Pith. sign in

REVIEW 4 major objections 4 minor 42 references

PIER: A Novel Metric for Evaluating What Matters in Code-Switching

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read Fine-tuning on monolingual data can improve WER while the actual code-switched words get worse, and a word-restricted metric called PIER is the way to see it.

desk verdict The PIER metric is a useful diagnostic, but the paper's headline divergence between WER and PIER is currently confounded by comparing different evaluation subsets. read the letter →

arxiv 2501.09512 v2 pith:TGY7PCWW submitted 2025-01-16 cs.CL cs.LG

classification cs.CLcs.LG
keywords code-switchingautomaticspeechrecognitionevaluationmetricworderrorratePIERfine-tuningmultilingualASRintra-word
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that standard Word Error Rate and Character Error Rate are misleading as headline metrics for code-switched speech recognition, because the dominant matrix language drowns out errors on the few embedded-language words. It introduces PIER, a WER-style metric that counts edit operations only at tagged positions of interest, and uses it to show that fine-tuning on non-code-switched monolingual data from both languages improves WER/CER on Spanish-English, Arabic-English, and Mandarin-English code-switching test sets while the error rate on the code-switched words themselves gets worse. If the claim holds, PIER gives a practical way to separate genuine code-switching ability from monolingual strength, and the large gap between PIER and WER indicates that current models are much worse at code-switching than ordinary WER suggests.

What carries the argument

The central object is PIER, the Point-of-Interest Error Rate, a variant of WER in which the reference and hypothesis are aligned and only edit operations whose source position lies in a pre-specified set I of point-of-interest words are counted, normalized by the size of I. It stays as close as possible to WER so that behavior is interpretable, while its purpose is to isolate errors at code-switched positions, typically embedded-language words. By tagging words as inter-word or intra-word switches, PIER also decomposes code-switching errors into these two types; for language pairs with different scripts the tags are obtained automatically, and for German-English DECM the paper supplies manual annotations.

What would settle it

Recompute WER on exactly the code-switched-only utterances used for the PIER table, with monolingual utterances removed, and check whether WER still improves after the same monolingual fine-tuning; if WER on that restricted subset also improves while PIER worsens, the claimed mismatch survives, and if WER on that subset does not improve, the mismatch is an artifact of comparing different test subsets.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that classical WER/CER can improve on a code-switching test set even as the words that actually switch languages are recognized worse, and that the PIER metric exposes this. Across Whisper large, Whisper small, and the CTC-based MMS model, fine-tuning on monolingual utterances from both matrix and embedded language typically lowers full-test-set WER or CER on Fisher, Arzen, and SEAME, while PIER on the embedded-language words rises in the same configurations, for example the agnostic Whisper large improving 5.06% relative WER on Fisher while its PIER worsens 39.47%. PIER also quantifies that intra-word code-switching, where a single word mixes morphemes from two languages, has higher error rates than inter-word switching on DECM and Arzen.

Load-bearing premise

The comparison assumes the WER and PIER numbers are computed on the same test utterances; PIER is reported after excluding monolingual utterances from the code-switching test sets, while the WER values it is contrasted with come from the full test sets, so a WER improvement driven by monolingual utterances would not contradict PIER's worsening.

