Pith. sign in

REVIEW 3 major objections 5 minor 55 references

NIRANTAR: Continual Learning with New Languages and Domains on Real-world Speech Data

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Nirantar, a 3,250-hour benchmark of naturally incremental Indian speech across 22 languages and 208 districts, shows that no existing continual-learning method handles language and domain shifts consistently.

desk verdict A genuinely useful ASR continual-learning benchmark with 1,720 new hours and a novel LIDIL scenario, but the 'natural episodes' claim is undercut by random assignment of data to episodes. read the letter →

arxiv 2507.00534 v1 pith:DPRYJOYM submitted 2025-07-01 cs.CL

classification cs.CL
keywords continuallearningautomaticspeechrecognitionmultilingualASRlow-resourcelanguageslanguage-incrementaldomain-incrementalLIDILbenchmark
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

The paper introduces Nirantar, a continual-learning benchmark for speech recognition built from real data collection rounds rather than artificially split datasets. It provides 3,250 hours of human-transcribed speech across 22 Indian languages and 208 districts, arranged into 12 natural episodes. These episodes support three evaluation setups: adding languages one at a time, adding districts one at a time, and adding both at once, a scenario the paper calls LIDIL. The central message is that none of the standard continual-learning methods performs well in all three setups, so the field still lacks a robust recipe for real-world multilingual speech recognition. If true, Nirantar gives researchers a common ground to develop and test such recipes.

What carries the argument

The central object is the episode timeline $T = \langle E_0, \dots, E_\tau \rangle$: a data batch $B = (l, d)$ is speech from a district $d$ in language $l$, and an episode groups batches collected in one period. Three timelines are cut from the same corpus—one that adds a language per episode, one that adds districts while all languages are present from $E_0$, and one that adds both languages and districts (LIDIL). The district-as-domain choice is what makes these timelines natural rather than synthetic, and the LIDIL scenario is the new evaluation mode the paper contributes. The evaluation machinery consists of AMER (average match error rate across all seen episodes), forward transfer, backward transfer, and intransigence, computed against a Conformer-L encoder with a CTC-RNNT decoder trained both incrementally and jointly.

What would settle it

Rerun the DIL and LIDIL evaluations with districts randomly re-assigned within each language to invented domain labels, holding hours and episode order fixed; if method rankings and AMER curves barely move, the district-as-domain premise is not doing the work the paper assigns it.

Watch

Extended reading notes

Core claim

On its own terms, the paper discovers that a large, genuinely episodic multilingual speech corpus can be organized into three continual-learning scenarios and that standard continual-learning methods fail it unevenly. Experience Replay is the strongest tested method in language-incremental learning, regularization-based methods such as EWC and MAS become competitive when only new domains are added, and the combined LIDIL scenario initially behaves like language-incremental learning and then like domain-incremental learning. Adapter-based methods prevent forgetting but add a prohibitive parameter for each language and domain, making them impractical at the scale of 22 languages and 208 districts. The paper treats this uneven performance as evidence that no single method is adequate and that LIDIL, where both new languages and new domains arrive together, remains an open problem.

Load-bearing premise

The claim depends on the assumption that each Indian district is a distinct ASR domain whose unique accents, vocabulary, and topics create a distribution shift comparable to what a deployed system would meet, and that the natural order of data collection is a fair deployment order.

Editorial extensions

If this is right

  • Replay with a 3% buffer is the strongest tested strategy for adding new languages, beating EWC and MAS on average error rate and backward transfer.
  • In domain-only addition, where the base model already covers all languages, methods improve over episodes and regularization becomes competitive, suggesting domain expansion is the easier scenario.
  • When languages and domains arrive together, the learning curve is initially language-incremental-like and then domain-incremental-like, confirming a genuinely hybrid scenario that existing benchmarks do not isolate.
  • Adapter-based continual learning is impractical at Nirantar's scale because supporting 22 languages and 208 domains would require one adapter per group and millions of extra parameters.
  • No single method is consistently good across the three scenarios, so the benchmark can serve as a common testbed for more robust continual-learning approaches.

