A Unified Denoising and Adaptation Framework for Self-Supervised Bengali Dialectal ASR

Imran; Swadhin Biswas; Tuhin Sheikh

REVIEW 5 major objections 6 minor 43 references

A Unified Denoising and Adaptation Framework for Self-Supervised Bengali Dialectal ASR

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Two-stage WavLM tuning sets the bar for noisy Bengali dialect ASR

desk verdict A useful recipe paper for low-resource dialectal ASR, but the empirical claims are held up by an undocumented data split and no error bars. read the letter →

arxiv 2509.00988 v1 pith:DBQTCBAJ submitted 2025-08-31 cs.SD eess.AS

Swadhin Biswas , Imran , Tuhin Sheikh This is my paper

classification cs.SDeess.AS

keywords BengaliASRdialectalspeechrecognitionself-supervisedlearningWavLMnoiserobustnessmulti-stagefine-tuninglow-resourcelanguagesCTC

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that Bengali dialectal speech recognition fails under real-world noise because standard fine-tuning ignores two coupled problems at once: dialect mismatch and acoustic corruption. It proposes a single recipe that addresses both: start from WavLM, a self-supervised model pre-trained with a masked speech denoising objective, and fine-tune in two stages—first on clean standard Bengali, then on dialectal audio with noise added on the fly. On Chittagonian and Sylheti test sets, the recipe lowers word error rates across every tested noise level, with the largest gains in the noisiest conditions. If these results hold, the same staged adaptation could be a practical pattern for other low-resource languages with many dialects.

What carries the argument

The machinery is the combination of WavLM's masked speech denoising and prediction pre-training objective—the model learns to predict clean pseudo-labels from masked, noisy input—with a two-stage fine-tuning schedule. Stage 1 adapts the full model to standard Bengali at a learning rate of 5e-5 for 10 epochs; Stage 2 continues at 1e-5 for 20 epochs on dialectal speech augmented online with noise at randomly chosen SNRs. A CTC head trained in both stages maps frame-level representations to character sequences. WavLM's denoising pre-training makes it a noise-aware starting point, and the staged schedule prevents the small dialectal dataset from destroying the general Bengali representation.

What would settle it

A fair check: repeat the two-stage recipe on OOD-Speech with a speaker-disjoint split reconstructed from speaker IDs, across at least three random seeds, and report the mean and spread of WER at 0 dB. If the margin between the two-stage WavLM and the single-stage WavLM baseline shrinks below the roughly 24-percentage-point gap reported at 0 dB on Chittagonian—or if test speakers appear among Stage 2 training speakers—the central claim would not survive.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that adapting a self-supervised speech model to a low-resource language with a single fine-tuning pass is the wrong tool when the language has dialects and the deployment environment is noisy. It reports that fine-tuning WavLM-Large first on about 94 hours of standard Bengali and then on dialectal Bengali with noise mixed in at 0–20 dB SNR produces the best word and character error rates on the Chittagonian and Sylheti splits of the OOD-Speech benchmark. The headline numbers are WER/CER of 10.1/5.2 on clean Chittagonian and 33.8/17.6 at 0 dB, compared with 58.0/31.0 for a single-stage WavLM fine-tune and 65.2/34.8 for Whisper-large-v2 at 0 dB. Ablations att

Load-bearing premise

The reported margins over the baselines rest on the assumption that the OOD-Speech dataset was split into non-overlapping train, development, and test partitions with no speaker overlap and no test audio seen during Stage 2 training, and that the single reported run is representative; the paper does not state split details or seed variance.

Editorial extensions

If this is right

The two-stage adaptation pattern—general-domain first, noisy dialect second—can be reused directly for other low-resource languages with high dialectal variation.
WavLM's denoising pre-training objective provides measurable robustness that fine-tuning alone cannot supply, as shown by the wav2vec 2.0 comparison under the same recipe.
Noise augmentation during dialectal specialization is the largest single component of the gain: removing it raises noisy Sylheti WER from 24.8% to 41.2% in the ablation.
Large multilingual models such as Whisper-large-v2 are not automatically the best starting point for this task; a specialized SSL model with staged tuning outperforms them, especially at low SNR.

