Pith. sign in

REVIEW 2 major objections 5 minor 37 references

A multimodal fusion of speech and CTC features improves both dialect ID and ASR at once for Indian languages, without the usual trade-off.

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-4.5

2026-07-12 06:31 UTC pith:UG2W2ZVB

load-bearing objection Solid empirical joint ASR-DID result on 33 Indian dialects that actually improves both tasks without the usual trade-off. the 2 major comments →

arxiv 2607.02862 v1 pith:UG2W2ZVB submitted 2026-07-03 cs.CL eess.AS

Jointly Improving Dialect Identification and ASR in Indian Languages using Multimodal Feature Fusion

Saurabh Kumar , Amartyaveer , Prasanta Kumar Ghosh This is my paper
classification cs.CL eess.AS
keywords automatic speech recognitiondialect identificationmultimodal feature fusionIndian languagesConformerCTC embeddingsgating mechanismlow-resource speech
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.

Indian languages often have many dialects that share phonetics and vocabulary, so identifying the dialect is harder than identifying the language, and most systems that try to do both end up trading one for the other. Earlier joint models either prepend dialect labels to the text (which hurts ASR when the label is wrong) or train dialect ID without feeding its features back into recognition. This paper shows a different route: extract dialect cues from Conformer speech features with a bottleneck encoder, extract complementary cues from CTC token embeddings with a small RoBERTa, fuse them with a learned gate and attention, then concatenate the refined dialect embeddings (detached) back into the ASR encoder. On eight languages covering thirty-three dialects the method reaches 81.63 percent average dialect accuracy together with 4.65 percent character and 17.73 percent word error rates, both better than the previous best joint baseline. The gains matter because they remove the forced choice between accurate dialect tagging and accurate transcription for low-resource, dialect-rich speech.

Core claim

Joint ASR and dialect identification need not trade off against each other when dialectal information is extracted multimodally from Conformer speech representations and CTC embeddings, fused by gating and attention, and then re-injected into the ASR encoder without prepending dialect tokens. The resulting model simultaneously raises average dialect accuracy to 81.63 percent and lowers average CER/WER to 4.65 percent / 17.73 percent across eight Indian languages and thirty-three dialects, with the largest relative ASR gains appearing precisely on utterances whose dialect is misclassified.

What carries the argument

ASR-BN-ROB multimodal fusion: a Bottleneck Encoder (1-D convolutions + bottleneck transformer) on Conformer outputs, a RoBERTa encoder on projected CTC embeddings, a sigmoid gate that adaptively mixes the two streams, an attention encoder that refines the fused representation, and a detached concatenation of the resulting dialect embeddings back into the ASR path.

Load-bearing premise

Detaching the dialect embeddings before they are concatenated into the ASR path fully blocks harmful gradient interference while still letting the fused features improve transcription; if that separation is incomplete or the features carry too little usable information, the joint gains disappear.

What would settle it

Re-train the identical architecture with the detachment removed (or replaced by a soft stop-gradient) and measure whether CER/WER on the same eight-language RESPIN splits rise above the reported 4.65/17.73 while dialect accuracy stays at or above 81.63 percent; if both metrics degrade or the ASR gain vanishes, the claim fails.

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

If this is right

  • Dialect-aware ASR systems no longer need to prepend dialect tokens, removing the error cascade that occurs when dialect prediction is wrong.
  • The same gated speech-plus-CTC fusion can be dropped into existing Conformer pipelines for other multi-dialect languages without redesigning the decoder.
  • Lower CER/WER on misclassified dialects implies that the fused embeddings supply useful acoustic-linguistic context even when the hard dialect decision is incorrect.
  • Joint optimization with a single weighted loss becomes practical for low-resource Indian languages where separate high-quality ASR and DID models are unavailable.

Where Pith is reading between the lines

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

  • The gating weights themselves may reveal which dialects rely more on acoustic versus lexical cues, offering a diagnostic for dialect typology.
  • Because the frontend is frozen IndicWav2Vec, further gains may be available by unfreezing later layers or by swapping in stronger multilingual SSL models.
  • The architecture should transfer to code-switched or conversational speech if the same multimodal path is retained, but that remains untested.
  • A single multilingual multi-dialect model using the same fusion could replace the eight separate language models, reducing deployment cost.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes a joint multi-task ASR-DID architecture (ASR-BN-ROB) for eight Indian languages covering 33 dialects. A Bottleneck Encoder extracts dialectal cues from Conformer speech representations while a RoBERTa encoder processes CTC embeddings; the two streams are fused by a learnable gate, refined by an Attention Encoder, and the resulting embeddings are detached and concatenated with Conformer outputs to enrich the ASR pathway. On the RESPIN read-speech corpus the method reports 81.63 % average DID accuracy together with 4.65 % CER / 17.73 % WER, statistically significantly better than the prior ASR-DID-ROB baseline that prepends dialect tokens, and without the ASR degradation that baseline exhibits on incorrectly classified utterances. Ablations (ASR-BN, ASR-ROB) isolate the contribution of each modality and of the gated fusion.

