Recognizing Every Voice: Towards Inclusive ASR for Rural Bhojpuri Women

Eldho Ittan George; Kaushal Bhogale; Mitesh M. Khapra; Nikhil Narasimhan; Sakshi Joshi; Tahir Javed

REVIEW 3 major objections 6 minor 32 references

Recognizing Every Voice: Towards Inclusive ASR for Rural Bhojpuri Women

T0 review · 3 major / 6 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 The paper shows that 25–30 seconds of seed audio per speaker, synthesized by a multilingual TTS model that never saw Bhojpuri, can produce 100 hours of training data and lower the word error rate on rural Bhojpuri women's speech by 4.7…

desk verdict A genuinely useful new benchmark for rural Bhojpuri ASR, but the headline 4.7 WER gain is not the synthetic data's isolated effect. read the letter →

arxiv 2506.09653 v1 pith:TEK7J5A6 submitted 2025-06-11 eess.AS

Sakshi Joshi , Eldho Ittan George , Tahir Javed , Kaushal Bhogale , Nikhil Narasimhan , Mitesh M. Khapra This is my paper

classification eess.AS

keywords automaticspeechrecognitionBhojpurilow-resourcelanguagessynthetictext-to-speechvoicecloningruralwomendigitalinclusion

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 tries to establish that useful speech-recognition training data for a very low-resource language can be manufactured from tiny, socially feasible voice samples. It introduces SRUTI, a one-hour benchmark of real speech from rural Bhojpuri women across agriculture, health, government schemes, and finance, and finds that current ASR models achieve about 34% word error rate on it. To get around the social and cultural barriers to large-scale recording, the authors synthesize 100 hours of Bhojpuri and 100 hours of Hindi audio from 25–30 second clips per speaker, using a text-to-speech model that was never trained on Bhojpuri. Adding this synthetic audio to real training data lowers word error rate on SRUTI from 33.8 to 29.1, a 4.7-point average gain. If correct, the approach gives a scalable, minimally intrusive recipe for improving ASR in other underserved languages and for making voice-based digital services accessible to rural women.

What carries the argument

The mechanism is a multilingual prompt-based text-to-speech system trained on 11 Indian languages, used zero-shot for Bhojpuri because Bhojpuri was absent from its training data. It takes a 25–30 second transcripted clip from a target speaker plus a new Bhojpuri sentence and returns synthetic speech in that speaker's voice, relying on transfer from related languages such as Hindi. This machinery converts a short, culturally feasible recording session into 100 hours of acoustic training data, and the same pipeline generates the synthetic Hindi audio from rural Hindi-speaking women.

What would settle it

A native-speaker listening test on a random sample of the synthetic Bhojpuri hours would settle the phonetic-fidelity question; if Bhojpuri raters flag frequent mispronunciations, the 4.7-point WER gain is better explained by textual overlap between synthetic training data and the benchmark than by genuine acoustic transfer.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that zero-shot prompt-based speech synthesis can bridge the ASR data gap for a language the synthesizer never saw. With only 39.4 minutes of real Bhojpuri speech from 100 women and short Hindi clips from 100 rural women, the authors generate 100 hours of synthetic Bhojpuri and 100 hours of synthetic Hindi. Adding these to the existing Bhojpuri corpora plus 376 hours of real Hindi improves average WER on SRUTI from 33.8 to 29.1, with per-domain gains of 2.7 to 6.4 points; synthetic Hindi produces the largest improvement, attributed to richer and more diverse Hindi text sources. The paper also claims that synthetic audio generated from in-domain text beats the same volume of general-domain synthetic audio, and that more distinct speakers in the synthetic set improve performance up to the fixed 100-hour audio cap.

Load-bearing premise

The load-bearing premise is that a text-to-speech model trained on 11 Indian languages but never on Bhojpuri can speak Bhojpuri correctly when prompted with a short voice sample; if the synthetic audio is Hindi-accented or mispronounced, the training data teaches the recognizer the wrong sounds.

