Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Can we train ASR systems on Code-switch without real code-switch data? Case study for Singapore's languages

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Synthetic phrase-mixed audio can train code-switched ASR without any real code-switching data.

desk verdict The synthetic-data idea is plausible, but the evaluation doesn't test the central claim because the test sets are themselves synthetic or LLM-generated, and the relative-improvement numbers are inflated by a nonstandard formula. read the letter →

arxiv 2506.14177 v1 pith:AQFCRS3L submitted 2025-06-17 cs.CL cs.AIcs.SDeess.AS

Tuan Nguyen , Huy-Dat Tran This is my paper
classification cs.CLcs.AIcs.SDeess.AS
keywords code-switchingautomaticspeechrecognitionsyntheticdataaugmentationphrase-levelmixinglow-resourcelanguagesSoutheastAsianmultilingualASRSingapore
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 tries to establish that high-performing automatic speech recognition (ASR) for code-switched speech can be built without any real recorded code-switching data, by fabricating the code-switched audio from monolingual speech. Its method mixes phrases across languages at the text level and splices the corresponding audio segments, producing 1,000 hours of synthetic code-switched audio per language pair. Fine-tuning large pretrained ASR models on monolingual data plus this synthetic audio consistently lowered error rates on test sets for Malay-English, Mandarin-Malay, and Tamil-English. The result matters because transcribed code-switched speech is scarce and expensive, so a purely synthetic route would make CS-ASR practical for under-resourced languages and multilingual societies such as Singapore.

What carries the argument

The load-bearing mechanism is a phrase-level mixing pipeline (Figure 1) that converts monolingual audio into synthetic code-switched audio. A sentence in one language is translated, aligned phrase-by-phrase with Fast Align, and 10–30% of its tokens are replaced by the other language's counterpart under a consecutive-word constraint to mimic Singaporean phrase-level switching; the corresponding audio segments are then located by forced alignment and spliced together with amplitude normalization. This pipeline produced the 3,000 synthetic training hours that, when added to 3,775 hours of monolingual speech during fine-tuning, carried the observed error reductions in the paper's experiments.

What would settle it

A natural, spontaneously code-switched test set for one of these language pairs—for example, transcribed Singaporean radio or podcast conversations recorded independently of any training source—run through the same MONO+CS fine-tuning recipe, would settle it: if the synthetic-augmented model fails to beat the monolingual baseline on such data, the central claim that real CS data is unnecessary collapses.

Watch

Extended reading notes

Core claim

The central claim is that phrase-mixed synthetic code-switched audio can stand in for real code-switching data in training: adding 1,000 hours of it per pair to monolingual fine-tuning improved WER/CER/MER on all three code-switch test sets for all three model families evaluated (Whisper-Large-v3-Turbo, MMS-1B-All, and SeamlessM4T-v2-Large). The authors report the largest relative gains on Malay-English (about 83% for Whisper Turbo-V3), substantial gains on Tamil-English, and smaller but real gains on Mandarin-Malay, where the augmentation also reduced strong monolingual Mandarin bias. They attribute the transfer to the distributional match between the synthetic training data and the test sets, measured by code-mixing index, switching frequency, and language balance, and they present these test sets as the first code-switch ASR benchmarks for under-resourced Southeast Asian language pairs.

Load-bearing premise

The load-bearing premise is that the test sets faithfully represent real-world code-switched speech; the BM-EN and ZH-BM tests are ChatGPT-generated texts read by Singaporeans, and the TA-EN test is spliced from the same audio sources used for training, so the experiments may measure matching of the synthetic training distribution rather than genuine code-switching ability.

Editorial extensions

If this is right

  • Fine-tuning with synthetic phrase-mixed CS audio improves code-switched ASR error rates on all three language pairs and across all three pretrained model families tested.
  • The largest gains occur for BM-EN, where the synthetic training distribution most closely matches the test set on code-mixing metrics, suggesting that measurable distributional alignment predicts transfer.
  • For Mandarin-Malay, the synthetic CS augmentation reduces a strong Mandarin bias in pretrained models, lowering substitution, insertion, and deletion errors on Malay segments.
  • For MMS-1B-All, full fine-tuning with a BPE vocabulary is required to benefit from the synthetic data; adapter-based fine-tuning shows little or no gain.
  • If the effect holds beyond these settings, the approach removes the need for transcribed real CS data, turning monolingual corpora plus a translation/alignment step into a scalable CS-ASR training resource.