Reading between the lines

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

The paper does not test whether one Stage-2 run can serve several dialects at once; a shared multidi-dialect specialization may be more parameter-efficient than per-dialect runs, but that remains untested.
Because the framework is purely acoustic, adding a Bengali language model at decoding time could compound the reported gains; the paper lists this as future work without quantifying it.
The reported WER gap depends on a clean train/test split of OOD-Speech with no speaker overlap; a strict speaker-disjoint split with multiple seeds would tell whether the margin over single-stage WavLM is robust.

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

Colleague,

Short version: this is a sensible application paper that combines WavLM, noise augmentation, and two-stage fine-tuning for Bengali dialectal ASR. The recipe is not conceptually new—each ingredient is known—but the specific combination is new and the benchmark numbers against wav2vec 2.0 and Whisper are genuinely useful for anyone working on low-resource ASR. The paper deserves a serious referee, but the current draft is not ready as-is.

What is actually good: the central idea is plausible and the ablations are the right ones. Removing Stage 1 hurts, removing noise augmentation hurts much more, and comparing WavLM to wav2vec2 with the same fine-tuning protocol isolates the backbone contribution. The qualitative examples in Table 4 are a nice touch—they show the model is not just noise-robust but that it keeps dialectal lexicon that the baseline normalizes away. The writing is clear and the related work is wide enough, though heavy on web citations.

Now the soft spots, in rough order of severity.

First, the stress-test concern about the OOD-Speech split is real and it is load-bearing. Table 1 says the dataset is ~20 hours, ~100 speakers, and is partitioned into train/dev/test, but the paper never says whether the split is speaker-disjoint. With that few speakers, an utterance-level random split would very likely put the same speaker in both training and test. The model can memorize speaker-specific pronunciations, and that alone could explain a large chunk of the WER gap—especially the 33.8% vs 58.0% at 0 dB. The authors need to state the split explicitly: speaker-disjoint or not, how many speakers per partition, and hours per partition. This is not a minor omission; it is the difference between a SOTA claim and a rumor.

Second, there are no error bars, no seeds, no repeated runs, and no significance testing. Given the margins in Table 2 are often large (e.g., 10.1 vs 12.8 clean), the result would probably survive replication, but without variance estimates I cannot tell if the smaller advantages (e.g., 11.3 vs 14.2 at 20 dB for Baseline 2) are real or noise. This is a standard expectation for an empirical paper and should be easy to fix.

Third, the draft is incomplete. The main figure is a placeholder, Algorithm 1 is referenced but Algorithm 2 appears, and the Whisper baseline is described as having two settings but only one row of numbers appears. Also, claims of 'multiple' dialects and 'comprehensive' benchmarking overstate the evidence: only Chittagonian and Sylheti are actually evaluated, and the limitations section admits this. The authors know this—the limitation is stated—but the framing in the abstract and introduction oversells it.

My take: the paper is a credible engineering contribution that could become a solid benchmark reference after the split question is answered and the experimental reporting is tightened. I would not desk-reject it. I would send it to review but with a clear request to the authors to document the split, add variance estimates, and fix the missing artifacts.

For a reading group: maybe, if someone is working on low-resource dialectal ASR. For citing in my own work: not yet, until the split is clarified.

—

Recommendation: accept for peer review, expect major revision.

Referee Report

5 major / 6 minor

Summary. The paper proposes a unified two-stage fine-tuning framework for Bengali dialectal ASR. The system starts from WavLM-Large, a self-supervised model with a masked speech denoising pre-training objective, and applies two sequential fine-tuning stages: Stage 1 adapts the model to standard Bengali using MCV and OpenSLR data, and Stage 2 specializes it on dialectal speech from OOD-Speech with online NOISEX-92 noise augmentation at SNRs between 0 and 20 dB. The model is evaluated on Chittagonian and Sylheti test sets under clean and four noisy conditions, reporting WER/CER. Comparisons include a single-stage WavLM baseline, a wav2vec2-XLSR baseline with the same two-stage strategy, and Whisper-large-v2. The paper reports consistent improvements, with the largest gains at low SNR, and includes ablations showing the contribution of Stage 1 and of noise augmentation.