Editorial extensions

If this is right

Adding synthetic Bhojpuri and synthetic Hindi audio to the real Bhojpuri baseline reduces WER on SRUTI from 33.8 to 29.1, moving the model closer to usable voice-based services in agriculture, health, governance, and finance.
Synthetic audio generated from domain-specific text outperforms the same volume of general-domain synthetic audio, so collecting or translating in-domain text should be a priority for dataset builders.
Increasing the number of distinct speakers in the synthetic data improves recognition at a fixed audio budget, though the effect may plateau once the total audio duration is capped.
Because only 25–30 seconds of seed audio is needed per speaker, the method bypasses the social and logistical obstacles that prevent long recording sessions with rural women.
Related-language synthetic audio yields the largest gains, indicating that target-language TTS coverage is not a prerequisite for useful synthetic augmentation.

Reading between the lines

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

As an extension beyond the paper, the same 25-second-seed pipeline should be testable on other underrepresented languages that have a close high-resource relative and some available text; a positive result would generalize the recipe beyond Bhojpuri.
As an extension beyond the paper, the largest gain appearing in the Health domain (6.4 points) suggests that domain-matched text, not language identity alone, drives most of the improvement; an experiment that varies only the text domain while holding speaker count and audio duration fixed would isolate that factor.
As an extension beyond the paper, because no listening or phoneme-level quality check is reported on the synthetic audio, a native-speaker rating study of synthetic Bhojpuri clips would determine whether the WER gain comes from genuine acoustic cloning or from topic overlap between synthetic training text and the benchmark.
As an extension beyond the paper, the voice-cloning step raises an implicit consent question, since seed clips were collected for transcription and synthesis was layered on top; scaling this approach would benefit from an explicit protocol stating how cloned audio may and may not be used.

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.

Desk Editor's Note

read the letter

Your reader's report is close to right, and the stress-test note lands. The benchmark is the real contribution: a 1-hour, 51-speaker SRUTI test set from rural Bhojpuri women across three UP districts, covering health, agriculture, governance, and finance, with careful community engagement and two-tier transcription. That fills a genuine gap, and the paper is honest about how hard the data is to collect. The ablations are also useful, especially speaker count at fixed duration and the in-domain text split. Releasing the code and data is the right move.

The soft spot is the headline. The abstract and intro say synthetic data alone gives the 4.7 WER gain, but Table 2 shows M1→M4 bundles 376 h of real Hindi with 200 h of synthetic speech. The isolated synthetic-Bhojpuri increment over the bilingual baseline is 1.8 (M2→M3), and the total synthetic increment is 4.2 (M2→M4). Neither is 4.7, and no run isolates synthetic Bhojpuri on top of M1. So the causal attribution in the abstract is overstated. That said, the ladder still supports the general claim that synthetic data helps: the gains are consistent across domains and prior cross-lingual TTS augmentation work points the same way. This is a framing error, not a fatal flaw.

Two smaller issues. First, the evaluation set is an hour of audio with no confidence intervals; some of the domain-level deltas (e.g., Finance 31.0→33.4 for M1→M2) are noise-level. A larger held-out set or error bars would sharpen the claims. Second, the load-bearing TTS model is an anonymized preprint, and no subjective or objective quality check is reported on the synthesized Bhojpuri. Since the paper explicitly leans on zero-shot transfer from Hindi, the absence of any validation of that transfer is a real gap, though again fixable.

This paper deserves a serious referee. The benchmark alone justifies it, and the augmentation recipe is practical for an underserved demographic. The referee should push for a corrected headline (report M2→M3 or M2→M4 as the synthetic effect, or run M1 with only synthetic Bhojpuri), confidence intervals on Table 2, and full details or release of the TTS. I'd take it to our reading group after those fixes, and I'd cite the SRUTI benchmark in any ASR work on Indic languages.