Reading between the lines

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

  • Inference: The same recipe could be applied to any language pair with adequate monolingual corpora and decent machine translation, shifting the bottleneck from collecting CS transcripts to translation and alignment quality.
  • Inference: Because the BM-EN and ZH-BM test sets are generated from ChatGPT text read aloud and the TA-EN test set is spliced from the same audio as the training data, the reported gains may overstate performance on spontaneously produced code-switched speech; a naturally occurring held-out corpus would be the stricter test.
  • Inference: The code-mixing statistics used in the paper could serve as a cheap pre-flight predictor: if a language pair's synthetic data diverges from a small real-sample reference as much as ZH-BM does here, the augmentation gains should be expected to shrink.
  • Inference: The same synthetic CS audio-text pairs could feed downstream tasks beyond ASR, such as code-switched speech translation or spoken language understanding, since the pipeline already produces aligned text and audio.
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 / 5 minor

Summary. The paper proposes a phrase-level mixing method that splices monolingual speech from Singapore's languages to create synthetic code-switched (CS) training data, and fine-tunes Whisper, MMS, and SeamlessM4T on monolingual data with and without 1,000 hours per language pair of synthetic CS audio. It evaluates the resulting models on newly constructed test sets for Malay-English (BM-EN), Mandarin-Malay (ZH-BM), and Tamil-English (TA-EN) and reports consistent WER/CER/MER improvements from adding synthetic CS data. The central claim is that high-performing CS ASR can be trained without real code-switching data.

Significance. If the central claim were established, the work would be a useful, scalable recipe for low-resource CS ASR: the synthetic-data pipeline is clearly described, the scale (3,000 hours) is substantial, and the comparison across three large pretrained models is a useful engineering data point. The release of new benchmark test sets for three under-resourced Southeast Asian language pairs is a genuine contribution, as are the detailed phrase-mixing statistics in Table 3. However, the evaluation design prevents the paper from actually demonstrating that real CS data is unnecessary: the test sets are themselves largely synthetic or LLM-processed, and the reported relative-improvement metric is nonstandard and changes the headline ranking. The paper's main value at present is as a synthetic-data-augmentation study with a candidate benchmark, not as a proof that real CS data can be dispensed with.