Editorial extensions

If this is right

  • If PIER is adopted, a model that merely improves recognition of the matrix language will no longer look better on code-switched data; the metric will credit only progress on switched words.
  • Reported PIER is consistently higher than WER on every dataset and model configuration, so the code-switched positions are a bottleneck that full-set WER understates.
  • Since fine-tuning on non-code-switched monolingual data worsens PIER while improving WER, code-switching performance needs either code-switched training data or a method that targets the switched positions explicitly.
  • PIER with coarse inter-word versus intra-word tagging shows intra-word code-switching is substantially harder, so systems and benchmarks should report these two error types separately.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same point-of-interest normalization could be applied to other rare or semantically load-bearing word classes in ASR, such as named entities, numbers, or dialect words, whenever those positions can be tagged, though the paper does not discuss those extensions.
  • Because PIER is computed over a small number of interest words per utterance, its per-utterance values will be noisy; reporting confidence intervals or significance tests is a natural extension the paper leaves implicit.
  • A sharper test of the paper's thesis would be to fine-tune on synthetic code-switched utterances obtained by splicing monolingual segments; PIER predicts that genuine switching ability would then improve alongside WER, separating the 'harder test subset' effect from the 'monolingual fine-tuning degrades mixing' effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper argues that WER is insufficient for evaluating code-switched ASR, because it is dominated by matrix-language words and can improve even when code-switched words get worse. It proposes Point-of-Interest Error Rate (PIER), a WER variant restricted to tagged words of interest, and instantiates it with embedded-language words. Experiments with Whisper-large, Whisper-small, and MMS on Fisher, Arzen, SEAME, and DECM claim to show that fine-tuning on monolingual data improves WER while PIER worsens, and that intra-word code-switching is harder than inter-word. The paper also provides an open-source implementation.

Significance. If the empirical claims hold, PIER is a simple, interpretable, and potentially useful complement to WER for code-switching research, and the fine-grained inter-word versus intra-word analysis is a useful addition. The paper has concrete strengths: it covers several language pairs and model architectures, it attempts to study both inter-word and intra-word switching, and it ships code. However, the central empirical demonstration is currently confounded by a mismatch between the evaluation subsets used for WER and PIER, so the significance of the claimed divergence is not yet established.

major comments (4)
  1. [Section V and Tables III/IV] Section V states that "we excluded the monolingual utterances from the code-switching test set" when computing PIER, but Table III reports WER values on the full test sets with no indication that a similar exclusion was applied. Consequently, the claimed divergence between improving WER and worsening PIER could simply be an artifact of comparing WER on all utterances with PIER on code-switched-only utterances. The authors should report WER on the same code-switched-only subset for both base and fine-tuned models; without this, the headline result is unverified.
  2. [Section IV, Eq. (1)] The definition of PIER relies on an alignment A and on the source position i_src of each edit operation, but the alignment algorithm is not specified. Standard Levenshtein alignments are not unique, and the treatment of insertions is unclear: the rule that insertions after the final reference word are counted only when the final reference word is a point of interest appears ad hoc, and the definition does not explain how insertions at other positions are mapped to reference-word indexes. Without a precise alignment protocol, PIER is not well-defined and the reported numbers are not reproducible.
  3. [Tables III and IV] No error bars, confidence intervals, significance tests, or multiple random seeds are reported, and the pattern across the tables is not uniform. For example, W-large in the X decoding improves both WER (30.57 to 28.01) and PIER (59.92 to 51.16) on Fisher, and W-large EN on Arzen improves both WER (103.81 to 102.81) and PIER (38.77 to 38.72). The paper's general claim that fine-tuning improves WER while PIER worsens therefore needs statistical support and an acknowledgment of the counterexamples.
  4. [Section V, point-of-interest tagging] The procedure for identifying English words as points of interest is not described for same-script data. For Fisher (Spanish-English), language identification cannot be done by script, yet the paper does not state whether it used a language-identification tool, manual annotation, or another method. This matters because PIER values depend entirely on the tagged word set; the authors should specify the procedure and, if annotation was manual, report inter-annotator agreement.
minor comments (4)
  1. [Table V] The Arzen row for W-large agnostic is formatted as "49.50, 105.71" with a comma rather than as two separate table cells, which makes the table difficult to read.
  2. [Throughout] The text contains typographical and formatting errors, including "its quite surprising" in Section III-A, "SotA" instead of "state-of-the-art", and the reference entry [12] which types "Waibe11" instead of "Waibel".
  3. [Related work] The closely related PolyWER metric [16] is cited but not compared or discussed; a short comparison would help position PIER relative to existing code-switching evaluation metrics.
  4. [Section II] Fine-tuning details are underspecified: the paper mentions early stopping after five consecutive non-improving evaluations, but does not report learning rates, numbers of update steps, batch sizes, or the amount of sampled fine-tuning data per experiment.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; PIER is a definitional metric and self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained: PIER is explicitly defined in Section IV as a WER restricted to tagged points of interest, and the fine-tuning experiments use public monolingual data and standard models. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. The self-citations to DECM [23] and earlier code-switching papers [20], [21] are used as data sources or background, not as load-bearing justifications for the metric. The main empirical claim—that WER improves while PIER worsens—is a measurement, not a quantity forced by construction. A separate validity concern, noted in Section V, is that PIER excludes monolingual utterances from the code-switching test set while Table III WER is reported on the full test sets; this confound is a correctness risk for the headline comparison, but it is not circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper introduces no fitted free parameters and no new entities. The evaluation rests on alignment and point-of-interest tagging assumptions that are only partially specified.