Reading between the lines

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

  • If the district-as-domain assumption holds, the same natural-episode construction could be applied to other geographically organized speech corpora, turning collection logs into continual-learning benchmarks without synthetic splits.
  • A natural next method suggested by the results is a hybrid controller that uses replay when a new language arrives and regularization when a new district arrives; the LIDIL curves imply such switching is worth testing.
  • The paper reports only three randomized LIDIL sequences; varying the composition of the base episode's languages and domains is an untested direction that could change the ranking of methods.
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

3 major / 5 minor

Summary. Nirantar is a dataset-and-benchmark paper for continual learning (CL) in multilingual ASR. It extends IndicVoices with 1,720 new hours of human-transcribed speech, for a total of 3,250 hours across 22 Indian languages and 208 (or 214, or 400) districts. The paper defines three CL scenarios: LIL, DIL, and a new LIDIL scenario in which both languages and domains arrive over 12 episodes. It evaluates incremental finetuning, EWC, MAS, ER, and adapters on a Conformer-L hybrid CTC-RNNT model using AMER, FWT, BWT, and IM. The main reported finding is that no single CL method performs consistently well across the three scenarios.

Significance. If the benchmark is as described, it would be a valuable resource: it is one of the few human-transcribed, low-resource multilingual ASR corpora organized for CL evaluation, and LIDIL is a useful scenario that combines language and domain shifts. The release of code, data, and models under CC-BY-4.0 is a concrete strength, and the use of standard CL metrics with MER is sensible. However, the paper's central novelty claim of 'natural episodes' is not supported by the episode-construction procedure in Section 3.4, and the headline district count is inconsistent across the paper; these issues must be resolved before the benchmark claims can be fully assessed.

major comments (3)
  1. [Abstract and §3.4] The abstract and introduction claim that Nirantar leverages data collected incrementally through 'natural episodes,' but Section 3.4 states that for DIL and LIDIL 'data batches are randomly assigned to episodes' and that LIL introduces languages 'in random order.' The timelines are therefore constructed re-orderings of a static corpus, not observed incremental deployment episodes; only the selection of E0 uses the collection structure. This contradicts the paper's stated novelty and means the reported AMER/BWT/FWT/IM trajectories in Figure 3 describe one arbitrary schedule. The randomized-sequence analysis mentioned in Section 5 is precisely the evidence needed to validate the conclusions, yet it is omitted; it should be reported, or the 'natural episodes' claims should be reworded.
  2. [Table 2 / §3.3 / Introduction] District counts are inconsistent: the Introduction says 400 districts, the Abstract and Section 3.3 say 208, and the #D column of Table 2 sums to 214. Since the number of domains is a headline quantity and is used to derive the DIL/LIDIL episode structures in Table 3, the correct number and the source of the discrepancies must be clarified.
  3. [§5 / Figure 3] The main comparative claim—'no single method performs consistently well'—is supported only by single-run trajectories in Figure 3. No error bars, number of seeds, or confidence intervals are reported. For a benchmark intended to support method selection, this is insufficient; at minimum, report variance across the three LIDIL random orders mentioned in Section 5 and ideally across multiple LIL and DIL orders, since those orders are constructed rather than observed.
minor comments (5)
  1. [Figure 3] The middle-panel caption reads 'Domain Incremental Learning (LIL)' but should read 'Domain Incremental Learning (DIL)'.
  2. [§2] The sentence 'However, real-world scenarios remain similar to ours remain underexplored' is ungrammatical and should be rewritten.
  3. [§4.3] The definition of FWT is ambiguous: it defines M ER_inc.ft_t but does not specify whether this is the MER of the incrementally fine-tuned model at episode t evaluated on episode t, or the MER of a model trained only on episode t. Please state the exact quantity.
  4. [References] Reference [16] cites 'Overcoming Catastrophic Forgetting in Graph Neural Networks' (AAAI 2021) for Elastic Weight Consolidation; the correct citation is Kirkpatrick et al., 'Overcoming catastrophic forgetting in neural networks,' PNAS 2017.
  5. [§3.3] The test-set description says 'a maximum of 15 minutes from each of the domains,' while the abstract says 'per domain-language pair.' Since a district is associated with a specific language, clarify whether the 15-minute cap is per district-language pair and how the resulting total of 50 hours is computed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark dataset, the three CL scenarios, and the baseline comparison are externally grounded, and no claim reduces to its own inputs; the 'natural episodes' wording is contradicted by the random timeline construction, but that is a validity caveat rather than a circular step.