Referee Report

3 major / 6 minor

Summary. This paper introduces SRUTI, a one-hour benchmark of 444 utterances from 51 rural Bhojpuri women in Uttar Pradesh, spanning agriculture, health, governance, and finance. To address severe data scarcity, the authors collect 25--30 seconds of seed audio from about 100 additional rural women and use a multilingual prompt-based TTS model, trained on 11 Indian languages but not Bhojpuri, to synthesize 100 hours of Bhojpuri speech and 100 hours of Hindi speech. They train four Conformer-L models: monolingual Bhojpuri (M1), plus 376 hours of real Hindi (M2), plus synthetic Bhojpuri (M3), plus synthetic Hindi (M4), and report that the final system improves average WER on SRUTI from 33.8 to 29.1, a 4.7-point gain. The paper also reports ablations on in-domain text and the number of seed speakers.

Significance. The paper addresses an important and underserved population, and its data-collection methodology, including community engagement and ethical considerations, is a genuine contribution. The model ladder is transparent, and the public release of the benchmark and code is valuable. If the causal claims were correct, the result would be significant: it would suggest that very small per-speaker seed recordings, combined with a related-language TTS, can produce useful ASR training data for a language the TTS was never trained on. However, the headline 4.7 WER improvement is not attributable to synthetic data alone, and the absence of any quality verification of the synthetic audio leaves the central mechanism unvalidated. The paper's own ablations show that synthetic data helps, but the magnitude and source of the improvement need to be restated and supported by additional experiments.

major comments (3)

[Abstract; Section 6.1; Table 2] The abstract and Section 6.1 attribute the full 4.7 WER improvement to "this synthetic data," but the M1-to-M4 comparison in Table 2 changes three factors simultaneously: 376 hours of real Hindi, 100 hours of synthetic Bhojpuri, and 100 hours of synthetic Hindi. The isolated effect of synthetic Bhojpuri over the bilingual baseline is M2-to-M3, a 1.8 WER gain (33.3 to 31.5), while the combined synthetic effect over that baseline is M2-to-M4, a 4.2 WER gain (33.3 to 29.1). Neither equals 4.7. The headline should be rephrased, and an ablation is needed in which synthetic Bhojpuri is added to M1 alone, and in which synthetic Bhojpuri and synthetic Hindi are added separately to M2, so that the contribution of the seed-audio pipeline can be separated from the real-Hindi corpus. In addition, the statement that synthetic data "substantially boosts performance across all domains" is contradicted by the Finance column in Table 2, where M2-to-M3 degrades from 33.4 to 34.4.
[Section 4; reference [11]] The method relies on the assumption that a TTS model trained on 11 Indian languages, with no Bhojpuri training data, produces phonetically correct Bhojpuri when prompted with a 25--30 second seed clip. The paper provides no quality check of the synthetic audio: there is no native-speaker listening test, no pronunciation error analysis, no acoustic comparison between synthetic and real Bhojpuri, and no proxy such as ASR confidence on synthetic utterances. This is load-bearing because the synthetic audio is used as training labels; if the TTS produces Hindi-accented or mispronounced Bhojpuri, the WER gains may reflect domain overlap between synthetic audio and benchmark text rather than genuine acoustic generalization. Please add an objective or subjective quality evaluation, and either identify the TTS model or make it available once the anonymity constraint is lifted, as reference [11] is currently an anonymized preprint that cannot be inspected.
[Abstract; Section 1; Section 4; Table 2] The abstract and conclusion state that "evaluation of current ASR models on SRUTI shows poor performance," but the experiments do not evaluate any existing pretrained ASR system on SRUTI. The paper only reports results for the authors' own Conformer-L models trained from existing Bhojpuri and Hindi datasets. To support the benchmark's stated purpose, please report at least one or two public off-the-shelf ASR systems (for example, a Whisper variant or an IndicWav2Vec model) on SRUTI. Without such external baselines, the claim that current ASR systems are inadequate for this demographic, and the significance of the synthetic-data gains, cannot be assessed against the broader field.

minor comments (6)