Significance. If the reported joint gains hold under broader conditions, the work supplies a practical, publicly released recipe that simultaneously advances both dialect identification and recognition for low-resource Indian languages—an area where previous multi-task systems typically traded one metric for the other. The detach-and-fuse design, the language-wise tables, confusion-matrix analysis, and paired t-tests constitute a clear empirical advance over the authors’ own prior baseline and over several published ASR-DID variants. Code and models are released, supporting reproducibility.

major comments (2)
  1. Section 2.1 and Table 3: the central claim that detaching the DID embeddings before concatenation both prevents harmful gradient interference and still supplies useful dialectal information rests on a single architectural choice. While the incorrect-DID subset shows a clear relative ASR improvement, the paper does not report an ablation that keeps the embeddings attached (or that freezes the DID block after pre-training). Without that control it remains possible that residual leakage or insufficient information flow contributes to the observed gains; a short attached-vs-detached comparison would make the design claim load-bearing rather than plausible.
  2. Section 3.2 and Tables 1–2: all results are obtained with a frozen IndicWav2Vec frontend and a single train/dev/test split of read speech. No confidence intervals, multiple random seeds, or spontaneous-speech evaluation are provided. Given that the free parameters (λ_CTC, γ_CE, bottleneck/RoBERTa dimensions, learning-rate schedule) are tuned on the same validation set, the statistical significance of the 1–2 % absolute gains could be overstated; at least seed-averaged means and standard deviations would strengthen the claim that the multimodal fusion is robust.
minor comments (5)
  1. Figure 1 caption and Section 2.2.1: the Bottleneck Encoder is described as replacing 2-D convolutions with 1-D convolutions “to better preserve temporal information,” yet no quantitative comparison with the original 2-D design of [26] is given; a one-sentence ablation or citation of the performance drop would clarify the design decision.
  2. Equation (3): the weighting coefficients λ_CTC and γ_CE are introduced without stating the final values used for the reported runs; listing them (or noting that they were selected by validation) would aid reproducibility.
  3. Table 1 and Figure 2: dialect labels are given only as D1–D5; a short mapping to the actual dialect names (or a reference to the RESPIN documentation) would make the confusion matrices more interpretable for readers unfamiliar with the corpus.
  4. Section 4.1: the claim of a 16.08 % average reduction in the standard deviation of dialect-wise accuracies is useful but the per-language standard deviations themselves are not tabulated; adding them would allow independent verification.
  5. Typographical consistency: “Conformer-based” vs “conformer”, “RoBERTa” vs “Roberta”, and occasional missing spaces after punctuation appear throughout; a light copy-edit pass would improve readability.

Circularity Check

0 steps flagged

No significant circularity; purely empirical multi-task architecture with held-out evaluation against external and prior baselines.

full rationale

The paper proposes a concrete multimodal architecture (Bottleneck Encoder on Conformer features + RoBERTa on CTC embeddings, gated fusion, attention refinement, detached concatenation into the ASR path) and a weighted multi-task loss, then reports measured DID accuracy, CER and WER on the held-out RESPIN test splits for eight languages. All claims of joint improvement are empirical comparisons (Tables 1–3, paired t-tests) against Base-ASR and several ASR-DID variants, including the authors’ own prior ASR-DID-ROB baseline [11]. No equation, uniqueness theorem, fitted constant, or self-citation is used to force or redefine the reported numbers; the prior work appears only as a comparison point, not as a load-bearing derivation step. Architectural choices (gating, detach, attention sizes) are stated as design decisions, not derived from the target metrics. The evaluation is therefore self-contained against external data and does not reduce any claimed result to its own inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 1 invented entities

Empirical deep-learning paper; the central claim rests on architectural choices, loss-weight hyper-parameters, a frozen SSL frontend, and the assumption that the RESPIN read-speech corpus adequately represents dialectal variation. No new physical entities are postulated; free parameters are the usual neural-network knobs.