Significance. If the empirical results hold, the paper makes a useful practical contribution: it provides a concrete recipe for adapting a noise-robust SSL backbone to a low-resource dialectal language and demonstrates that explicit noise augmentation during dialectal specialization matters. The choice of WavLM is well motivated, and the ablations in Table 3 give independent evidence that both the two-stage scheme and noise augmentation are important. However, the significance is currently bounded by three problems: the evaluation covers only two dialects despite broader claims; the OOD-Speech train/dev/test split is undocumented; and no experimental variability (seeds, error bars, significance tests) is reported. The Whisper baseline is described in two settings but only one row of numbers appears, making the headline comparison ambiguous. These gaps are fixable and do not undermine the plausibility of the central mechanism, but they must be addressed before the SOTA claim can be considered reliable.

major comments (5)

[§4.1, Table 1] The OOD-Speech split is completely undocumented. Section 4.1 says only that the dataset 'is partitioned into training, development, and test sets', and Table 1 lists 'Train (Stage 2), Dev, Test' but gives no per-split hours, speaker counts, or partition scheme. With roughly 20 hours and ~100 speakers total, an utterance-level random split could easily place the same speaker in both training and test. This is load-bearing: speaker overlap between Stage 2 training and evaluation could inflate the WER improvements in Table 2 through speaker memorization, especially for dialectal pronunciations. Please report exact split sizes, speaker counts, and explicitly state whether the splits are speaker-disjoint. If not, re-run with a speaker-disjoint partition.
[§5.1, Table 2] No error bars, seeds, or repeated runs are reported. All WER/CER numbers in Table 2 appear to come from a single run. On a small dataset, the reported margins (e.g., 33.8% vs 49.5% at 0 dB on Dialect A) could be within run-to-run variance. Please report mean and standard deviation over at least 3-5 seeds, or provide significance testing, and state the number of test utterances per condition. This is necessary to support the claim of consistent superiority.
[§4.4, Table 2] The Whisper baseline is ambiguous. Section 4.4 states that Whisper-large-v2 is evaluated in two settings—zero-shot and fine-tuned—but Table 2 shows only a single row labeled 'Baseline 3 (Whisper-large-v2)' with no indication of which setting is reported. If these are zero-shot numbers, comparing them to a fine-tuned specialized model is apples-to-oranges; if they are fine-tuned, the fine-tuning protocol (data, learning rate, epochs) is missing. Please clarify and, if both settings were run, report both rows.
[Abstract, §4.1, Table 2] The abstract and introduction claim the framework is evaluated on 'a comprehensive benchmark comprising multiple Bengali dialects', but the results in Table 2 cover only two dialects (Chittagonian and Sylheti). Table 1 lists 'Chittagonian, Sylheti, etc.' without specifying which other dialects exist in OOD-Speech or why they are excluded. The stated SOTA claim is therefore only supported for two dialects. Please either restrict the claims to the evaluated dialects or add results for additional dialects; otherwise the 'comprehensive' wording overstates the evidence.
[§4.4, Baseline 1] The single-stage Baseline 1 is described as fine-tuned on a 'combined dataset of standard and dialectal Bengali speech', but the exact composition is not given. It is unclear whether this baseline uses the same OOD-Speech training split as Stage 2, whether it sees the same number of epochs, and whether any noise augmentation is used (the text says no). These details are necessary to ensure that the multi-stage vs. single-stage comparison is not confounded by data amount or training schedule. Please specify the exact data and hyperparameters used for Baseline 1.

minor comments (6)

[§3.5] Section 3.5 says the method is 'formalized in Algorithm 1', but no Algorithm 1 is present in the manuscript; a stray heading 'Algorithm 2:' appears instead. Either include the pseudocode or remove the reference.
[§3.1.2] The formula for attention with gated relative position bias is ambiguous: 'softmax(QK^T + b_rel / sqrt(d_k))V' is missing parentheses. It should be 'softmax((QK^T + b_rel)/sqrt(d_k))V' or equivalent.
[§3.4] The CTC example says paths are mapped to the target 'helo'; presumably this should be 'hello' to match the example path. Please correct.
[Figure 1] Figure 1 is described in the text but is not actually included; the caption placeholder says 'A line graph would be inserted here'. Please include the figure.
[References] The reference list is dominated by web-page access links with access date July 24, 2025, which predates the arXiv submission date of the manuscript. Many entries are duplicated (e.g., items 3 and 4; items 8 and 11). Please replace these with standard venue citations and deduplicate.
[Abstract and §6.3] The limitations section honestly acknowledges that only a limited subset of dialects was evaluated and that only NOISEX-92 noise was used, but the abstract and introduction do not reflect these caveats. Please align the claims with the scope of the experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical training study; its claims are tested against held-out data and ablations, with no equation-level or self-citation circularity.