major comments (4)
  1. [Section 3.3, Table 3] The evaluation does not test the central claim because the test sets are not natural, spontaneous code-switched speech. BM-EN and ZH-BM are 'collected from ChatGPT conversations (text) spoken by Singaporeans' (Section 3.3), meaning they are read speech of LLM-generated text with no spontaneous disfluencies, unplanned switch points, or conversational prosody. The TA-EN set is constructed by splicing sentences from MUCS (TA) and IMDA3 (EN) using NeMo's sentence-level CS scripts, and Table 3 shows its I-Index is 6.50 versus 37.33 for the phrase-mixed training data, so it is not even within-utterance code-switching. Because both the training set and these test sets are generated by similar pipeline decisions, the similarity in CMI/I-Index/M-Index reported in Section 4.1 may reflect design choices rather than fidelity to real-world CS. The conclusion in Section 5 that 'high-performing ASR systems ... can be effectively trained without real code-switching data' is therefore not supported by the evidence. The paper needs at least one naturally occurring CS test set (for example, an external BM-EN corpus, or SEAME for a related pair) before this claim can stand.
  2. [Table 1, Section 4.3] The relative-improvement formula used in Table 1 appears to be (baseline - new)/new instead of the standard (baseline - new)/baseline. This inflates the reported gains and changes the cross-pair ranking. For example, Whisper on BM-EN: baseline 30.67, +CS 16.75, reported 83.1% improvement; the standard computation gives (30.67-16.75)/30.67 = 45.4%. For Whisper on TA-EN: baseline 114.6, +CS 51.37, reported 123%; the standard computation gives 55.2%. Under the standard metric, TA-EN shows a larger relative gain than BM-EN for Whisper, contradicting the abstract's ordering 'BM-EN showing highest gains, then TA-EN and ZH-BM.' The paper should state the formula explicitly and re-report the main results with the standard baseline-denominator metric.
  3. [Section 3.3, Table 2] The TA-EN test set overlaps with the MONO training data: Table 2 lists IMDA 1-3 and 5-6 (750h) as part of the English MONO corpus, and Section 3.3 states that the TA-EN test set is spliced from MUCS (TA) and IMDA3 (EN). The English side of the test set therefore comes from the same corpus and acoustic conditions as the training data, giving MONO and MONO+CS an advantage on the English portions of TA-EN that would not transfer to unseen English speakers. The authors must report whether any specific speaker or utterance overlap exists between the MONO training data and the TA-EN test set, and if not, explain why the same corpus is nevertheless not a training/test leakage.
  4. [Section 4.3] The claim of 'consistent performance improvements across all models in the MONO+CS setting' is contradicted by the table itself. For SeamlessM4T-v2 on BM-EN, +CS gives 14.58 WER while +MONO gives 13.08, so adding synthetic CS data degrades performance relative to monolingual fine-tuning. The paper's own Section 4.3 later concedes this exception. The characterization should be revised to state the actual pattern, including which model/pair combinations show degradation.
minor comments (5)
  1. [Table 1] The table header is ambiguous: 'Data Model # Multilingual Code-Switch MONO CS' does not clearly label the rightmost two columns as averages of the monolingual and code-switch columns, and the caption does not explain this. Please add explicit column labels and a note on how the averages are computed.
  2. [Table 3] The labels 'Real BM-EN' and 'Real ZH-BM' are misleading: these test sets are read speech of ChatGPT-generated text, which is not natural spontaneous code-switching. Please rename them, e.g., 'Read ChatGPT BM-EN' and 'Read ChatGPT ZH-BM.'
  3. [Section 3.3] The sentence claiming 'these test sets are the first benchmarks for code-switch evaluation in under-resourced Southeast Asian languages' should be qualified, since SEAME is already a widely used Mandarin-English CS benchmark in the same geographical region (as the paper itself notes in Table 3). The claim of novelty should be scoped to the specific language pairs.
  4. [Contributions, Section 3.4] The paper lists 'Publicly Available Benchmark Datasets' as a contribution but says the datasets are available only 'upon email request.' This limits reproducibility and conflicts with the word 'publicly.' Please provide a permanent repository link or a clear statement of access conditions.
  5. [Section 2.2, Section 3.2] The phrase-mixing generation uses Google Translate and Mesolitica's translation model, but there is no evaluation of translation or alignment quality, which could directly affect the naturalness of the synthetic CS audio. A small human or automatic quality check on the generated CS text would strengthen the method.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's empirical claim rests on generated benchmarks, but no prediction is equivalent to its inputs by construction.

full rationale

I traced the claimed derivation chain: synthetic phrase-mixed CS audio is generated from monolingual MONO corpora via translation, Fast Align, HMM-DNN alignment, and audio splicing; models are fine-tuned on MONO plus synthetic CS; and improvements are measured on the paper's CS test sets. I find no step where a prediction is mathematically or statistically forced by an input. The BM-EN and ZH-BM test sets are ChatGPT-scripted read speech, and the TA-EN test set is synthesized by splicing MUCS and IMDA3 audio using NeMo's sentence-level CS scripts, while synthetic training CS is phrase-level and generated from MONO sources. These are real external-validity and source-overlap concerns: for TA-EN especially, the training and test distributions share a splicing methodology and at least one source corpus (IMDA3 is in the MONO training set), so gains may partly reflect distribution matching rather than genuine code-switching ability. However, the test sets are not used to fit the models, the BM-EN and ZH-BM test sets are not produced by the same Google-Translate/Fast-Align pipeline as the training data, and no load-bearing self-citation or uniqueness theorem is invoked. The relative-improvement percentages in Table 1 use an unusual denominator (the post-fine-tuned WER rather than the baseline), which can inflate and reorder apparent gains, but that is a reporting/metric-construction issue, not circularity. The paper's own closing admission that future work should enhance 'the linguistic realism of synthetic data' further indicates the limitation is benchmark fidelity, not a logical reduction of the conclusion to its inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The contributions are empirical; the central claim rests on a pipeline of chosen design parameters (mixing ratio, data volume, normalization) and on domain assumptions about what synthetic data and semi-synthetic test sets represent. No new theoretical entities are introduced.