free parameters (4)
  • lambda_CTC
    Balances CTC vs. attention ASR losses (Eq. 3); value chosen empirically on validation data.
  • gamma_CE
    Scales the dialect cross-entropy term; set to 5 after validation tuning.
  • bottleneck_dim / RoBERTa_hidden / attention_dims
    Architectural widths (32, 64, 256, 1024) selected by hand on validation performance.
  • learning_rate / warmup / dropout
    Optimizer and regularization settings (0.002, 1.5e4 steps, 0.1) tuned on development set.
axioms (3)
  • domain assumption Frozen IndicWav2Vec layers 7–11 supply sufficiently rich phonetic features for both ASR and DID.
    Stated in §3.2; upstream parameters remain frozen throughout training.
  • ad hoc to paper Detaching dialect embeddings before concatenation prevents gradient interference while still transferring useful information.
    Explicit design choice in §2.1; no formal proof that detachment is optimal.
  • domain assumption RESPIN train/dev/test splits (following [11]) are representative of real dialectal speech.
    Data section §3.1; all experiments use the same fixed splits.
invented entities (1)
  • ASR-BN-ROB multimodal fusion block (Bottleneck Encoder + RoBERTa + gate + Attention Encoder with detach) no independent evidence
    purpose: Jointly extract and inject dialectal features into the ASR encoder without prepending dialect tokens.
    The composite module is the paper’s main technical contribution; individual sub-components exist in prior work but the combination and detach mechanism are new here.

pith-pipeline@v1.1.0-grok45 · 16608 in / 2619 out tokens · 25854 ms · 2026-07-12T06:31:00.285558+00:00 · methodology

0 comments
read the original abstract

Automatic Speech Recognition (ASR) and Dialect Identification (DID) are crucial for Indian languages, many of which are low-resource and exhibit significant dialectal differences. Existing methods often optimize ASR or DID individually, resulting in performance trade-offs. In this work, we propose a multimodal framework that jointly improves ASR and DID. Our method employs a Bottleneck Encoder to extract dialectal features from Conformer-based speech representations and a RoBERTa encoder to process ASR-generated CTC embeddings. A gating mechanism merges these features, followed by an attention encoder to refine the representations. The learned embeddings are concatenated with Conformer outputs to enhance ASR features. Evaluated on eight Indian languages with thirty-three dialects, our method achieves an average DID accuracy of 81.63% and average CER and WER of 4.65% and 17.73%, respectively. These results highlight the effectiveness of our method for joint ASR-DID modeling.

Figures

Figures reproduced from arXiv: 2607.02862 by Amartyaveer, Prasanta Kumar Ghosh, Saurabh Kumar.

Figure 1
Figure 1. Figure 1: Illustration of the (a) proposed architecture and (b) bottleneck encoder (left) and attention encoder (right). 2. Proposed Method Our proposed method employs multimodal feature fusion to jointly enhance Automatic Speech Recognition (ASR) and Di￾alect Identification (DID) for Indian languages. As shown in Figure 1a, the architecture consists of two primary components: the ASR Block and the DID Block, integr… view at source ↗
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 1 canonical work pages

  1. [1]

    These results highlight the effectiveness of our method for joint ASR-DID modeling

    73%, respectively. These results highlight the effectiveness of our method for joint ASR-DID modeling. Index Terms: Automatic speech recognition, dialect identifica- tion, Indian languages

  2. [2]

    AS R sys- tems rely on robust acoustic and linguistic models to transc ribe speech accurately, while DID facilitates the adaptation of ASR models to different dialectal variations

    Introduction Automatic Speech Recognition (ASR) and Dialect Identifica- tion (DID) are crucial for advancing speech technology, par - ticularly in linguistically diverse regions like India. AS R sys- tems rely on robust acoustic and linguistic models to transc ribe speech accurately, while DID facilitates the adaptation of ASR models to different dialectal...

    Pith/arXiv arXiv 2026

  3. [3]

    Proposed Method Our proposed method employs multimodal feature fusion to jointly enhance Automatic Speech Recognition (ASR) and Di- alect Identification (DID) for Indian languages. As shown in Figure 1a, the architecture consists of two primary compone nts: the ASR Block and the DID Block, integrating speech and text- based dialectal cues for effective rep...

  4. [4]

    Experiments 3.1. Datasets We use a subset of the RESPIN dataset covering eight Indian languages: Bhojpuri (bh), Bengali (bn), Chhattisgarhi (ch ), Kannada (kn), Magahi (mg), Maithili (mt), Marathi (mr), and Telugu (te). The train-test splits follow [11] 2, with approxi- mately 140 − 175 hours of training data, 2 hours of develop- ment data, and 6 − 8 hour...