full rationale

The paper's central claim is an empirical performance comparison (Table 2) of a two-stage fine-tuned WavLM-Large against baselines on OOD-Speech test sets. This is not a derivation: the framework is evaluated, not derived from its own inputs. The choice of WavLM is motivated by its denoising pre-training objective, but the claimed benefit is tested empirically by comparing WavLM+Our FT against Wav2Vec2+Our FT (Baseline 2), which is an independent check rather than a consequence of definition. The two-stage adaptation and noise augmentation are ablated in Table 3; removing either component degrades WER, providing further independent evidence. No fitted parameter is renamed as a prediction, and no load-bearing argument reduces to a self-citation: WavLM, wav2vec 2.0, and Whisper are external published models, and the authors do not rely on their own prior work to justify the framework. The Limitations section (Sec. 6.3) honestly acknowledges restricted dialect and noise coverage, which is a scope/correctness caveat, not circularity. The only substantive concern—undisclosed OOD-Speech train/dev/test partition details and possible speaker overlap—is an experimental-validity risk about the trustworthiness of the empirical results, not circularity of the kind that would make the paper's result equivalent to its input by construction. Therefore the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper contributes a training recipe, not a new theory. Its central claim rests on two classes of assumptions: (1) the validity of cross-lingual transfer from English-pretrained WavLM to Bengali, and (2) the integrity and representativeness of the OOD-Speech benchmark and NOISEX-92 noise simulation. The free parameters are the hand-chosen hyperparameters of the recipe, which were not optimized or analyzed for sensitivity.

free parameters (6)

Stage 1 learning rate = 5e-5
Chosen by hand for general-domain adaptation; no search described (Section 4.3).
Stage 2 learning rate = 1e-5
Chosen by hand for dialectal specialization; no search described (Section 4.3).
Epochs per stage = 10 (Stage 1), 20 (Stage 2)
Handpicked; no early stopping or validation-based selection described (Section 4.3).
SNR augmentation set = {0, 5, 10, 20} dB
Randomly sampled for online augmentation; no optimization over SNR levels (Section 4.2).
Noise corpus = NOISEX-92
Chosen as the standard noise corpus; other noise types not tested (Section 4.1).
Model size = WavLM-Large
Largest available variant; no comparison with Base or Medium (Section 4.3).

assumptions (5)

domain assumption WavLM pre-trained on 94k hours of English transfers to Bengali via fine-tuning
Relies on cross-lingual transfer of SSL speech representations, a standard but unverified assumption in low-resource ASR (Sections 3.2, 4.1).
domain assumption OOD-Speech train/dev/test splits are non-overlapping and representative
No split statistics or speaker overlap checks are provided; the entire evaluation hinges on this (Section 4.1, Table 1).
domain assumption NOISEX-92 noise mixed at SNRs 0-20 dB is a valid proxy for real-world noise
Used for augmentation and evaluation; real environments may have non-stationary noise not covered (Sections 4.2, 6.3).
domain assumption CTC decoding with the Bengali character set is the right output representation
No language model is used; the paper relies on CTC character output as the full ASR pipeline (Section 3.4).
standard math Standard math for softmax attention, cross-entropy, and CTC forward-backward
Invoked throughout Section 3 without proof; these are established results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Unified Denoising and Adaptation Framework for Self-Supervised Bengali Dialectal ASR." pith.science (2026). https://pith.science/paper/DBQTCBAJ

@misc{pith2026250900988,
  author       = {Pith},
  title        = {Pith review of: A Unified Denoising and Adaptation Framework for Self-Supervised Bengali Dialectal ASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DBQTCBAJ}},
  note         = {Machine review of arXiv:2509.00988}
}