full rationale

Nirantar is an empirical benchmark paper rather than a derivation, so the circularity patterns that matter here (self-definitional equations, fitted inputs renamed as predictions, imported uniqueness theorems, ansatz smuggled by self-citation) largely do not apply. The load-bearing claims are: (1) a 3,250-hour dataset of which 1,720 hours are newly released; (2) three CL scenarios including the new LIDIL; (3) an empirical finding that no single CL method performs consistently well. The dataset claim is externally verifiable: the new hours are released under CC-BY-4.0, the base hours come from the publicly released IndicVoices corpus (ref. [14]), and Section 3.3 states the collection procedure (local coordinators, three recording tasks, in-house transcription) directly in this paper rather than importing it solely by citation. The CL comparison inherits hyperparameters from an external benchmark, CL-MASR (ref. [21]) — EWC lambda=5 and alpha=0.5, MAS alpha=1 and lambda=0.5, ER with a 3% buffer — and evaluates on held-out test samples (15 minutes per domain-language pair, Section 3.3) using standard metrics AMER, FWT, BWT, and IM (Section 4.3). No parameter is fitted to the evaluation data and then reported as a prediction; the 'no single method performs consistently well' conclusion is an observed ordering of independent baselines, not an output forced by construction. Two caveats belong to validity and evidence rather than circularity. First, the abstract's 'natural episodes' claim is not supported by Section 3.4, which randomly orders languages in LIL and randomly assigns data batches to episodes in DIL and LIDIL, so the reported trajectories may depend on one arbitrary schedule. Second, the paper's own robustness check over three randomized LIDIL sequences is asserted but omitted ('Due to space constraints, these results are not included'), leaving that dependence unverified. Neither caveat is a reduction of a claimed result to its own inputs, so the circularity score is 0.

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

The central deliverable is a dataset, so most assumptions concern what counts as a domain and what counts as a natural episode. No free parameters are fitted to force a derivation; the listed hyperparameters are inherited from prior work and affect the comparative conclusion that no method wins everywhere.

free parameters (4)
  • EWC hyperparameters = lambda = 5, alpha = 0.5
    Set following CL-MASR (reference [21]); not fitted in this paper, but influences regularization-based results that contribute to the claim that no single method wins.
  • MAS hyperparameters = alpha = 1, lambda = 0.5
    Also inherited from CL-MASR (reference [21]); affects the regularization-based comparison in DIL and LIDIL.
  • Experience replay buffer size = 3% of data per episode
    Chosen following previous work; the buffer size materially affects the replay method's performance in LIL.
  • Adapter bottleneck dimension = 64
    Hand-chosen network size for the LIL adapter experiments; affects the reported parameter growth and adapter results.
assumptions (4)
  • domain assumption A district defines a distinct domain for ASR.
    Section 3.3: each district is treated as a distinct domain due to unique vocabulary, accents, and local interests. The DIL and LIDIL scenarios depend on this.
  • domain assumption Natural collection phases map directly to continual learning episodes.
    Section 1 and Section 3.4 argue that periodic gaps between data collection batches create a natural CL setting. The benchmark's validity depends on this mapping.
  • domain assumption Human transcriptions are accurate enough for ASR evaluation.
    Section 3.3 mentions an in-house transcription team and rigorous quality control, but no inter-annotator agreement or error rate is reported.
  • domain assumption Hyperparameters from prior CL-MASR work transfer to this dataset.
    Section 4.1 sets EWC, MAS, and ER hyperparameters by following [21], which may not be optimal for Nirantar's 22 languages and 208 districts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NIRANTAR: Continual Learning with New Languages and Domains on Real-world Speech Data." pith.science (2026). https://pith.science/paper/DPRYJOYM

@misc{pith2026250700534,
  author       = {Pith},
  title        = {Pith review of: NIRANTAR: Continual Learning with New Languages and Domains on Real-world Speech Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DPRYJOYM}},
  note         = {Machine review of arXiv:2507.00534}
}
read the original abstract