assumptions (3)
  • standard math Edit-distance alignment between reference and hypothesis is a faithful representation of ASR errors.
    PIER counts aligned edit operations, so any alignment ambiguity alters the PIER value.
  • domain assumption Points of interest, here embedded-language words, can be reliably identified in test references.
    For Fisher, both languages share the Latin script, and the tagging relies on dataset annotations or an undescribed automatic tagger.
  • ad hoc to paper Insertions after the last reference word are counted only when the final reference word is a point of interest.
    This edge-case rule in the PIER definition in Section IV is not given a linguistic or statistical justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PIER: A Novel Metric for Evaluating What Matters in Code-Switching." pith.science (2026). https://pith.science/paper/TGY7PCWW

@misc{pith2026250109512,
  author       = {Pith},
  title        = {Pith review of: PIER: A Novel Metric for Evaluating What Matters in Code-Switching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TGY7PCWW}},
  note         = {Machine review of arXiv:2501.09512}
}
read the original abstract

Code-switching, the alternation of languages within a single discourse, presents a significant challenge for Automatic Speech Recognition. Despite the unique nature of the task, performance is commonly measured with established metrics such as Word-Error-Rate (WER). However, in this paper, we question whether these general metrics accurately assess performance on code-switching. Specifically, using both Connectionist-Temporal-Classification and Encoder-Decoder models, we show fine-tuning on non-code-switched data from both matrix and embedded language improves classical metrics on code-switching test sets, although actual code-switched words worsen (as expected). Therefore, we propose Point-of-Interest Error Rate (PIER), a variant of WER that focuses only on specific words of interest. We instantiate PIER on code-switched utterances and show that this more accurately describes the code-switching performance, showing huge room for improvement in future work. This focused evaluation allows for a more precise assessment of model performance, particularly in challenging aspects such as inter-word and intra-word code-switching.

Figures

Figures reproduced from arXiv: 2501.09512 by the authors.

Figure 1
Figure 1. Comparison of WER and PIER behaviour after fine [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 35 canonical work pages

  1. [1]

    & Others Code Switching an d Code Mixing in Teaching and Learning of English as a Second Langua ge: Building on Knowledge

    Ezeh, N., Umeh, I., Anyanwu, E. & Others Code Switching an d Code Mixing in Teaching and Learning of English as a Second Langua ge: Building on Knowledge. English Language Teaching . 15, 106-106 (2022)

  2. [2]

    & Aji, A

    Zhang, R., Cahyawijaya, S., Cruz, J., Winata, G. & Aji, A. Multilingual large language models are not (yet) code-switchers. ArXiv Preprint arXiv:2305.14235. (2023)

  3. [3]

    & Watanabe, S

    Y an, B., Wiesner, M., Klejch, O., Jyothi, P . & Watanabe, S . Towards zero-shot code-switched speech recognition. ICASSP 2023-2023 IEEE International Conference On Acoustics, Speech And Signal P rocessing (ICASSP). pp. 1-5 (2023)

  4. [4]