read the original abstract

Automatic Speech Recognition (ASR) for Bengali, the world's fifth most spoken language, remains a significant challenge, critically hindering technological accessibility for its over 270 million speakers. This challenge is compounded by two persistent and intertwined factors: the language's vast dialectal diversity and the prevalence of acoustic noise in real-world environments. While state-of-the-art self-supervised learning (SSL) models have advanced ASR for low-resource languages, they often lack explicit mechanisms to handle environmental noise during pre-training or specialized adaptation strategies for the complex phonetic and lexical variations across Bengali dialects. This paper introduces a novel, unified framework designed to address these dual challenges simultaneously. Our approach is founded on the WavLM model, which is uniquely pre-trained with a masked speech denoising objective, making it inherently robust to acoustic distortions. We propose a specialized multi-stage fine-tuning strategy that first adapts the model to general-domain standard Bengali to establish a strong linguistic foundation and subsequently specializes it for noise-robust dialectal recognition through targeted data augmentation. The framework is rigorously evaluated on a comprehensive benchmark comprising multiple Bengali dialects under a wide range of simulated noisy conditions, from clean audio to low Signal-to-Noise Ratio (SNR) levels. Experimental results demonstrate that the proposed framework significantly outperforms strong baselines, including standard fine-tuned wav2vec 2.0 and the large-scale multilingual Whisper model. This work establishes a new state-of-the-art for this task and provides a scalable, effective blueprint for developing practical ASR systems for other low-resource, high-variation languages globally.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

[1]

(PDF) BanglaDialecto: An End -to-End AI-Powered Regional Speech Standardization, accessed on July 24, 2025, https://www.researchgate.net/publication/385920290_BanglaDialecto_An_End -to- End_AI-Powered_Regional_Speech_Standardization

work page arXiv 2025
[2]

BAAD: A multipurpose dataset for automatic Bangla offensive speech recognition - PMC, accessed on July 24, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC10070523/

work page 2025
[4]

arxiv.org, accessed on July 24, 2025, https://arxiv.org/html/2507.01931v1

arXiv 2025
[5]

Enabling ASR for Low -Resource Languages: A Comprehensive Dataset Creation Approach, accessed on July 24, 2025, https://www.researchgate.net/publication/381153027_Enabling_ASR_for_Low - Resource_Languages_A_Comprehensive_Dataset_Creation_Approach

work page arXiv 2025
[6]

mushrafi88/asr_bangla: Automatic Speech Recognition system using Wav2Vec - XLSR for Bangla - GitHub, accessed on July 24, 2025, https://github.com/mushrafi88/asr_bangla

work page 2025
[7]

(PDF) Automatic Speech Recognition for Biomedical Data in Bengali Language, accessed on July 24, 2025, https://www.researchgate.net/publication/381579766_Automatic_Speech_Recognit ion_for_Biomedical_Data_in_Bengali_Language

work page arXiv 2025
[8]

Challenges and Opportunities of Speech Recognition for Bengali Language - ResearchGate, accessed on July 24, 2025, https://www.researchgate.net/publication/354887634_Challenges_and_Opportuniti es_of_Speech_Recognition_for_Bengali_Language

work page arXiv 2025
[9]

An Overview of Bengali Speech Recognition: Methods, Challenges, and Future Direction, accessed on July 24, 2025, https://scispace.com/papers/an -overview -of- bengali-speech -recognition -methods -challenges -14mpau2a

work page 2025

Show all 43 references

[10]

OOD-Speech: A Large Bengali Speech Recognition Dataset for Out -of-Distribution Benchmarking | Papers With Code, accessed on July 24, 2025, https://paperswithcode.com/paper/ood -speech -a-large-bengali-speech -recognition

2025
[11]

Challenges and Opportunities of Speech Recognition for Bengali Language - arXiv, accessed on July 24, 2025, https://arxiv.org/pdf/2109.13217

2025 arXiv
[12]

- arXiv, accessed on July 24, 2025, https://arxiv.org/pdf/2110.13900

WavLM: Large-Scale Self-Supervised Pre -Training for Full ... - arXiv, accessed on July 24, 2025, https://arxiv.org/pdf/2110.13900

2025
[13]

A Survey of Multilingual Models for Automatic Speech Recognition | Papers With Code, accessed on July 24, 2025, https://paperswithcode.com/paper/a -survey -of- multilingual -models-for-automatic

2025
[15]

Text-guided HuBERT: Self -Supervised Speech Pre -training via Generative Adversarial Networks - arXiv, accessed on July 24, 2025, https://arxiv.org/html/2402.15725v1

2025 arXiv
[16]

