjung Yeo, William Chen, Shinji Watanabe, and David R. Mortensen. 2026. An empirical recipe for universal phone recognition. arXiv:2603.29042

work page internal anchor Pith review arXiv 2026
[4]

Mortensen

Shikhar Bharadwaj, Chin-Jou Li, Yoonjae Kim, Kwanghee Choi, Eunjung Yeo, Ryan Soh-Eun Shim, Hanyu Zhou, Brendon Boldt, Karen Rosero Jacome, Kalvin Chang, Darsh Agrawal, Keer Xu, Chao-Han Huck Yang, Jian Zhu, Shinji Watanabe, and David R. Mortensen. 2026. PRiSM: Benchmarking phone realization in speech models. arXiv:2601.14046

work page internal anchor Pith review arXiv 2026
[5]

William Chen, Wangyou Zhang, Yifan Peng, Xinjian Li, Jinchuan Tian, Jiatong Shi, Xuankai Chang, Soumi Maiti, Karen Livescu, and Shinji Watanabe. 2024. Towards robust speech representation learning for thousands of languages. Proceedings of EMNLP

2024
[6]

Kevin Glocker, Aaricia Herygers, and Munir Georges. 2023. Allophant: Cross-lingual phoneme recognition with articulatory attributes. Interspeech, pages 2258--2262

2023
[7]

Alex Graves, Santiago Fernandez, Faustino Gomez, and J\"urgen Schmidhuber. 2006. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. Proceedings of ICML, pages 369--376

2006
[8]

Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. 2020. Conformer: Convolution-augmented Transformer for speech recognition. Interspeech

2020
[9]

Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing

2021
[10]

Nat Jeffries, Evan King, Manjunath Kudlur, Guy Nicholson, James Wang, and Pete Warden. 2024. Moonshine: Speech recognition for live transcription and voice commands. arXiv:2410.15608

work page arXiv 2024
[11]

Han, and Shinji Watanabe

Kwangyoun Kim, Felix Wu, Yifan Peng, Jing Pan, Prashant Sridhar, Kyu J. Han, and Shinji Watanabe. 2022. E-Branchformer: Branchformer with enhanced merging for speech recognition. arXiv:2210.00077

work page arXiv 2022
[12]

Black, and Florian Metze

Xinjian Li, Siddharth Dalmia, Juncheng Li, Matthew Lee, Jiahong Yuan, Wei-Ning Hsu, Yao-Hung Hubert Tsai, Alan W. Black, and Florian Metze. 2020. Universal phone recognition with a multilingual allophone system. ICASSP

2020
[13]

Xinjian Li, Juncheng Li, Florian Metze, and Alan W. Black. 2021. Hierarchical phone recognition with compositional phonetics. Interspeech, pages 2461--2465

2021
[14]

Mortensen, Alan W

Xinjian Li, Florian Metze, David R. Mortensen, Alan W. Black, and Shinji Watanabe. 2022. Phone inventories and recognition for every language. Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1061--1067

2022
[15]

Mortensen, and Shinji Watanabe

Chin-Jou Li, Kalvin Chang, Shikhar Bharadwaj, Eunjung Yeo, Kwanghee Choi, Jian Zhu, David R. Mortensen, and Shinji Watanabe. 2025. POWSM: A phonetic open Whisper-style speech foundation model. arXiv:2510.24992

work page arXiv 2025
[16]

Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. International Conference on Learning Representations

2019
[17]

Mortensen, Siddharth Dalmia, and Patrick Littell

David R. Mortensen, Siddharth Dalmia, and Patrick Littell. 2016. PanPhon: A resource for mapping IPA segments to articulatory feature vectors. Proceedings of COLING

2016
[18]

Mortensen, Xinjian Li, Patrick Littell, Alexis Michaud, Shruti Rijhwani, Antonios Anastasopoulos, Alan W

David R. Mortensen, Xinjian Li, Patrick Littell, Alexis Michaud, Shruti Rijhwani, Antonios Anastasopoulos, Alan W. Black, Florian Metze, and Graham Neubig. 2020. AlloVera: A multilingual allophone database. Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 5329--5336

2020
[19]

Yifan Peng, Siddhant Arora, Yosuke Higuchi, Yifan Shao, Jiatong Shi, Xuankai Chang, and Shinji Watanabe. 2022. Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding. Proceedings of ICML