    2048

  5. [5]

    The evaluation con- siders DID accuracy, ASR Character Error Rate (CER), and Word Error Rate (WER)

    Results and Discussion In this section, we analyze the performance of existing meth - ods and our proposed multimodal feature fusion approach for joint DID and ASR in Indian languages. The evaluation con- siders DID accuracy, ASR Character Error Rate (CER), and Word Error Rate (WER). We also examine the impact of dialect- informed ASR and the effectivenes...

  6. [6]

    08%), Chhattisgarhi (ch: 79

    46% with an increase of 1. 08%), Chhattisgarhi (ch: 79. 38% with an increase of 2. 38%), and Telugu (te: 82. 23% with an increase of 1. 17%). These results underscore the advantages of multimodal fusion for DID. To further validate our approach, we compare the confu- sion matrices of the baseline and proposed methods in Figure 2. Our model not only improv...

  7. [7]

    By integrating a bottleneck encoder on Conformer outputs and a RoBERTa encoder on CTC embed- dings, our method improves both DID accuracy and ASR per- formance

    Conclusion We propose a novel multimodal feature fusion approach for joint dialect identification and automatic speech recognit ion in Indian languages. By integrating a bottleneck encoder on Conformer outputs and a RoBERTa encoder on CTC embed- dings, our method improves both DID accuracy and ASR per- formance. Experimental results demonstrate that our AS...

  8. [8]

    65% CER, and 17

    63% DID accuracy, 4. 65% CER, and 17. 73% WER. Addi- tionally, we analyze the impact of incorrect DID prediction s on ASR and show that our model effectively mitigates this degra - dation. Future work will explore the impact of multimodal fu - sion on multilingual, multi-dialect ASR in linguistically diverse scenarios, aiming to enhance adaptability acros...

  9. [9]

    We thank the RESPIN team and our project partner, Navana Tech, for their contributio ns to data collection

    Acknowledgment This work was supported by the RESPIN project, funded by the Bill & Melinda Gates Foundation. We thank the RESPIN team and our project partner, Navana Tech, for their contributio ns to data collection

  10. [10]

    Gated Multi Encoders and Multitask Objectives for Di - alectal Speech Recognition in Indian Languages,

    S. Udupa, J. Bandekar, G. Deekshitha, S. Kumar, P . K. Ghos h, S. Badiger, A. Singh, S. Murthy, P . Pai, S. Raghavan, and R. Nana- vati, “Gated Multi Encoders and Multitask Objectives for Di - alectal Speech Recognition in Indian Languages,” in Proc. of IEEE Automatic Speech Recognition and Understanding W ork- shop (ASRU), 2023, pp. 1–8

    2023

  11. [11]

    Spoken language identification using deep learning,

    G. Singh, S. Sharma, V . Kumar, M. Kaur, M. Baz, and M. Ma- sud, “Spoken language identification using deep learning,” Com- putational Intelligence and Neuroscience , vol. 2021, no. 1, p. 5123671, 2021

    2021

  12. [12]

    Deep learning for spoken language identifi cation,

    G. Montavon, “Deep learning for spoken language identifi cation,” in NIPS W orkshop on deep learning for speech recognition and related applications. Citeseer, 2009, pp. 1–4

    2009

  13. [13]

    Joint ASR and language identification using RNN-T: An effi- cient approach to dynamic language switching,

    S. Punjabi, H. Arsikere, Z. Raeesy, C. Chandak, N. Bhave, A. Bansal, M. M¨ uller, S. Murillo, A. Rastrow, A. Stolcke et al. , “Joint ASR and language identification using RNN-T: An effi- cient approach to dynamic language switching,” in Proc. ICASSP. IEEE, 2021, pp. 7218–7222

    2021

  14. [14]

    Ant Multilingual Recognitio n Sys- tem for OLR 2021 Challenge,

    A. Lyu, Z. Wang, and H. Zhu, “Ant Multilingual Recognitio n Sys- tem for OLR 2021 Challenge,” in Proc. Interspeech, 2022, pp. 3684–3688

    2021

  15. [15]

    Streaming End-to-End Multilingu al Speech Recognition with Joint Language Identification,