    & Niesler , T

    Biswas, A., Yılmaz, E., Westhuizen, E., Wet, F. & Niesler , T. Code- switched automatic speech recognition in five South African languages. Computer Speech & Language . 71 pp. 101262 (2022)

  5. [5]

    & Ury, W

    Scotton, C. & Ury, W. Bilingual strategies: The social fu nctions of code- switching. (Walter de Gruyter, Berlin/New Y ork Berlin, New Y ork,1977)

  6. [6]

    A Survey of Code-switching: Linguistic and Social Perspectives for Language Technologies

    Do˘ gru¨ oz, A., Sitaram, S., Bullock, B. & Toribio, A. A su rvey of code- switching: Linguistic and social perspectives for languag e technologies. ArXiv Preprint arXiv:2301.01967 . (2023)

  7. [7]

    & Waibel, A

    Ugan, E., Mediani, M., Al Jawabra, O., Khader, A., Liu, Y . & Waibel, A. Modular Design of a Front-End and Back-End Speech-to-Spe ech Translation Application for Psychiatric Treatment of Refu gees. 2023 IEEE Global Humanitarian Technology Conference (GHTC) . pp. 128- 131 (2023)

  8. [8]

    & Torres- Smith, K

    Hourin, S., Binder, J., Y aeger, D., Gamerdinger, C., Wil son, K. & Torres- Smith, K. undefined. SPEECH-TO-SPEECH TRANSLATION TOOL (S2ST2) Limited Utility Assessment Final Report . pp. 1-77 (2013)

Show all 42 references
  1. [9]

    & Fuegen, C

    Waibel, A. & Fuegen, C. Simultaneous translation of open domain lectures and speeches. (2012), US Patent 8,090,570

  2. [10]

    Speech translators for humanitarian projec ts

    Waibel, A. Speech translators for humanitarian projec ts. Spoken Lan- guage Technologies for Under-Resourced Languages . pp. 4-5 (2010)

  3. [11]

    & Others End-to- End Evaluation for Low-Latency Simultaneous Speech Translati on

    Huber, C., Dinh, T., Mullov, C., Pham, N., Nguyen, T., Re tkowski, F., Constantin, S., Ugan, E., Liu, D., Li, Z. & Others End-to- End Evaluation for Low-Latency Simultaneous Speech Translati on. ArXiv Preprint arXiv:2308.03415 . (2023)

  4. [12]

    & Others C HIL: Computers in the human interaction loop

    Waibe11, A., Steusloff, H., Stiefelhagen, R. & Others C HIL: Computers in the human interaction loop. (Citeseer,2005)

  5. [13]

    & Waibel, A

    Stuker, S., Schultz, T., Metze, F. & Waibel, A. Multilin gual articulatory features. 2003 IEEE International Conference On Acoustics, Speech, And Signal Processing, 2003. Proceedings.(ICASSP’03). . 1 pp. I-I (2003)

  6. [14]

    & Waibel, A

    M¨ uller, M., St¨ uker, S. & Waibel, A. Language adaptive multilingual CTC speech recognition. Speech And Computer: 19th International Conference, SPECOM 2017, Hatfield, UK, September 12-16, 201 7, Proceedings 19. pp. 473-482 (2017)

  7. [15]

    & Zaiem, S

    Abdallah, A., Kabboudi, A., Kanoun, A. & Zaiem, S. Lever aging Data Collection and Unsupervised Learning for Code-Switch ed Tunisian Arabic Automatic Speech Recognition. ICASSP 2024-2024 IEEE In- ternational Conference On Acoustics, Speech And Signal Pro cessing (ICASSP). pp. ...