free parameters (3)
  • phrase mixing ratio = 10-30% tokens replaced per sentence
    Hand-chosen replacement range to mimic phrase-level Singaporean code-switching; affects the entire synthetic CS distribution and is never ablated.
  • synthetic data volume per language pair = 1000 hours
    Chosen scale for synthetic CS training data; no ablation shows sensitivity, so gains could depend on this volume.
  • relative improvement normalization = (baseline-new)/new
    The paper uses the new value as denominator when reporting gains, which inflates numbers; this choice directly forces the headline ranking of language-pair gains.
assumptions (4)
  • domain assumption Phrase-level mixing mimics natural Singaporean code-switching patterns
    Section 2.2 relies on [17] for the claim that Singaporean CS is phrase-level; the whole synthetic generation strategy depends on this.
  • domain assumption Spliced audio with amplitude normalization is acoustically natural enough for ASR fine-tuning
    Section 2.2 and Figure 1 assume concatenated forced-aligned segments yield usable training audio; no naturalness test is provided.
  • domain assumption The benchmark test sets are representative of real-world code-switching
    Section 3.3 labels ChatGPT-read and spliced test sets as 'real-world'; this representativeness is unvalidated.
  • domain assumption WER/CER/MER comparisons are fair across models with different tokenizers and vocabularies
    Section 3.3 and Table 1 compare MMS with character or BPE vocabularies against Whisper and SeamlessM4T, which may bias metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can we train ASR systems on Code-switch without real code-switch data? Case study for Singapore's languages." pith.science (2026). https://pith.science/paper/AQFCRS3L

@misc{pith2026250614177,
  author       = {Pith},
  title        = {Pith review of: Can we train ASR systems on Code-switch without real code-switch data? Case study for Singapore's languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQFCRS3L}},
  note         = {Machine review of arXiv:2506.14177}
}
read the original abstract

Code-switching (CS), common in multilingual settings, presents challenges for ASR due to scarce and costly transcribed data caused by linguistic complexity. This study investigates building CS-ASR using synthetic CS data. We propose a phrase-level mixing method to generate synthetic CS data that mimics natural patterns. Utilizing monolingual augmented with synthetic phrase-mixed CS data to fine-tune large pretrained ASR models (Whisper, MMS, SeamlessM4T). This paper focuses on three under-resourced Southeast Asian language pairs: Malay-English (BM-EN), Mandarin-Malay (ZH-BM), and Tamil-English (TA-EN), establishing a new comprehensive benchmark for CS-ASR to evaluate the performance of leading ASR models. Experimental results show that the proposed training strategy enhances ASR performance on monolingual and CS tests, with BM-EN showing highest gains, then TA-EN and ZH-BM. This finding offers a cost-effective approach for CS-ASR development, benefiting research and industry.

Figures

Figures reproduced from arXiv: 2506.14177 by the authors.

Figure 1
Figure 1. Phrase-mixed CS Data Generation Pipeline for ASR Training ALL, and SeamlessM4T-v2-Large), revealing consistent per￾formance improvements across multiple under-resourced lan￾guage pairs. • Publicly Available Benchmark Datasets: We release novel evaluation test sets for three under-resourced language pairs to foster research in code-switching ASR. Datasets and model APIs are available upon email request. 2. Methods 2.… view at source ↗
Figure 2
Figure 2. MER components of ZH-BM testset. Include Correctness, Insertion rate, Deletion rate, and Substitution rate. similar across all metrics (CMI: 38.37 vs. 36.88, I-Index: 46.81 vs. 39.44, M-Index: 23.47 vs. 24.73). However, for ZH-BM, the training and test sets are less consistent. The training set shows higher switching frequency but lesser language balance compared to the test set. In contrast, compared to training da… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 45 canonical work pages

  1. [1]

    Despite its widespread nature, research on CS, especially in Automatic Speech Recognition (ASR), significantly lags behind that of monolingual language processing

    Introduction Code-switching (CS), the natural and fluid alternation between two or more languages within a single conversation or utterance, is a pervasive linguistic phenomenon worldwide, particularly as multilingualism grows [1]. Despite its widespread nature, research on CS, especially in Automatic Speech Recognition (ASR), significantly lags behind th...

    arXiv 2025

  2. [2]

    Code-Switching in Singapore Singapore is a culturally and linguistically diverse nation, rep- resenting the broader Southeast Asian region within a single country

    Methods 2.1. Code-Switching in Singapore Singapore is a culturally and linguistically diverse nation, rep- resenting the broader Southeast Asian region within a single country. The primary language of communication is English, which is supported by government policies that promote Man- darin Chinese, Bahasa Malay, and Tamil as official languages [18]. As ...