    C. Zhang, B. Li, T. Sainath, T. Strohman, S. Mavandadi, S. - Y . Chang, and P . Haghani, “Streaming End-to-End Multilingu al Speech Recognition with Joint Language Identification,” in Proc. Interspeech, 2022, pp. 3223–3227

    2022

  16. [16]

    Im- proving Massively Multilingual ASR with Auxiliary CTC Obje c- tives,

    W. Chen, B. Y an, J. Shi, Y . Peng, S. Maiti, and S. Watanabe, “Im- proving Massively Multilingual ASR with Auxiliary CTC Obje c- tives,” in Proc. ICASSP, 2023, pp. 1–5

    2023

  17. [17]

    A Configurable Multilin gual Model is All Y ou Need to Recognize All Languages,

    L. Zhou, J. Li, E. Sun, and S. Liu, “A Configurable Multilin gual Model is All Y ou Need to Recognize All Languages,” in Proc. ICASSP, 2022, pp. 6422–6426

    2022

  18. [18]

    End- to-end Japanese Multi-dialect Speech Recognition and Dial ect Identification with Multi-task Learning,

    R. Imaizumi, R. Masumura, S. Shiota, and H. Kiya, “End- to-end Japanese Multi-dialect Speech Recognition and Dial ect Identification with Multi-task Learning,” APSIPA Transactions on Signal and Information Processing , vol. 11, no. 1, pp. –, 2022. [Online]. Available: http://dx.doi.org/10.1561/116.00000045

    work page doi:10.1561/116.00000045 2022

  19. [19]

    Low-resource speech recognition and dialect ident ification of Irish in a multi-task framework,

    L. Lonergan, M. Qian, N. N. Chiar´ ain, C. Gobl, and A. N. C ha- saide, “Low-resource speech recognition and dialect ident ification of Irish in a multi-task framework,” in The Speaker and Language Recognition W orkshop (Odyssey), 2024, pp. 67–73

    2024

  20. [20]

    Im- proving Dialect Identification in Indian Languages Using Mu lti- modal Features from Dialect Informed ASR,

    Amartyaveer, S. Kumar, S. Sharma, S. Udupa, S. Badiger, A. Singh, D. G, J. Bandekar, S. Murthy, and P . Kumar Ghosh, “Im- proving Dialect Identification in Indian Languages Using Mu lti- modal Features from Dialect Informed ASR,” in Proc. ICASSP, 2025, pp. 1–5

    2025

  21. [21]

    Building an ASR System for Indian (Punjabi) language and its evaluation for Malwa and Majha dialect: Preliminary Results,

    V . Bhardwaj, V . Kukreja, N. Kaur, and N. Modi, “Building an ASR System for Indian (Punjabi) language and its evaluation for Malwa and Majha dialect: Preliminary Results,” in Proc. of In- ternational Conference on Computing Communication and Net - working Technologies (ICCCNT), 2021, pp. 1–5

    2021

  22. [22]

    Multilingual low resource Indian language speech recogni tion and spell correction using Indic BERT,

    M. C. S. Priya, D. K. Renuka, L. A. Kumar, and S. L. Rose, “Multilingual low resource Indian language speech recogni tion and spell correction using Indic BERT,” S¯ adhan¯ a, vol. 47, no. 227, 2022

    2022

  23. [23]

    DuDe: Dual- Decoder Multilingual ASR for Indian Languages using Common Label Set,

    A. Arunkumar, M. D. Batra, and S. Umesh, “DuDe: Dual- Decoder Multilingual ASR for Indian Languages using Common Label Set,” ArXiv, vol. abs/2210.16739, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:253237346

    Pith/arXiv arXiv 2022

  24. [24]

    ASRoIL: a comprehensive survey for automatic speech recognition of Indian languages,

    A. Singh, V . Kadyan, M. Kumar, and N. Bassan, “ASRoIL: a comprehensive survey for automatic speech recognition of Indian languages,” Artif. Intell. Rev. , vol. 53, no. 5, p. 3673–3704, June 2020. [Online]. Available: https://doi.o rg/10. 1007/s10462-019-09775-8

    2020

  25. [25]

    MU CS 2021: Multilingual and Code-Switching ASR Challenges for L ow Resource Indian Languages,

    A. Diwan, R. V aideeswaran, S. Shah, A. Singh, S. Raghava n, S. Khare, V . Unni, S. Vyas, A. Rajpuria, C. Y arra, A. Mittal, P . K. Ghosh, P . Jyothi, K. Bali, V . Seshadri, S. Sitaram, S. Bharad- waj, J. Nanavati, R. Nanavati, and K. Sankaranarayanan, “MU CS 2021: Multilingual and Code-Switching ASR Challenges for L ow Resource Indian Languages,” in Proc...