WavLM: Large-Scale Self-Supervised Pre -Training for Full Stack Speech Processing | Request PDF - ResearchGate, accessed on July 24, 2025, https://www.researchgate.net/publication/355664615_WavLM_Large -Scale_Self - Supervised_Pre -Training_for_Full_Stack_Speech_Processing

2025
[17]

wav2vec 2.0: A Framework for Self -Supervised Learning of Speech Representations - NIPS, accessed on July 24, 2025, https://proceedings.neurips.cc/paper/2020/file/92d1e1eb1cd6f9fba3227870bb6d7f0 7-Paper.pdf

2025
[18]

wav2vec 2.0: A Framework for Self -Supervised Learning of Speech Representations, accessed on July 24, 2025, https://www.researchgate.net/publication/342377803_wav2vec_20_A_Framework_ for_Self -Supervised_Learning_of_Speech_Representations

2025
[19]

wav2vec 2.0: A Framework for Self -Supervised Learning of Speech ..., accessed on July 24, 2025, https://arxiv.org/pdf/2006.11477

2025 arXiv
[20]

wav2vec -S: Adapting Pre-trained Speech Models for Streaming - ACL Anthology, accessed on July 24, 2025, https://aclanthology.org/2024.findings -acl.681.pdf

2025
[21]

Multi-resolution HuBERT: Multi -resolution Speech Self -Supervised ..., accessed on July 24, 2025, https://arxiv.org/pdf/2310.02720

2025 arXiv
[22]

AV-HuBERT with Multi -Resolution Attention - Stanford University, accessed on July 24, 2025, https://web.stanford.edu/class/archive/cs/cs224n/cs224n.1234/final - reports/final -report -169407002.pdf

2025
[23]

[2309.13860] Fast -HuBERT: An Efficient Training Framework for Self -Supervised Speech Representation Learning - arXiv, accessed on July 24, 2025, https://arxiv.org/abs/2309.13860

2025 arXiv
[24]

WavLM: Large-Scale Self-Supervised Pre -Training for Full Stack Speech Processing, accessed on July 24, 2025, https://www.catalyzex.com/paper/wavlm - large-scale-self-supervised -pre

2025
[25]

Attention Is All You Need - Wikipedia, accessed on July 24, 2025, https://en.wikipedia.org/wiki/Attention_Is_All_You_Need

2025
[26]

Attention is All you Need - NIPS, accessed on July 24, 2025, https://papers.neurips.cc/paper/7181 -attention -is-all-you-need.pdf

2025
[27]

Fine-Tuning ASR models for Very Low -Resource Languages: A Study on Mvskoke - ACL Anthology, accessed on July 24, 2025, https://aclanthology.org/2024.acl -srw.16/

2025
[28]

End-to-End Low-Resource Automatic Speech Recognition for Second Language Learners - Aaltodoc, accessed on July 24, 2025, https://aaltodoc.aalto.fi/bitstreams/b1d48859 -9778-4f63-b3d0- 98a83f50e907/download

2025
[29]

End-to-End Bengali Speech Recognition using DeepSpeech - ResearchGate, accessed on July 24, 2025, https://www.researchgate.net/publication/337940431_End -to- End_Bengali_Speech_Recognition_using_DeepSpeech

2025
[30]

End-to-End Bengali Speech Recognition | Request PDF - ResearchGate, accessed on July 24, 2025, https://www.researchgate.net/publication/344335433_End -to- End_Bengali_Speech_Recognition

2025
[31]

Recent Advancement in Speech Recognition for Bangla: A Survey, accessed on July 24, 2025, https://thesai.org/Publications/ViewPaper?Volume=12&Issue=3&Code=IJACSA& SerialNo=65

2025
[32]

Bengali Speech Recognition: An Overview (2022) | 3 Citations - SciSpace, accessed on July 24, 2025, https://scispace.com/papers/bengali -speech-recognition - an-overview -yu1kzme2

2022
[33]

Recent Advancement in Speech Recognition for Bangla: A Survey - ResearchGate, accessed on July 24, 2025, https://www.researchgate.net/publication/350548169_Recent_Advancement_in_Sp eech_Recognition_for_Bangla_A_Survey

2025
[34]

Recent Advancement in Speech Recognition for Bangla: A Survey, accessed on July 24, 2025, https://thesai.org/Downloads/Volume12No3/Paper_65 - Recent_Advancement_in_Speech_Recognition.pdf