    work page

  3. [3]

    Translation: We employ Google Translate for BM-EN and TA-EN pairs, and Mesolitica’s translation model 1 for ZH- EN, to optimize for speed and accuracy

    work page

  4. [4]

    Textual Alignment: We replace BERT-based alignment with FAST ALIGN [20] for improved scalability

    work page

  5. [5]

    Phrase-Mixed Replacement: Singaporean code-switching occurs primarily at the phrase level rather than the word level [17], with speakers naturally alternating between full phrases while maintaining academic terminology. To better reflect this phenomenon, we replace the fixed 20% lexicon substitu- tion with a more flexible 10-30% mixing per sentence, cap- ...

    work page

  6. [6]

    This adjustment addresses the distortion observed when using the original energy-based normalization

    Audio Splicing: We improve Speech Collage [16] by in- corporating NeMo’s amplitude-based normalization 2, which 1hf.co/mesolitica/translation-nanot5-base-malaysian-cased 2github.com/NVIDIA/NeMo/tree/main/scripts/ produces more natural speech output. This adjustment addresses the distortion observed when using the original energy-based normalization. 2.3. ...

    work page

  7. [7]

    M ONO Data We utilize monolingual data from four languages: EN, ZH, BM, and TA, incorporating both local and non-local accents

    Experiments 3.1. M ONO Data We utilize monolingual data from four languages: EN, ZH, BM, and TA, incorporating both local and non-local accents. Local accents refer to those associated with Singapore or similar re- gions. All datasets, except for the Internal dataset (which is sampled from our proprietary data), are listed in Table 2. We merged all the da...

    work page

  8. [8]

    Results 4.1. Analysis on Code-Switch Characteristics in Phrase- mixed Training Sets and Test Sets We assessed code-switching (CS) patterns using three metrics: the Code-Mixing Index (CMI) considers both the number of switches and language distribution [16]; the I-Index for how often switching occurs; and the M-Index for language balance [14]. Table 3 show...

    work page

Show all 46 references
  1. [9]

    Conclusion This study demonstrates that high-performing ASR systems for multilingual and code-switched speech can be effectively trained without real code-switching data. By leveraging phrase- mixed synthetic data validated on new benchmarks for Sin- gaporean language pairs (B...

  2. [10]

    World population bilingual percentage,

    Newsdle, “World population bilingual percentage,” https:// www.newsdle.com/blog/world-population-bilingual-percentage, accessed: 2024-12-24

    2024

  3. [11]

    Effectiveness of mining audio and text pairs from public data for improving asr systems for low-resource languages,

    K. S. Bhogale, A. Raman, T. Javed et al. , “Effectiveness of mining audio and text pairs from public data for improving asr systems for low-resource languages,” 2022. [Online]. Available: https://arxiv.org/abs/2208.12666

    2022 arXiv

  4. [12]

    reducing multilingual context con- fusion for end-to-end code-switching automatic speech recogni- tion,

    S. Zhang, J. Yi, Z. Tian et al., “reducing multilingual context con- fusion for end-to-end code-switching automatic speech recogni- tion,” in Interspeech 2022, 2022, pp. 3894–3898

    2022

  5. [13]

    Reducing language confusion for code-switching speech recognition with token-level language diarization,

    H. Liu, H. Xu, L. P. Garcia et al., “Reducing language confusion for code-switching speech recognition with token-level language diarization,” in ICASSP 2023 - 2023 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  6. [14]

    Lae-st-moe: Boosted language- aware encoder using speech translation auxiliary task for e2e code-switching asr,

    G. Ma, W. Wang, Y . Li et al., “Lae-st-moe: Boosted language- aware encoder using speech translation auxiliary task for e2e code-switching asr,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2023, pp. 1–8

    2023

  7. [15]

    An effective mixture- of-experts approach for code-switching speech recognition lever- aging encoder disentanglement,

    T.-T. Yang, H.-W. Wang, Y .-C. Wanget al., “An effective mixture- of-experts approach for code-switching speech recognition lever- aging encoder disentanglement,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2024, p...