We introduce Nirantar, a comprehensive framework for evaluating continual learning (CL) in multilingual and multi-domain ASR. Designed to reflect real-world CL challenges, Nirantar leverages data collected incrementally across 22 languages and 208 districts in India through natural episodes. This enables evaluation across Language-Incremental (LIL), Domain-Incremental (DIL), and the novel Language-Incremental Domain-Incremental Learning (LIDIL) scenarios. Unlike prior work that relies on simulated episodes, Nirantar presents dynamic, non-uniform language and domain shifts, making it an ideal testbed for CL research. With 3250 hours of human-transcribed speech, including 1720 hours newly introduced in this work, our framework enables systematic benchmarking of CL methods. We evaluate existing approaches and demonstrate that no single method performs consistently well, underscoring the need for more robust CL strategies.

Figures

Figures reproduced from arXiv: 2507.00534 by the authors.

Figure 1
Figure 1. Illustration of Language-Incremental Domain￾Incremental Learning: A practical scenario showing the ad￾dition of both new languages and domains in each episode of speech data collection. The episodic nature of this data collection, with periodic gaps between batches, creates a natural setting for continual learning (CL). Leveraging this, we introduce Nirantar, a CL framework designed for three scenarios: Language-Inc… view at source ↗
Figure 2
Figure 2. Evolution of vocabulary and domains across episodes a maximum of 15 minutes from each of the domains resulting in a total of 50 hours across languages. Since the test data con￾tains samples from every district, we can evaluate the forward and backward transfer of CL approaches. 3.4. Continual Learning Playground Nirantar comprises of three distinct timelines corresponding to LIL, DIL and LIDIL scenarios respectively… view at source ↗
Figure 3
Figure 3. Comparison of various CL methods: (top) Lan￾guage Incremental Learning (LIL), (middle) Domain Incre￾mental Learning (LIL) and (bottom) Language-Incremental Domain-Incremental Learning (LIDIL) and Backward Transfer in DIL, indicating that regularization￾based methods are well-suited for domain-incremental learning. LIDIL: In [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 51 canonical work pages

  1. [1]

    NIRANTAR: Continual Learning with New Languages and Domains on Real-world Speech Data

    Introduction There is a growing trend towards training massive multilingual speech models on large datasets [1, 2] aggregated across multi- ple languages [3, 4, 5]. Given the high computational demands, continual training is essential as new datasets covering addi- tional languages, domains, or demographics are introduced over time [1, 6]. To address this...

  2. [2]

    Prior work includes domain-specific ASR sub-models [19] and monolingual hybrid CTC-transformer adaptation [20], both fo- cusing on domain-incremental setups

    Related work Continual Learning (CL) in ASR has been explored mainly in Language-Incremental and Domain-Incremental Learning [18]. Prior work includes domain-specific ASR sub-models [19] and monolingual hybrid CTC-transformer adaptation [20], both fo- cusing on domain-incremental setups. CL-MASR [21] ex- amines CL strategies in a multilingual setting, emp...

  3. [3]

    We now introduce definitions which will be used through the paper

    N IRANTAR : CL on Real-World Data This section introduces Nirantar, a playground for continual learning in ASR with new languages and domains. We now introduce definitions which will be used through the paper. 3.1. Definitions Data Batch (B): A data batch, represented as an ordered tuple B = (l, d), is the outcome of a single data collection activity for ...

  4. [4]

    Continual Learning Methods Below, we list down all the approaches considered in this work

    Experimental setup 4.1. Continual Learning Methods Below, we list down all the approaches considered in this work. Incremental Finetuning (Inc. FT): Given a base model m0, we sequentially finetune modelsm1≤t≤τ using the data batches Table 3: Statistics showing district counts per language in LIL, DIL and LIDIL scenarios. Each cell carries the number of di...

  5. [5]

    Results and Discussions LIL: Referring to Figure 3 (top), we observe a steady increase in AMER as new languages are introduced for Incremental FT, which is undesirable. Both regularization-based approaches, EWC and MAS, struggle to retain knowledge of previously learned languages, as shown by the trends in the Forward Trans- fer (FWT) across episodes. In ...