[Table 1; Section 2] The dataset name is spelled inconsistently: "Shruitlipi" in Table 1 and "Shrutilipi" in the text and reference [5]. Please unify the spelling.
[Section 9] There is a typo in the section heading: "Acknoweldgments" should be "Acknowledgments."
[Section 6.2; Figure 3] The text says the model trained with in-domain text "significantly outperformed" the general-domain model, but Figure 3 shows no error bars, confidence intervals, or statistical tests. Given the small benchmark size, please add uncertainty estimates or state the difference descriptively without the word "significantly."
[Section 4] The synthetic Bhojpuri text was obtained using Google Translate with acknowledged translation errors, and it is unclear whether these transcripts were manually corrected before being used as ASR training labels. Please clarify this and, if no correction was made, discuss the potential effect of incorrect transcripts on the trained model.
[Abstract; Section 4] The abstract says "approximately 100 rural women" for seed audio, but the method also uses 25--30 second samples from 100 rural Hindi-speaking women in the same districts to generate synthetic Hindi. Please make clear in the abstract or introduction that the pipeline involves both Bhojpuri and Hindi seed speakers.
[Section 5; Table 1] The training data for M1 is described as 133.4 hours from four sources, but no per-source hour breakdown is given. Adding a breakdown would help readers understand the composition of the baseline and the relative scale of the synthetic additions.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular reduction found; the headline 4.7 WER gain is confounded by simultaneously added real Hindi data, but that is an attribution error rather than a definitional circularity.

full rationale

Walking the claimed derivation chain: SRUTI is a held-out benchmark from 51 speakers, while the seed clips for synthetic data come from 100 distinct speakers, and the additional 15.1 hours from benchmark speakers are explicitly excluded from training. No model is trained on the test set, and no reported number is defined as a function of the target result. The synthetic-data pipeline in Section 4 rests on the stated assumption that the prompt-based TTS model transfers from related languages such as Hindi; the paper does not claim this follows from reference [11] by construction, and [11] is cited only as the source of the model. The abstract's 4.7 WER claim is not the synthetic-only effect: Table 2 compares M1 with M4, which simultaneously adds 376 hours of real Hindi, 100 hours of synthetic Bhojpuri, and 100 hours of synthetic Hindi. The paper's own ladder yields M2-to-M3 = 1.8 WER and M2-to-M4 = 4.2 WER for the synthetic blocks, so the headline number is a causal-attribution overstatement rather than a definitional reduction. The only circularity-adjacent concern is reference [11], labeled 'Blinded for review', which is unverifiable and potentially self-authored; however, the paper does not invoke it as a proof of transfer, and the synthetic-data result is empirically testable through the reported ablations. Therefore no step reduces to its input by construction, and the paper is not circular under the stated criteria, though the attribution error and the unverifiable TTS reference are correctness and reproducibility risks. This yields a score of 1.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper contributes no theoretical entities; its new artifacts are the SRUTI benchmark (data, not a postulated entity) and a synthetic-data pipeline built on an unpublished TTS. The ledger's free parameters are design choices (100 speakers, 100 hours, 25-30 second seeds, in-domain split) that are not varied jointly and partly confound each other. The axioms are four domain assumptions plus one ad hoc choice (Google Translate text) that the paper itself flags but does not validate. The key unstated dependency is the TTS model, cited in anonymized form as reference [11].

free parameters (4)

Number of seed speakers = 100
Chosen by hand as the practical recruitment cap across three districts; the ablation in Section 6.3 varies 40/60/80/100 but keeps total duration at 100 hours, so 'more speakers' cannot be separated from 'less repetition per speaker'.
Synthetic data volume = 100 hours per language (Bhojpuri plus Hindi)
Cap on generated data; the speaker-count ablation holds duration fixed, so gains from more data and gains from more speakers are confounded.
Seed prompt duration = 25-30 seconds
Introduced in the abstract and Section 3.6 as the minimal viable voice sample; no ablation on prompt length is reported.
In-domain versus general text split = 32.5 h in-domain / 67.5 h general
Chosen split between GPT-4o plus Google Translate domain text and scraped web text; Section 6.2 compares all-general versus mixed but does not vary the ratio.