    2024

  8. [16]

    Towards one model to rule all: Multilingual strategy for dialectal code- switching arabic asr,

    S. A. Chowdhury, A. Hussein, A. Abdelali et al. , “Towards one model to rule all: Multilingual strategy for dialectal code- switching arabic asr,” in Interspeech 2021, 2021, pp. 2466–2470

    2021

  9. [17]

    Language-specific characteristic assistance for code-switching speech recognition,

    T. Song, Q. Xu, M. Ge et al., “Language-specific characteristic assistance for code-switching speech recognition,” in Interspeech 2022, 2022, pp. 3924–3928

    2022

  10. [18]

    Camel: Cross- attention enhanced mixture-of-experts and language bias for code-switching speech recognition,

    H. Wang, X. Wan, N. Zheng et al. , “Camel: Cross- attention enhanced mixture-of-experts and language bias for code-switching speech recognition,” 2024. [Online]. Available: https://arxiv.org/abs/2412.12760

    2024 arXiv

  11. [19]

    Enhancing code-switching speech recognition with interactive language biases,

    H. Liu, L. P. Garcia, X. Zhang et al., “Enhancing code-switching speech recognition with interactive language biases,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 10 886–10 890

    2024

  12. [20]

    Leverag- ing data collection and unsupervised learning for code-switched tunisian arabic automatic speech recognition,

    A. A. B. Abdallah, A. Kabboudi, A. Kanoun et al. , “Leverag- ing data collection and unsupervised learning for code-switched tunisian arabic automatic speech recognition,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), ...

    2024

  13. [21]

    Adapting openai’s whisper for speech recognition on code-switch mandarin-english seame and asru2019 datasets,

    Y . Yang, Y . Peng, H. Huanget al., “Adapting openai’s whisper for speech recognition on code-switch mandarin-english seame and asru2019 datasets,” in 2024 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2024, accessed: 2024-12-24

    2024

  14. [22]

    Indonesian- English code-switching speech recognition using the machine speech chain based semi-supervised learning,

    R. V . M. Tazakka, D. Lestari, A. Purwarianti et al., “Indonesian- English code-switching speech recognition using the machine speech chain based semi-supervised learning,” in Proceedings of the 3rd Annual Meeting of the Special Interest Group on Under- resourced Languages @ L...

    2024

  15. [23]

    Multilingual and code- switching speech recognition,

    A. Ali, S. Chowdhury, L. Ondel et al., “Multilingual and code- switching speech recognition,” 2022 Eighth Frederick Jelinek Memorial Summer Workshop, Johns Hopkins University, Tech. Rep., August 2023. [Online]. Available: https://www.clsp.jhu. edu/wp-content/uploads/2023/08/JS...

    2022

  16. [24]

    Unsupervised code-switched text generation from parallel text,

    J. Chi, B. Lu, J. Eisner et al., “Unsupervised code-switched text generation from parallel text,” inINTERSPEECH 2023, 2023, pp. 1419–1423

    2023

  17. [25]

    Speech collage: Code- switched audio generation by collaging monolingual corpora,

    A. Hussein, D. Zeinali, O. Klejch et al., “Speech collage: Code- switched audio generation by collaging monolingual corpora,” in ICASSP 2024 - 2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) , 2024, pp. 12 006– 12 010

    2024

  18. [26]

    Bolton and W

    K. Bolton and W. Botha, Multilingualism and Language Mixing Among Singapore University Students. Cham: Springer Interna- tional Publishing, 2019, pp. 43–61

    2019

  19. [27]

    Languages of singapore - a detailed guide to singapore languages,

    Holidify, “Languages of singapore - a detailed guide to singapore languages,” https://www.holidify.com/pages/ singapore-languages-629.html, accessed: 2024-12-24

    2024

  20. [28]

    Bilingual policy,

    NLB, “Bilingual policy,” https://www. nlb.gov.sg/main/article-detail?cmsuuid= 82fbbca5-e8e2-40cc-b944-fbb2bd2367fe, accessed: 2024- 12-25

    2024

  21. [29]

    A simple, fast, and effective reparameterization of IBM model 2,

    C. Dyer, V . Chahuneau, and N. A. Smith, “A simple, fast, and effective reparameterization of IBM model 2,” in Proceedings of the 2013 Conference of the North American Chapter of the As- sociation for Computational Linguistics: Human Language Tech- nologies, L. Vanderwende, H....