    2021

  26. [26]

    Exploring the Impact of Back-End Net work on Wav2vec 2.0 for Dialect Identification,

    Q. Luo and R. Zhou, “Exploring the Impact of Back-End Net work on Wav2vec 2.0 for Dialect Identification,” in Proc. Interspeech, 2023, pp. 5356–5360

    2023

  27. [27]

    On the Robustness of Arabic Speech Dialect Identification,

    Peter Sullivan and AbdelRahim Elmadany and Muhammad Abdul-Mageed, “On the Robustness of Arabic Speech Dialect Identification,” in Proc. Interspeech, 2023, pp. 5326–5330

    2023

  28. [28]

    North S´ ami Dia lect Identification with Self-supervised Speech Models,

    S. Kakouros and K. Hiovain-Asikainen, “North S´ ami Dia lect Identification with Self-supervised Speech Models,” in Proc. In- terspeech, 2023, pp. 5306–5310

    2023

  29. [29]

    Self-Supervised Adaptive Pre-Training of Multil ingual Speech Models for Language and Dialect Identification,

    Shaik, Mohammed Maqsood and Klakow, Dietrich and Abdul lah, Badr M., “Self-Supervised Adaptive Pre-Training of Multil ingual Speech Models for Language and Dialect Identification,” in Proc. ICASSP, 2024, pp. 11 436–11 440

    2024

  30. [30]

    A Parameter-Efficient Learning Approach to Arabic Dialect Identification with Pre-Trained General-Pu rpose Speech Model,

    Srijith Radhakrishnan and Chao-Han Huck Y ang and Sumee r Ah- mad Khan and Narsis A. Kiani and David Gomez-Cabrero and Jesper N. Tegner, “A Parameter-Efficient Learning Approach to Arabic Dialect Identification with Pre-Trained General-Pu rpose Speech Model,” in Proc. Interspeech, 2023, pp. 1958–1962

    2023

  31. [31]

    Arabic Dialect Identificatio n for Travel and Twitter Text,

    P . Mishra and V . Mujadia, “Arabic Dialect Identificatio n for Travel and Twitter Text,” in Proc. of the F ourth Arabic Natural Language Processing W orkshop. Florence, Italy: Association for Computational Linguistics, August 2019, pp. 234–238. [Online]. Available: https://aclanthology.org/W19-4628

    2019

  32. [33]

    Available: https://arxiv.org/abs/2009.12622

    [Online]. Available: https://arxiv.org/abs/2009.12622

    Pith/arXiv arXiv 2009

  33. [34]

    Unsupervised deep language and dialect identifica tion for short texts,

    K. Goswami, R. Sarkar, B. R. Chakravarthi, T. Fransen, a nd J. P . McCrae, “Unsupervised deep language and dialect identifica tion for short texts,” in Proceedings of the 28th International Confer- ence on Computational Linguistics , 2020, pp. 1606–1617

    2020

  34. [35]

    Bottleneck Transformers for Visual Recogniti on,

    A. Srinivas, T.-Y . Lin, N. Parmar, J. Shlens, P . Abbeel, and A. V aswani, “Bottleneck Transformers for Visual Recogniti on,” in Proc. of IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR) , 2021, pp. 16 514–16 524

    2021

  35. [36]

    Disentangled Repre- sentations in Local-Global Contexts for Arabic Dialect Ide ntifica- tion,

    Z. Alhakeem, S.-I. Jang, and H.-G. Kang, “Disentangled Repre- sentations in Local-Global Contexts for Arabic Dialect Ide ntifica- tion,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 879–890, 2024

    2024

  36. [37]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach,

    Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov, “RoBERTa: A Robustly Optimized BERT Pretraining Approach,” ArXiv, vol. abs/1907.11692, 2019. [Online]. Available: https: //api.semanticscholar.org/CorpusID:198953378

    Pith/arXiv arXiv 1907

  37. [38]

    Towards Building ASR Systems for the Next Billion Users,

    T. Javed, S. Doddapaneni, A. Raman, K. S. Bhogale, G. Ramesh, A. Kunchukuttan, P . Kumar, and M. M. Khapra, “Towards Building ASR Systems for the Next Billion Users,” CoRR, vol. abs/2111.03945, 2021. [Online]. Available: https: //arxiv.org/abs/2111.03945

    Pith/arXiv arXiv 2021