assumptions (5)

domain assumption Zero-shot cross-lingual TTS transfer from Hindi-family languages to Bhojpuri produces phonetically usable Bhojpuri speech
Stated in Section 4: 'We rely on the assumption that the model's exposure to related languages, such as Hindi, will allow for meaningful transfer.' If false, all 100 hours of synthetic Bhojpuri audio teach wrong acoustics.
domain assumption 376 hours of Hindi improve Bhojpuri ASR rather than degrade it
Used in model M2 (Section 5); empirically the average gain over M1 is only 0.5 WER and Finance and Agriculture worsen, so the benefit is weak and domain-dependent.
domain assumption One hour from 51 speakers in three districts is a representative evaluation of rural Bhojpuri women
Section 3.7; no validation against held-out districts or speakers, and the 15.1 hours of same-speaker data is excluded from training but not used to enlarge evaluation.
ad hoc to paper Google Translate Bhojpuri text, despite acknowledged translation errors, provides correct-enough transcripts for synthetic training audio
Section 4 dismisses translation errors because the goal is vocabulary coverage; in ASR the transcript must match the audio, and no check on translated text quality is reported.
domain assumption Conformer-L with CTC plus RNN-T trained on 133.4 hours for 100 epochs is a representative baseline for the claim
Section 5; there is no comparison to a stronger pretrained ASR or to longer training, so the measured gains are specific to this setup.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Recognizing Every Voice: Towards Inclusive ASR for Rural Bhojpuri Women." pith.science (2026). https://pith.science/paper/TEK7J5A6

@misc{pith2026250609653,
  author       = {Pith},
  title        = {Pith review of: Recognizing Every Voice: Towards Inclusive ASR for Rural Bhojpuri Women},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TEK7J5A6}},
  note         = {Machine review of arXiv:2506.09653}
}

read the original abstract

Digital inclusion remains a challenge for marginalized communities, especially rural women in low-resource language regions like Bhojpuri. Voice-based access to agricultural services, financial transactions, government schemes, and healthcare is vital for their empowerment, yet existing ASR systems for this group remain largely untested. To address this gap, we create SRUTI ,a benchmark consisting of rural Bhojpuri women speakers. Evaluation of current ASR models on SRUTI shows poor performance due to data scarcity, which is difficult to overcome due to social and cultural barriers that hinder large-scale data collection. To overcome this, we propose generating synthetic speech using just 25-30 seconds of audio per speaker from approximately 100 rural women. Augmenting existing datasets with this synthetic data achieves an improvement of 4.7 WER, providing a scalable, minimally intrusive solution to enhance ASR and promote digital inclusion in low-resource language.

Figures

Figures reproduced from arXiv: 2506.09653 by the authors.

**Figure 2.** Number of hours of data in target domains (left) and non-target domains (right) in SRUTI 4. Synthetic Data Generation Since it is relatively easy to collect short audio samples from multiple speakers, we explored whether this data could be used to generate additional speech samples in the same voices. To achieve this, we employed a multilingual prompt-based speech synthesis model trained on 11 Indian languages [11].… view at source ↗

**Figure 3.** (Left) Model’s performance with and without indomain synthetic data. (Right) Effect of number of speakers in the synthetic data on model performance. were generated using text from targeted domains (this version is reported as the default model 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

32 extracted references · 19 canonical work pages

[11]

Indicsuperb: A speech processing universal performance benchmark for indian languages,

T. Javed, K. S. Bhogale, A. Raman, A. Kunchukuttan, P. Kumar, and M. M. Khapra, “Indicsuperb: A speech processing universal performance benchmark for indian languages,” 2022. [Online]. Available: https://arxiv.org/abs/2208.11761

arXiv 2022
[1]

that which is heard

Introduction Automatic Speech Recognition (ASR) is a critical technology for digital inclusion, especially for populations with low literacy who rely on speech to access services. In India, where 68.91% of the population resides in rural areas, rural women face sig- nificant barriers to digital participation. 2 ASR can bridge this gap by enabling access t...

arXiv 2011
[2]

How- ever, as shown in Table 1, Bhojpuri remains severely underrep- resented in these efforts

Related Work Indic Datasets.Currently, there are several large-scale dataset collection efforts for Indian languages, such as, Vaani [1], SpringINX [4], IndicV oices [2], Kathbath [3] and ULCA [6], along with web-mined datasets such as Shruitlipi [5]. How- ever, as shown in Table 1, Bhojpuri remains severely underrep- resented in these efforts. This poses...