    2013

  22. [30]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu et al., “Robust speech recognition via large-scale weak supervision,” 2022

    2022

  23. [31]

    Scaling speech technology to 1,000+ languages,

    V . Pratap, A. Tjandra, B. Shi et al., “Scaling speech technology to 1,000+ languages,” 2023

    2023

  24. [32]

    Seamless: Multilingual expressive and streaming speech translation,

    S. Communication, L. Barrault, Y .-A. Chung et al., “Seamless: Multilingual expressive and streaming speech translation,” 2023

    2023

  25. [33]

    Ml-superb 2.0: Benchmark- ing multilingual speech models across modeling constraints, lan- guages, and datasets,

    J. Shi, S.-H. Wang, W. Chen et al., “Ml-superb 2.0: Benchmark- ing multilingual speech models across modeling constraints, lan- guages, and datasets,” in Interspeech 2024, 2024, pp. 1230–1234

    2024

  26. [34]

    Librispeech: An asr cor- pus based on public domain audio books,

    V . Panayotov, G. Chen, D. Poveyet al., “Librispeech: An asr cor- pus based on public domain audio books,” in 2015 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  27. [35]

    Building the singapore english national speech corpus,

    J. X. Koh, A. Mislan, K. Khoo et al. , “Building the singapore english national speech corpus,” in Interspeech 2019, 2019, pp. 321–325

    2019

  28. [36]

    Aishell-1: An open-source mandarin speech corpus and a speech recognition baseline,

    H. Bu, J. Du, X. Na et al., “Aishell-1: An open-source mandarin speech corpus and a speech recognition baseline,” inOriental CO- COSDA 2017, 2017, p. Submitted

    2017

  29. [37]

    Asr datasets - magic data technology,

    Magic Data Technology, “Asr datasets - magic data technology,” 2024, accessed: 2024-09-13. [Online]. Available: https: //www.magicdatatech.com/datasets/asr

    2024

  30. [38]

    Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,

    B. Zhang, H. Lv, P. Guo et al., “Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,” in Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022

    2022

  31. [39]

    Subword dictionary learning and segmentation techniques for automatic speech recognition in tamil and kannada,

    M. A, B. Pilar, and R. A. G, “Subword dictionary learning and segmentation techniques for automatic speech recognition in tamil and kannada,” 2022

    2022

  32. [40]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis et al. , “Common voice: A massively-multilingual speech corpus,” 2020

    2020

  33. [41]

    Multilingual and code-switching asr challenges for low resource indian languages,

    A. Diwan, R. Vaideeswaran, S. Shah et al. , “Multilingual and code-switching asr challenges for low resource indian languages,” Proceedings of Interspeech, 2021

    2021

  34. [42]

    Seame: a mandarin- english code-switching speech corpus in south-east asia,

    D.-C. Lyu, T.-P. Tan, E. S. Chng et al. , “Seame: a mandarin- english code-switching speech corpus in south-east asia,” in In- terspeech 2010, 2010, pp. 1986–1989

    2010

  35. [43]

    Tradition or innovation: A comparison of modern asr methods for forced alignment,

    R. Rousso, E. Cohen, J. Keshet et al., “Tradition or innovation: A comparison of modern asr methods for forced alignment,” in Interspeech 2024, 2024, pp. 1525–1529

    2024

  36. [44]

    Open-source Multi- speaker Speech Corpora for Building Gujarati, Kannada, Malay- alam, Marathi, Tamil and Telugu Speech Synthesis Systems,

    F. He, S.-H. C. Chu, O. Kjartansson et al., “Open-source Multi- speaker Speech Corpora for Building Gujarati, Kannada, Malay- alam, Marathi, Tamil and Telugu Speech Synthesis Systems,” in Proceedings of The 12th Language Resources and Evaluation Conference (LREC). Marseille, F...

    2020

  37. [45]

    pyannote.audio: neural building blocks for speaker diarization,

    H. Bredin, R. Yin, J. M. Coria, G. Gelly, P. Korshunov, M. Lavechin, D. Fustes, H. Titeux, W. Bouaziz, and M.-P. Gill, “pyannote.audio: neural building blocks for speaker diarization,” in ICASSP 2020, IEEE International Conference on Acoustics, Speech, and Signal Processing, B...

    2020

  38. [46]

    Effects of di- alectal code-switching on speech modules: A study using egyp- tian arabic broadcast speech,

    S. A. Chowdhury, Y . Samih, M. Eldesouki et al., “Effects of di- alectal code-switching on speech modules: A study using egyp- tian arabic broadcast speech,” in Interspeech 2020 , 2020, pp. 2382–2386

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.