  6. [6]

    Conclusion We presented Nirantar, a novel data framework designed to fa- cilitate training and evaluation of continual learning (CL) meth- ods in multilingual and multidomain settings. This dataset con- tains 3250 hours of human-transcribed speech data, including 1720 hours released from this study, organized into 12 episodes featuring diverse language an...

  7. [8]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila et al. , “Common voice: A massively-multilingual speech corpus,” in LREC. European Language Resources As- sociation, 2020, pp. 4218–4222

  8. [9]

    V oxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation,

    C. Wang et al. , “V oxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation,” in ACL/IJCNLP (1). ACL, 2021, pp. 993–1003

Show all 55 references
  1. [10]

    Pseudo-labeling for massively multilingual speech recognition,

    L. Lugosch et al. , “Pseudo-labeling for massively multilingual speech recognition,” in ICASSP. IEEE, 2022, pp. 7687–7691

  2. [11]

    In LIL, each episode adds all data batches of a single lan- guage, with languages introduced in random order. For DIL and LIDIL, data batches are randomly assigned to episodes, ensur- ing uniform distribution of data batches while still maintaining non-uniformity in training h...

  3. [12]

    Robust speech recognition via large-scale weak supervision,

    A. Radford et al., “Robust speech recognition via large-scale weak supervision,” in ICML, vol. 202. PMLR, 2023, pp. 28 492– 28 518

  4. [13]

    Google USM: scaling automatic speech recogni- tion beyond 100 languages,

    Y . Zhang et al., “Google USM: scaling automatic speech recogni- tion beyond 100 languages,” CoRR, vol. abs/2303.01037, 2023

  5. [14]

    SPRING-INX: A multilingual indian language speech corpus,

    N. R et al. , “SPRING-INX: A multilingual indian language speech corpus,” CoRR, vol. abs/2310.14654, 2023

  6. [15]

    A wholistic view of continual learning with deep neural networks: Forgotten lessons and the bridge to active and open world learning,

    M. Mundt et al., “A wholistic view of continual learning with deep neural networks: Forgotten lessons and the bridge to active and open world learning,” Neural Networks , vol. 160, pp. 306–336, 2023

  7. [16]

    An empirical investigation of catastrophic forgeting in gradient-based neural networks,

    I. J. Goodfellow et al., “An empirical investigation of catastrophic forgeting in gradient-based neural networks,” in ICLR (Poster), 2014

  8. [17]

    Continual learning through synaptic intelligence,

    F. Zenke et al., “Continual learning through synaptic intelligence,” in ICML, vol. 70. PMLR, 2017, pp. 3987–3995

  9. [18]

    The CLEAR benchmark: Continual learning on real-world imagery,

    Z. Lin et al. , “The CLEAR benchmark: Continual learning on real-world imagery,” inNeurIPS Datasets and Benchmarks, 2021

  10. [19]

    Learning multiple visual domains with residual adapters,

    S. Rebuffi et al., “Learning multiple visual domains with residual adapters,” in NIPS, 2017, pp. 506–516

  11. [20]

    Learn continually, generalize rapidly: Life- long knowledge accumulation for few-shot learning,

    X. Jin et al. , “Learn continually, generalize rapidly: Life- long knowledge accumulation for few-shot learning,” in EMNLP (Findings). ACL, 2021, pp. 714–729

  12. [21]

    Indicvoices: Towards building an inclusive mul- tilingual speech dataset for indian languages,

    T. Javed et al., “Indicvoices: Towards building an inclusive mul- tilingual speech dataset for indian languages,” in ACL (Findings). ACL, 2024, pp. 10 740–10 782

  13. [22]

    Experience replay for continual learning,

    D. Rolnick et al. , “Experience replay for continual learning,” in NeurIPS, 2019, pp. 348–358

  14. [23]

    Overcoming catastrophic forgetting in graph neural networks,

    H. Liu et al., “Overcoming catastrophic forgetting in graph neural networks,” in AAAI. AAAI Press, 2021, pp. 8653–8661

  15. [24]