2025
[35]

[2404.00656] WavLLM: Towards Robust and Adaptive Speech Large Language Model, accessed on July 24, 2025, https://arxiv.org/abs/2404.00656

2025 arXiv
[36]

Signal to Noise Ratio Formula - GeeksforGeeks, accessed on July 24, 2025, https://www.geeksforgeeks.org/physics/signal -to-noise-ratio-formula/

2025
[37]

Normalization through Fine -tuning: Understanding Wav2vec 2.0 Embeddings for Phonetic Analysis - arXiv, accessed on July 24, 2025, https://arxiv.org/html/2503.04814v1

2025 arXiv
[38]

Connectionist temporal classification - Wikipedia, accessed on July 24, 2025, https://en.wikipedia.org/wiki/Connectionist_temporal_classification

2025
[39]

Connectionist Temporal Classification - GeeksforGeeks, accessed on July 24, 2025, https://www.geeksforgeeks.org/nlp/connectionist -temporal -classification/

2025
[40]

Connectionist Temporal Classification(CTC) | by Abhishek Kumar Pandey - Medium, accessed on July 24, 2025, https://medium.com/@akp83540/connectionist -temporal -classification -ctc- 722bbb767e62

2025
[41]

Breaking down the CTC Loss - Sewade Ogun's Website, accessed on July 24, 2025, https://ogunlao.github.io/blog/2020/07/17/breaking -down-ctc-loss.html

2025
[42]

AdamW Explained | Papers With Code, accessed on July 24, 2025, https://paperswithcode.com/method/adamw

2025
[43]

AdamW - Cornell University Computational Optimization Open Textbook, accessed on July 24, 2025, https://optimization.cbe.cornell.edu/index.php?title=AdamW

2025
[44]

Low-resource track - IWSLT, accessed on July 24, 2025, https://iwslt.org/2025/low -resource

2025
[45]

EMBARGOED until 1 August 2025., accessed on July 24, 2025, https://mro.massey.ac.nz/items/11a54abf -87c6-45d3- b3b1-ffbee4812aa6

End-to-end automatic speech recognition for low -resource languages : a thesis submitted in partial fulfillment for the degree of Doctor of Philosophy in Computer Science at the School of Mathematical and Computational Sciences, Massey University, Auckland, New Zealand. EMBARG...

2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

(PDF) BanglaDialecto: An End -to-End AI-Powered Regional Speech Standardization, accessed on July 24, 2025, https://www.researchgate.net/publication/385920290_BanglaDialecto_An_End -to- End_AI-Powered_Regional_Speech_Standardization

work page arXiv 2025

[2] [2]

BAAD: A multipurpose dataset for automatic Bangla offensive speech recognition - PMC, accessed on July 24, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC10070523/

work page 2025

[3] [4]

arxiv.org, accessed on July 24, 2025, https://arxiv.org/html/2507.01931v1

arXiv 2025

[4] [5]

Enabling ASR for Low -Resource Languages: A Comprehensive Dataset Creation Approach, accessed on July 24, 2025, https://www.researchgate.net/publication/381153027_Enabling_ASR_for_Low - Resource_Languages_A_Comprehensive_Dataset_Creation_Approach

work page arXiv 2025

[5] [6]

mushrafi88/asr_bangla: Automatic Speech Recognition system using Wav2Vec - XLSR for Bangla - GitHub, accessed on July 24, 2025, https://github.com/mushrafi88/asr_bangla

work page 2025

[6] [7]

(PDF) Automatic Speech Recognition for Biomedical Data in Bengali Language, accessed on July 24, 2025, https://www.researchgate.net/publication/381579766_Automatic_Speech_Recognit ion_for_Biomedical_Data_in_Bengali_Language

work page arXiv 2025

[7] [8]

Challenges and Opportunities of Speech Recognition for Bengali Language - ResearchGate, accessed on July 24, 2025, https://www.researchgate.net/publication/354887634_Challenges_and_Opportuniti es_of_Speech_Recognition_for_Bengali_Language

work page arXiv 2025

[8] [9]

An Overview of Bengali Speech Recognition: Methods, Challenges, and Future Direction, accessed on July 24, 2025, https://scispace.com/papers/an -overview -of- bengali-speech -recognition -methods -challenges -14mpau2a

work page 2025