work page
[3]

SRUTI: A Benchmark for Rural Bhojpuri Women In this section, we describe our data collection effort. 3.1. Selection of Target domains To understand the specific needs of rural women, we began by engaging with local communities through village visits. Through these interactions, we identified four key domains that are critical to enhancing digital inclusio...

work page
[4]

To achieve this, we employed a multilingual prompt-based speech synthesis model trained on 11 Indian languages [11]

Synthetic Data Generation Since it is relatively easy to collect short audio samples from multiple speakers, we explored whether this data could be used to generate additional speech samples in the same voices. To achieve this, we employed a multilingual prompt-based speech synthesis model trained on 11 Indian languages [11]. It uses the English F5 model ...

work page
[5]

Experimental setup We trained four models using varying amounts of data, as de- scribed below. Each model utilized the Conformer-L architec- ture [17] with 130M parameters and a hybrid CTC + RNN-T loss function [18].The models were trained for 100 epochs with a peak learning rate of 5e-4, using the Noam learning rate sched- uler and Adam optimizer. Traini...

work page
[6]

Results and Analysis We now summarize the results of different models on SRUTI. 6.1. Performance on SRUTI Lower WERs.From Table 2, we first observe that the overall WER numbers are relatively high for this underserved demo- graphic and low-resource language (29+ WER even in our best set up). For context, the best reported WER for Hindi on the IndicV oices...

work page
[7]

Conclusion In this work, we address the lack of ASR data and benchmarks for rural Bhojpuri-speaking women. Through field visits and community engagement, we overcome social barriers and col- lect data in four key domains,health,agriculture,governance, andfinance, that are vital for digital inclusion. We discuss chal- lenges in data collection process requ...

work page

Show all 32 references

[8]

We grate- fully acknowledge Yotta Infrastructure for providing access to GPU resources that enabled large-scale model training

Acknoweldgments We would like to thank the Bill & Melinda Gates Foundation for their generous support in making this work possible. We grate- fully acknowledge Yotta Infrastructure for providing access to GPU resources that enabled large-scale model training. We would also lik...
[9]

Vaani: Capturing the language landscape for an inclu- sive digital india (phase 1),

V . Team, “Vaani: Capturing the language landscape for an inclu- sive digital india (phase 1),” https://vaani.iisc.ac.in/, 2025

2025
[10]

Indicvoices: Towards building an inclusive multilingual speech dataset for indian languages,

T. Javed, J. A. Nawale, E. I. George, S. Joshi, K. S. Bhogale, D. Mehendale, I. V . Sethi, A. Ananthanarayanan, H. Faquih, P. Palit, S. Ravishankar, S. Sukumaran, T. Panchagnula, S. Murali, K. S. Gandhi, A. R, M. K. M, C. V . Vaijayanthi, K. S. R. Karunganni, P. Kumar, and M. ...

2024 arXiv
[12]

Spring-inx: A multilingual indian language speech corpus by spring lab, iit madras,

N. R, M. S, J. F, A. Gangwar, M. N. J, S. Umesh, R. Sarab, A. K. Dubey, G. Divakaran, S. V . K, and S. V . Gangashetty, “Spring-inx: A multilingual indian language speech corpus by spring lab, iit madras,” 2023. [Online]. Available: https://arxiv.org/abs/2310.14654

2023 arXiv
[13]

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, S. Doddapaneni, A. Kunchukuttan, P. Kumar, and M. M. Khapra, “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
[14]

ULCA ASR Dataset Corpus,

Open-Speech-EkStep, “ULCA ASR Dataset Corpus,” https: //github.com/Open-Speech-EkStep/ULCA-asr-dataset-corpus, 2023, [Online; accessed 20-Feb-2025]

2023
[15]

A simple baseline for domain adaptation in end to end asr systems using synthetic data,

R. Joshi and A. Singh, “A simple baseline for domain adaptation in end to end asr systems using synthetic data,” inProceedings of The Fifth Workshop on e-Commerce and NLP (ECNLP 5). Association for Computational Linguistics, 2022, p. 244–249. [Online]. Available: http://dx.doi...