2022
[20]

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. Proceedings of ICML

2023
[21]

Ahn, Shreya Prakash, M \'a rton Soskuthy, Vered Shwartz, and Jian Zhu

Farhan Samir, Emily P. Ahn, Shreya Prakash, M \'a rton Soskuthy, Vered Shwartz, and Jian Zhu. 2025. A comparative approach for auditing multilingual phonetic transcript archives. Transactions of the Association for Computational Linguistics, 13:595--612

2025
[22]

Kathleen Siminyu, Kelly Davis, Gayatri Bhat, and Alan W. Black. 2021. Phoneme recognition through fine tuning of phonetic representations: A case study on a low-resource language. Interspeech

2021
[23]

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. 2021. RoFormer: Enhanced Transformer with rotary position embedding. arXiv:2104.09864

work page internal anchor Pith review Pith/arXiv arXiv 2021
[24]

Chihiro Taguchi, Yusuke Sakai, Parisa Haghani, and David Chiang. 2023. Universal automatic phonetic transcription into the International Phonetic Alphabet. Interspeech, pages 2548--2552

2023
[25]

Gomez, ukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems

2017
[26]

Qiantong Xu, Alexei Baevski, and Michael Auli. 2022. Simple and effective zero-shot cross-lingual phoneme recognition. Interspeech, pages 2113--2117

2022
[27]

Saierdaer Yusuyin, Te Ma, Hao Huang, Wenbo Zhao, and Zhijian Ou. 2024. Whistle: Data-efficient multilingual and crosslingual speech recognition via weakly phonetic supervision. arXiv:2406.02166

work page arXiv 2024
[28]

Mortensen

Jian Zhu, Farhan Samir, Eleanor Chodroff, and David R. Mortensen. 2025. ZIPA: A family of efficient models for multilingual phone recognition. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics

2025

[1] [1]

Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2022. XLS-R: Self-supervised cross-lingual speech representation learning at scale. Interspeech

2022

[2] [2]

Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems

2020

[3] [3]

Mortensen

Shikhar Bharadwaj, Chin-Jou Li, Kwanghee Choi, Eunjung Yeo, William Chen, Shinji Watanabe, and David R. Mortensen. 2026. An empirical recipe for universal phone recognition. arXiv:2603.29042

work page internal anchor Pith review arXiv 2026

[4] [4]

Mortensen

Shikhar Bharadwaj, Chin-Jou Li, Yoonjae Kim, Kwanghee Choi, Eunjung Yeo, Ryan Soh-Eun Shim, Hanyu Zhou, Brendon Boldt, Karen Rosero Jacome, Kalvin Chang, Darsh Agrawal, Keer Xu, Chao-Han Huck Yang, Jian Zhu, Shinji Watanabe, and David R. Mortensen. 2026. PRiSM: Benchmarking phone realization in speech models. arXiv:2601.14046

work page internal anchor Pith review arXiv 2026

[5] [5]

William Chen, Wangyou Zhang, Yifan Peng, Xinjian Li, Jinchuan Tian, Jiatong Shi, Xuankai Chang, Soumi Maiti, Karen Livescu, and Shinji Watanabe. 2024. Towards robust speech representation learning for thousands of languages. Proceedings of EMNLP

2024

[6] [6]

Kevin Glocker, Aaricia Herygers, and Munir Georges. 2023. Allophant: Cross-lingual phoneme recognition with articulatory attributes. Interspeech, pages 2258--2262

2023

[7] [7]

Alex Graves, Santiago Fernandez, Faustino Gomez, and J\"urgen Schmidhuber. 2006. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. Proceedings of ICML, pages 369--376

2006

[8] [8]

Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. 2020. Conformer: Convolution-augmented Transformer for speech recognition. Interspeech

2020

[9] [9]

Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing

2021

[10] [10]

Nat Jeffries, Evan King, Manjunath Kudlur, Guy Nicholson, James Wang, and Pete Warden. 2024. Moonshine: Speech recognition for live transcription and voice commands. arXiv:2410.15608

work page arXiv 2024

[11] [11]