  8. [16]

    & Aldarmak i, H

    Kadaoui, K., Ali, M., Toyin, H., Mohammed, I. & Aldarmak i, H. PolyWER: A Holistic Evaluation Framework for Code-Switche d Speech Recognition. Findings Of The Association F or Computational Linguis- tics: EMNLP 2024 . pp. 6144-6153 (2024,11),

  9. [17]

    & Ali, A

    Hamed, I., Hussein, A., Chellah, O., Chowdhury, S., Mub arak, H., Sitaram, S., Habash, N. & Ali, A. Benchmarking evaluation me trics for code-switching automatic speech recognition. 2022 IEEE Spoken Language Technology W orkshop (SLT). pp. 999-1005 (2023)

  10. [18]

    & Chien, J

    Aditya, B., Rohmatillah, M., Tai, L. & Chien, J. Attenti on-Guided Adaptation for Code-Switching Speech Recognition. ICASSP 2024- 2024 IEEE International Conference On Acoustics, Speech An d Signal Processing (ICASSP) . pp. 10256-10260 (2024)

  11. [19]

    & Khudanpur, S

    Liu, H., Garcia, L., Zhang, X., Khong, A. & Khudanpur, S. En- hancing code-switching speech recognition with interacti ve language biases. ICASSP 2024-2024 IEEE International Conference On Acousti cs, Speech And Signal Processing (ICASSP) . pp. 10886-10890 (2024)

  12. [20]

    & Waibel, A

    Ugan, E., Huber, C., Hussain, J. & Waibel, A. Language-a gnostic Code-Switching in Sequence-To-Sequence Speech Recogniti on. ArXiv Preprint arXiv:2210.08992 . (2022)

  13. [21]

    & Waibel, A

    Huber, C., Ugan, E. & Waibel, A. Code-switching without switch- ing: Language agnostic end-to-end speech translation. ArXiv Preprint arXiv:2210.01512. (2022)

  14. [22]

    & Y ong, Z

    Do˘ gru¨ oz, A., Sitaram, S. & Y ong, Z. Representativeness as a Forgotten Lesson for Multilingual and Code-switched Data Collection and Prepa- ration. Findings Of The Association F or Computational Linguistics : EMNLP 2023 . pp. 5751-5767 (2023,12),

  15. [23]

    & Waibel, A

    Ugan, E., Pham, N. & Waibel, A. DECM: Evaluating Bilingu al ASR Performance on a Code-switching/mixing Benchmark. Proceedings Of The 2024 Joint International Conference On Computational L inguistics, Language Resources And Evaluation (LREC-COLING 2024) . pp. 4468- 4475 (2024)

  16. [24]

    C ¸ etino˘ glu,¨O., Schulz, S. & Vu, N. Challenges of Computational Processing of Code-Switching. Proceedings Of The Second W orkshop On Computational Approaches To Code Switching . pp. 1-11 (2016,11),

  17. [25]

    & Waibel, A

    Y aman, D., Eyiokur, F., B¨ armann, L., Akti, S., Ekenel, H. & Waibel, A. Audio-Visual Speech Representation Expert for Enhanced Talking Face Video Generation and Evaluation. Proceedings Of The IEEE/CVF Conference On Computer Vision And Pattern Recognition . pp. 6003- 6013 (2024)

  18. [26]

    & Niehues, J

    Dinh, T. & Niehues, J. Perturbation-based qe: An explai nable, unsu- pervised word-level quality estimation method for blackbo x machine translation. ArXiv Preprint arXiv:2305.07457 . (2023)

  19. [27]

    A call for clarity in reporting BLEU scores

    Post, M. A call for clarity in reporting BLEU scores. ArXiv Preprint arXiv:1804.08771. (2018)

  20. [28]

    & L ¨ o¨ of, J

    Schmidt, R., Pires, T., Peitz, S. & L ¨ o¨ of, J. Non-autor egressive neural machine translation: A call for clarity. ArXiv Preprint arXiv:2205.10577. (2022)

  21. [29]

    & Bentivogli, L

    Gaido, M., Papi, S., Negri, M. & Bentivogli, L. Speech Tr anslation with Speech Foundation Models and Large Language Models: What is There and What is Missing?. ArXiv Preprint arXiv:2402.12025 . (2024)

  22. [30]