    Memory aware synapses: Learning what (not) to forget,

    R. Aljundi et al., “Memory aware synapses: Learning what (not) to forget,” inECCV (3), vol. 11207. Springer, 2018, pp. 144–161

  16. [25]

    Three types of incremental learning,

    G. M. van de Ven et al. , “Three types of incremental learning,” Nat. Mac. Intell., vol. 4, no. 12, pp. 1185–1197, 2022

  17. [26]

    Continual learning in automatic speech recogni- tion,

    S. Sadhu et al., “Continual learning in automatic speech recogni- tion,” in INTERSPEECH. ISCA, 2020, pp. 1246–1250

  18. [27]

    Towards lifelong learning of end-to-end ASR,

    H. Chang et al., “Towards lifelong learning of end-to-end ASR,” in Interspeech. ISCA, 2021, pp. 2551–2555

  19. [28]

    CL-MASR: A continual learning benchmark for multilingual ASR,

    L. D. Libera et al., “CL-MASR: A continual learning benchmark for multilingual ASR,” IEEE ACM Trans. Audio Speech Lang. Process., vol. 32, pp. 4931–4944, 2024

  20. [29]

    Core50: a new dataset and benchmark for continuous object recognition,

    V . Lomonaco et al. , “Core50: a new dataset and benchmark for continuous object recognition,” in CoRL, vol. 78. PMLR, 2017, pp. 17–26

  21. [30]

    Multi-label continual learning for the medical domain: A novel benchmark,

    M. Ceccon et al., “Multi-label continual learning for the medical domain: A novel benchmark,” CoRR, vol. abs/2404.06859, 2024

  22. [31]

    A comprehensive survey of continual learn- ing: Theory, method and application,

    L. Wang et al. , “A comprehensive survey of continual learn- ing: Theory, method and application,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 8, pp. 5362–5383, 2024

  23. [32]

    Dark experience for general continual learning: a strong, simple baseline,

    P. Buzzega et al., “Dark experience for general continual learning: a strong, simple baseline,” in NeurIPS, 2020

  24. [33]

    Efficient lifelong learning with A-GEM,

    A. Chaudhry et al., “Efficient lifelong learning with A-GEM,” in ICLR (Poster). OpenReview.net, 2019

  25. [34]

    Using adapters to overcome catastrophic for- getting in end-to-end automatic speech recognition,

    S. V . Eeckt et al., “Using adapters to overcome catastrophic for- getting in end-to-end automatic speech recognition,” in ICASSP. IEEE, 2023, pp. 1–5

  26. [35]

    Progressive neural networks,

    A. A. Rusu et al., “Progressive neural networks,” arXiv preprint arXiv:1606.04671, 2016

  27. [36]

    Packnet: Adding multiple tasks to a single net- work by iterative pruning,

    A. Mallya et al., “Packnet: Adding multiple tasks to a single net- work by iterative pruning,” in CVPR. Computer Vision Founda- tion / IEEE Computer Society, 2018, pp. 7765–7773

  28. [37]

    Interspeech 2018 low resource au- tomatic speech recognition challenge for indian languages,

    B. M. L. Srivastava et al. , “Interspeech 2018 low resource au- tomatic speech recognition challenge for indian languages,” in SLTU. ISCA, 2018, pp. 11–14

  29. [38]

    Crowd-sourced speech corpora for ja- vanese, sundanese, sinhala, nepali, and bangladeshi bengali,

    O. Kjartansson et al. , “Crowd-sourced speech corpora for ja- vanese, sundanese, sinhala, nepali, and bangladeshi bengali,” in SLTU. ISCA, 2018, pp. 52–55

  30. [39]

    Open-source multi-speaker speech corpora for building gujarati, kannada, malayalam, marathi, tamil and telugu speech synthesis systems,

    F. He et al. , “Open-source multi-speaker speech corpora for building gujarati, kannada, malayalam, marathi, tamil and telugu speech synthesis systems,” in LREC. European Language Re- sources Association, 2020, pp. 6494–6503

  31. [40]

    MUCS 2021: Multilingual and code-switching ASR challenges for low resource indian languages,

    A. Diwan et al., “MUCS 2021: Multilingual and code-switching ASR challenges for low resource indian languages,” in Inter- speech. ISCA, 2021, pp. 2446–2450