2022 doi
[16]

Text generation with speech synthesis for asr data augmentation,

Z. Huang, G. Keren, Z. Jiang, S. Jain, D. Goss-Grubbs, N. Cheng, F. Abtahi, D. Le, D. Zhang, A. D’Avirro, E. Campbell-Taylor, J. Salas, I.-E. Veliche, and X. Chen, “Text generation with speech synthesis for asr data augmentation,” 2023. [Online]. Available: https://arxiv.org/a...

2023 arXiv
[17]

Enhancing low-resource asr through versatile tts: Bridging the data gap,

G. Yang, F. Yu, Z. Ma, Z. Du, Z. Gao, S. Zhang, and X. Chen, “Enhancing low-resource asr through versatile tts: Bridging the data gap,” 2024. [Online]. Available: https://arxiv.org/abs/2410.16726

2024 arXiv
[18]

Asr data augmentation in low-resource settings using cross-lingual multi-speaker tts and cross-lingual voice conversion,

E. Casanova, C. Shulby, A. Korolev, A. C. Junior, A. da Silva Soares, S. Alu ´ısio, and M. A. Ponti, “Asr data augmentation in low-resource settings using cross-lingual multi-speaker tts and cross-lingual voice conversion,” 2023. [Online]. Available: https://arxiv.org/abs/2204.00618

2023 arXiv
[19]

Blinded for review,

Anonymous, “Blinded for review,”Blinded for Review, 2025

2025
[20]

F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,”arXiv preprint arXiv:2410.06885, 2024

2024 arXiv
[21]

Bloom library: Multimodal datasets in 300+ languages for a variety of downstream tasks,

C. Leong, J. Nemecek, J. Mansdorfer, A. Filighera, A. Owodunni, and D. Whitenack, “Bloom library: Multimodal datasets in 300+ languages for a variety of downstream tasks,” 2022. [Online]. Available: https://arxiv.org/abs/2210.14712

2022 arXiv
[22]

No language left behind: Scaling human-centered machine translation,

N. Team, M. R. Costa-juss `a, J. Cross, O. C ¸ elebi, M. Elbayad, K. Heafield, K. Heffernan, E. Kalbassi, J. Lam, D. Licht, J. Maillard, A. Sun, S. Wang, G. Wenzek, A. Youngblood, B. Akula, L. Barrault, G. M. Gonzalez, P. Hansanti, J. Hoffman, S. Jarrett, K. R. Sadagopan, D. R...

2022 arXiv
[23]

Zampieri, P

M. Zampieri, P. Nakov, N. Ljube ˇsi´c, J. Tiedemann, S. Malmasi, and A. Ali, Eds.,Proceedings of the Fifth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2018). Santa Fe, New Mexico, USA: Association for Computational Linguistics, Aug. 2018. [Online]. A...

2018
[24]

English-bhojpuri smt system: Insights from the karaka model,

A. K. Ojha, “English-bhojpuri smt system: Insights from the karaka model,”arXiv preprint arXiv:1905.02239, 2019

1905 arXiv
[25]

Conformer: Convolution-augmented transformer for speech recognition,

A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,”
[27]

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

V . Noroozi, S. Majumdar, A. Kumar, J. Balam, and B. Ginsburg, “Stateful conformer with cache-based inference for streaming automatic speech recognition,” inIEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2024, Seoul, Republic of Korea, April 1...

2024
[28]

Annotated speech corpus for low resource indian languages: Awadhi, bhojpuri, braj and magahi,

R. Kumar, S. Singh, S. Ratan, M. Raj, S. Sinha, b. lahiri, V . Se- shadri, K. Bali, and A. K. Ojha, “Annotated speech corpus for low resource indian languages: Awadhi, bhojpuri, braj and magahi,” inProceedings of Speech for Social Good Workshop, Interspeech 2022, 2022

2022
[29]

I. I. of Science, “Syspin,” https://syspin.iisc.ac.in/, accessed: Feb 18, 2025

2025
[30]