    & Sutskever, I

    Radford, A., Kim, J., Xu, T., Brockman, G., McLeavey, C. & Sutskever, I. Robust speech recognition via large-scale weak supervis ion. Interna- tional Conference On Machine Learning . pp. 28492-28518 (2023)

  23. [31]

    & Others Sca ling speech technology to 1,000+ languages

    Pratap, V ., Tjandra, A., Shi, B., Tomasello, P ., Babu, A ., Kundu, S., Elkahky, A., Ni, Z., Vyas, A., Fazel-Zarandi, M. & Others Sca ling speech technology to 1,000+ languages. Journal Of Machine Learning Research. 25, 1-52 (2024)

  24. [32]

    & Esteve, Y

    Rousseau, A., Del´ eglise, P . & Esteve, Y . TED-LIUM: an A utomatic Speech Recognition dedicated corpus.. LREC. pp. 125-129 (2012)

  25. [33]

    & Collobert , R

    Pratap, V ., Xu, Q., Sriram, A., Synnaeve, G. & Collobert , R. Mls: A large-scale multilingual dataset for speech research. ArXiv Preprint arXiv:2012.03411. (2020)

  26. [34]

    & Zhang, Y

    Ali, A., Bell, P ., Glass, J., Messaoui, Y ., Mubarak, H., Renals, S. & Zhang, Y . The MGB-2 challenge: Arabic multi-dialect broadc ast media recognition. 2016 IEEE Spoken Language Technology W orkshop (SLT)

  27. [35]

    & Zheng, H

    Bu, H., Du, J., Na, X., Wu, B. & Zheng, H. Aishell-1: An ope n- source mandarin speech corpus and a speech recognition base line. 2017 20th Conference Of The Oriental Chapter Of The Internat ional Coordinating Committee On Speech Databases And Speech I/O S ystems And Assessmen...

  28. [36]

    & Weber, G

    Ardila, R., Branson, M., Davis, K., Henretty, M., Kohle r, M., Meyer, J., Morais, R., Saunders, L., Tyers, F. & Weber, G. Com - mon voice: A massively-multilingual speech corpus. ArXiv Preprint arXiv:1912.06670. (2019)

  29. [37]

    & Paulik, M

    Weller, O., Sperber, M., Pires, T., Setiawan, H., Golla n, C., Telaar, D. & Paulik, M. End-to-end speech translation for code switche d speech. ArXiv Preprint arXiv:2204.05076 . (2022)

  30. [38]

    & Abdennadher, S

    Hamed, I., Vu, N. & Abdennadher, S. ArzEn: A speech corpu s for code-switched Egyptian Arabic-English. Proceedings Of The Twelfth Language Resources And Evaluation Conference . pp. 4237-4246 (2020)

  31. [39]

    Lyu, D., Tan, T., Chng, E. & Li, H. SEAME: a Mandarin-Engl ish code- switching speech corpus in south-east asia.. Interspeech. 10 pp. 1986- 1989 (2010)

  32. [40]

    Y e, S., Chen, S., Hu, X. & Xu, X. SC-MoE: Switch Conformer Mixture of Experts for Unified Streaming and Non-streaming Code-Swi tching ASR. ArXiv Preprint arXiv:2406.18021 . (2024)

  33. [41]

    & Chen, B

    Y ang, T., Wang, H., Wang, Y ., Lin, C. & Chen, B. An Effecti ve Mixture- Of-Experts Approach For Code-Switching Speech Recognitio n Leverag- ing Encoder Disentanglement. ICASSP 2024-2024 IEEE International Conference On Acoustics, Speech And Signal Processing (ICA SSP). pp. 1...

  34. [42]

    Boerner, Stephen Deems, Thomas R

    Timothy J. Boerner, Stephen Deems, Thomas R. Furlani, S helley L. Knuth, and John Towns. 2023. ACCESS: Advancing Innovatio n: NSF’s Advanced Cyberinfrastructure Coordination Ecosyst em: Services & Support. Practice and Experience in Advanced Research Computing (PEARC ’23)” . (2023)

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.