  32. [41]

    Indicsuperb: A speech processing universal per- formance benchmark for indian languages,

    T. Javed et al., “Indicsuperb: A speech processing universal per- formance benchmark for indian languages,” in AAAI. AAAI Press, 2023, pp. 12 942–12 950

  33. [42]

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

    K. S. Bhogale et al., “Effectiveness of mining audio and text pairs from public data for improving ASR systems for low-resource languages,” in ICASSP. IEEE, 2023, pp. 1–5

  34. [43]

    Gram vaani ASR challenge on spontaneous telephone speech recordings in regional variations of hindi,

    A. Bhanushali et al., “Gram vaani ASR challenge on spontaneous telephone speech recordings in regional variations of hindi,” in INTERSPEECH. ISCA, 2022, pp. 3548–3552

  35. [44]

    Subword dictionary learning and segmenta- tion techniques for automatic speech recognition in tamil and kan- nada,

    M. Ayyavu et al. , “Subword dictionary learning and segmenta- tion techniques for automatic speech recognition in tamil and kan- nada,” CoRR, vol. abs/2207.13331, 2022

  36. [45]

    Automatic speech recognition in sanskrit: A new speech corpus and modelling insights,

    D. Adiga et al., “Automatic speech recognition in sanskrit: A new speech corpus and modelling insights,” in ACL/IJCNLP (Find- ings). ACL, 2021, pp. 5039–5050

  37. [46]

    The IIIT-H indic speech databases,

    K. Prahallad et al. , “The IIIT-H indic speech databases,” in IN- TERSPEECH. ISCA, 2012, pp. 2546–2549

  38. [47]

    Crowdsourcing speech data for low-resource languages from low-income workers,

    B. Abraham et al., “Crowdsourcing speech data for low-resource languages from low-income workers,” in LREC. European Lan- guage Resources Association, 2020, pp. 2819–2826

  39. [48]

    Vistaar: Diverse benchmarks and training sets for indian language ASR,

    K. S. Bhogale et al., “Vistaar: Diverse benchmarks and training sets for indian language ASR,” inINTERSPEECH. ISCA, 2023, pp. 4384–4388

  40. [49]

    Resources for Indian languages,

    A. Baby et al. , “Resources for Indian languages,” in CBBLR – Community-Based Building of Language Resources. Tribun EU, 2016, pp. 37–43

  41. [50]

    Svarah: Evaluating english ASR systems on in- dian accents,

    T. Javed et al., “Svarah: Evaluating english ASR systems on in- dian accents,” in INTERSPEECH. ISCA, 2023, pp. 5087–5091

  42. [51]

    SPIRE-SIES: A spontaneous indian english speech corpus,

    A. Singh et al. , “SPIRE-SIES: A spontaneous indian english speech corpus,” in O-COCOSDA. IEEE, 2023, pp. 1–6

  43. [52]

    LAHAJA: A robust multi-accent benchmark for evaluating hindi ASR systems,

    T. Javed et al. , “LAHAJA: A robust multi-accent benchmark for evaluating hindi ASR systems,” CoRR, vol. abs/2408.11440, 2024

  44. [53]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati et al., “Conformer: Convolution-augmented transformer for speech recognition,” in Interspeech. ISCA, 2020, pp. 5036– 5040

  45. [54]

    Stateful conformer with cache-based inference for streaming automatic speech recognition,

    V . Noroozi et al., “Stateful conformer with cache-based inference for streaming automatic speech recognition,” in ICASSP. IEEE, 2024, pp. 12 041–12 045

  46. [55]

    Uncertainty-aware balancing for multilingual and multi-domain neural machine translation training,

    M. Wu et al., “Uncertainty-aware balancing for multilingual and multi-domain neural machine translation training,” inEMNLP (1). Association for Computational Linguistics, 2021, pp. 7291–7305

  47. [56]

    From WER and RIL to MER and WIL: im- proved evaluation measures for connected speech recognition,

    A. C. Morris et al., “From WER and RIL to MER and WIL: im- proved evaluation measures for connected speech recognition,” in INTERSPEECH. ISCA, 2004, pp. 2765–2768

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.