Han, and Shinji Watanabe

Kwangyoun Kim, Felix Wu, Yifan Peng, Jing Pan, Prashant Sridhar, Kyu J. Han, and Shinji Watanabe. 2022. E-Branchformer: Branchformer with enhanced merging for speech recognition. arXiv:2210.00077

work page arXiv 2022

[12] [12]

Black, and Florian Metze

Xinjian Li, Siddharth Dalmia, Juncheng Li, Matthew Lee, Jiahong Yuan, Wei-Ning Hsu, Yao-Hung Hubert Tsai, Alan W. Black, and Florian Metze. 2020. Universal phone recognition with a multilingual allophone system. ICASSP

2020

[13] [13]

Xinjian Li, Juncheng Li, Florian Metze, and Alan W. Black. 2021. Hierarchical phone recognition with compositional phonetics. Interspeech, pages 2461--2465

2021

[14] [14]

Mortensen, Alan W

Xinjian Li, Florian Metze, David R. Mortensen, Alan W. Black, and Shinji Watanabe. 2022. Phone inventories and recognition for every language. Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1061--1067

2022

[15] [15]

Mortensen, and Shinji Watanabe

Chin-Jou Li, Kalvin Chang, Shikhar Bharadwaj, Eunjung Yeo, Kwanghee Choi, Jian Zhu, David R. Mortensen, and Shinji Watanabe. 2025. POWSM: A phonetic open Whisper-style speech foundation model. arXiv:2510.24992

work page arXiv 2025

[16] [16]

Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. International Conference on Learning Representations

2019

[17] [17]

Mortensen, Siddharth Dalmia, and Patrick Littell

David R. Mortensen, Siddharth Dalmia, and Patrick Littell. 2016. PanPhon: A resource for mapping IPA segments to articulatory feature vectors. Proceedings of COLING

2016

[18] [18]

Mortensen, Xinjian Li, Patrick Littell, Alexis Michaud, Shruti Rijhwani, Antonios Anastasopoulos, Alan W

David R. Mortensen, Xinjian Li, Patrick Littell, Alexis Michaud, Shruti Rijhwani, Antonios Anastasopoulos, Alan W. Black, Florian Metze, and Graham Neubig. 2020. AlloVera: A multilingual allophone database. Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 5329--5336

2020

[19] [19]

Yifan Peng, Siddhant Arora, Yosuke Higuchi, Yifan Shao, Jiatong Shi, Xuankai Chang, and Shinji Watanabe. 2022. Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding. Proceedings of ICML

2022

[20] [20]

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. Proceedings of ICML

2023

[21] [21]

Ahn, Shreya Prakash, M \'a rton Soskuthy, Vered Shwartz, and Jian Zhu

Farhan Samir, Emily P. Ahn, Shreya Prakash, M \'a rton Soskuthy, Vered Shwartz, and Jian Zhu. 2025. A comparative approach for auditing multilingual phonetic transcript archives. Transactions of the Association for Computational Linguistics, 13:595--612

2025

[22] [22]

Kathleen Siminyu, Kelly Davis, Gayatri Bhat, and Alan W. Black. 2021. Phoneme recognition through fine tuning of phonetic representations: A case study on a low-resource language. Interspeech

2021

[23] [23]

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. 2021. RoFormer: Enhanced Transformer with rotary position embedding. arXiv:2104.09864

work page internal anchor Pith review Pith/arXiv arXiv 2021

[24] [24]

Chihiro Taguchi, Yusuke Sakai, Parisa Haghani, and David Chiang. 2023. Universal automatic phonetic transcription into the International Phonetic Alphabet. Interspeech, pages 2548--2552

2023

[25] [25]

Gomez, ukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems

2017

[26] [26]

Qiantong Xu, Alexei Baevski, and Michael Auli. 2022. Simple and effective zero-shot cross-lingual phoneme recognition. Interspeech, pages 2113--2117

2022

[27] [27]

Saierdaer Yusuyin, Te Ma, Hao Huang, Wenbo Zhao, and Zhijian Ou. 2024. Whistle: Data-efficient multilingual and crosslingual speech recognition via weakly phonetic supervision. arXiv:2406.02166

work page arXiv 2024

[28] [28]

Mortensen

Jian Zhu, Farhan Samir, Eleanor Chodroff, and David R. Mortensen. 2025. ZIPA: A family of efficient models for multilingual phone recognition. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics

2025