The iit bombay english-hindi parallel corpus,

A. Kunchukuttan, P. Mehta, and P. Bhattacharyya, “The iit bombay english-hindi parallel corpus,” 2018. [Online]. Available: https://arxiv.org/abs/1710.02855

2018 arXiv
[31]

Indictrans2: Towards high-quality and accessible machine translation models for all 22 scheduled indian languages,

J. Gala, P. A. Chitale, R. AK, V . Gumma, S. Doddapaneni, A. Kumar, J. Nawale, A. Sujatha, R. Puduppully, V . Raghavan, P. Kumar, M. M. Khapra, R. Dabre, and A. Kunchukuttan, “Indictrans2: Towards high-quality and accessible machine translation models for all 22 scheduled indi...

2023 arXiv
[32]

Rasa: Building expressive speech synthesis systems for indian languages in low-resource settings,

P. Srinivasa Varadhan, A. Sankar, G. Raju, and M. M. Khapra, “Rasa: Building expressive speech synthesis systems for indian languages in low-resource settings,” inInterspeech 2024, 2024, pp. 1830–1834

2024
[2020]

Available: https://arxiv.org/abs/2005.08100

[Online]. Available: https://arxiv.org/abs/2005.08100

2005 arXiv

Pith tools

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

[1] [11]

Indicsuperb: A speech processing universal performance benchmark for indian languages,

T. Javed, K. S. Bhogale, A. Raman, A. Kunchukuttan, P. Kumar, and M. M. Khapra, “Indicsuperb: A speech processing universal performance benchmark for indian languages,” 2022. [Online]. Available: https://arxiv.org/abs/2208.11761

arXiv 2022

[2] [1]

that which is heard

Introduction Automatic Speech Recognition (ASR) is a critical technology for digital inclusion, especially for populations with low literacy who rely on speech to access services. In India, where 68.91% of the population resides in rural areas, rural women face sig- nificant barriers to digital participation. 2 ASR can bridge this gap by enabling access t...

arXiv 2011

[3] [2]

How- ever, as shown in Table 1, Bhojpuri remains severely underrep- resented in these efforts

Related Work Indic Datasets.Currently, there are several large-scale dataset collection efforts for Indian languages, such as, Vaani [1], SpringINX [4], IndicV oices [2], Kathbath [3] and ULCA [6], along with web-mined datasets such as Shruitlipi [5]. How- ever, as shown in Table 1, Bhojpuri remains severely underrep- resented in these efforts. This poses...

work page

[4] [3]

SRUTI: A Benchmark for Rural Bhojpuri Women In this section, we describe our data collection effort. 3.1. Selection of Target domains To understand the specific needs of rural women, we began by engaging with local communities through village visits. Through these interactions, we identified four key domains that are critical to enhancing digital inclusio...

work page

[5] [4]

To achieve this, we employed a multilingual prompt-based speech synthesis model trained on 11 Indian languages [11]

Synthetic Data Generation Since it is relatively easy to collect short audio samples from multiple speakers, we explored whether this data could be used to generate additional speech samples in the same voices. To achieve this, we employed a multilingual prompt-based speech synthesis model trained on 11 Indian languages [11]. It uses the English F5 model ...

work page

[6] [5]

Experimental setup We trained four models using varying amounts of data, as de- scribed below. Each model utilized the Conformer-L architec- ture [17] with 130M parameters and a hybrid CTC + RNN-T loss function [18].The models were trained for 100 epochs with a peak learning rate of 5e-4, using the Noam learning rate sched- uler and Adam optimizer. Traini...

work page

[7] [6]

Results and Analysis We now summarize the results of different models on SRUTI. 6.1. Performance on SRUTI Lower WERs.From Table 2, we first observe that the overall WER numbers are relatively high for this underserved demo- graphic and low-resource language (29+ WER even in our best set up). For context, the best reported WER for Hindi on the IndicV oices...

work page

[8] [7]

Conclusion In this work, we address the lack of ASR data and benchmarks for rural Bhojpuri-speaking women. Through field visits and community engagement, we overcome social barriers and col- lect data in four key domains,health,agriculture,governance, andfinance, that are vital for digital inclusion. We discuss chal- lenges in data collection process requ...

work page