Pith. sign in

REVIEW 3 major objections 8 minor 30 references

Advancing STT for Low-Resource Real-World Speech

T0 review · 3 major / 8 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 Fine-tuning Whisper on a new 303-hour corpus of spontaneous Swiss German broadcast speech cuts word error rate by 19–33% and raises BLEU by up to 40% over zero-shot, reaching 17.1% WER and 74.8 BLEU.

desk verdict Useful new broadcast corpus for Swiss German STT, but the Whisper-derived ground truth makes the reported gains partly a measure of convergence to the labeling pipeline. read the letter →

arxiv 2506.08836 v1 pith:TOTUV6A3 submitted 2025-06-10 cs.CL cs.HC

Flavio D'Intino , Hans-Peter Hutter This is my paper
classification cs.CLcs.HC
keywords SwissGermanspeech-to-textlow-resourcelanguagesWhisperfine-tuningspontaneousspeechbroadcastcorpusworderrorrateBLEUscore
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 is trying to establish that the main barrier to speech-to-text for Swiss German is not model architecture but training data: a 303-hour corpus of spontaneous, long-form broadcast speech lets Whisper models beat both their zero-shot performance and models trained on controlled sentence-level corpora. The gains are measured with word error rate (WER), the share of words that must be fixed to match the reference, and BLEU, an n-gram overlap score. Fine-tuning on the new SRB-300 corpus cuts WER by 19–33% and raises BLEU by 8–40% relative to zero-shot Whisper, with the best model at 17.1% WER and 74.8 BLEU. The same test set drops the strongest prior Swiss German system from its reported 14.0 WER and 74.7 BLEU to 44.4 WER and 37.5 BLEU, showing that sentence-level corpora do not transfer to real-world speech. If the claim holds, the recipe—collect realistic long-form audio and fine-tune a large multilingual model—should transfer to other low-resource languages.

What carries the argument

The central object is the SRB-300 corpus: 303 hours of audio from 39 regional Swiss radio and TV stations, cut into samples of 3.4 to 30 seconds with a mean of 25.4 seconds, and split into training, validation, and test sets of 76%, 10%, and 14%, with six stations reserved for testing only. The samples keep the original chronological order of the broadcasts, so consecutive samples can be stitched into long audio for evaluation. The corpus does the work of supplying the variability missing from earlier corpora—spontaneous disfluencies, multiple speakers, background noise, and all major Swiss dialect regions—while matching Whisper's 30-second input window without zero-padding. The fine-tuning procedure is the second half of the machinery: a standard Whisper adaptation recipe using the German language tag, AdamW with learning rate $5\times10^{-6}$ for small and medium models and $1\times10^{-6}$ for large models, gradient accumulation over 32 steps, and 4–12 epochs.

What would settle it

Take a held-out set of Swiss German broadcast audio that no Whisper model has touched, have translators produce Standard German references without any automatic pre-transcription, and compare the best fine-tuned model against zero-shot Whisper; if the WER gap collapses, the reported gains partly measure self-agreement with Whisper-derived labels.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that realistic long-form audio, not more read-aloud sentences, is what unlocks Swiss German speech-to-text. The SRB-300 dataset provides that audio: 303 hours from 39 radio and TV stations, sampled into 30-second chunks that keep whole utterances, and split so that six stations appear only in the test set. Fine-tuning Whisper models of every size on this corpus improves both metrics over zero-shot, by 19% to 33% in WER and 8% to 40% in BLEU; the large-v3 model reaches 17.1% WER and 74.8 BLEU. By contrast, fine-tuning Whisper medium on the sentence-level STT4SG-350 corpus makes it worse on realistic audio than the zero-shot model (31.4 vs 27.2 WER), and adding STT4SG-350 to SRB-300 yields little further gain. The authors conclude that realistic data is the active ingredient and that their dataset addresses a gap that existing Swiss German corpora leave open.

Load-bearing premise

The load-bearing premise is that the manually corrected Whisper pre-transcriptions give an unbiased Standard German gold standard, so the WER and BLEU gains measure true quality rather than the models learning to agree with Whisper's own output; a secondary fragility is that test and training sets can share speakers because no unique speaker IDs exist.

Editorial extensions

If this is right

  • Every Whisper size tested improves with SRB-300 fine-tuning: WER falls 19–33% and BLEU rises 8–40% relative to zero-shot, including on broadcast stations with no training data.
  • The best adapted model, large-v3, reaches 17.1% WER and 74.8 BLEU on the realistic test set, matching the level the strongest prior systems report only on their own clean sentence-level test sets.
  • Fine-tuning on the sentence-level STT4SG-350 corpus alone makes Whisper worse on realistic audio (medium WER rises from 27.2 to 31.4), and adding it to SRB-300 adds little, so realistic data is the active ingredient.
  • Long-audio use works: concatenating consecutive test samples into broadcasts of up to one hour gives performance similar to the 30-second test set, so the gains are not an artifact of short segments.
  • The turbo variant offers a practical speed-quality trade-off: after fine-tuning it reaches 18.7 WER and 71.8 BLEU at roughly 40% of the large-v3 inference time.

Reading between the lines

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

  • A natural test the authors do not run is whether the gain is partly self-distillation: since the gold transcripts began as Whisper large-v3 output, a reference set produced without any automatic pre-transcription would show how much of the improvement is real audio-to-text learning rather than agreement with Whisper's own output distribution.
  • Because the corpus cannot be publicly released and lacks unique speaker IDs, independent confirmation would require building a similar broadcast corpus; a small professionally transcribed held-out set would also settle the speaker-overlap concern.
  • The same recipe should transfer to other low-resource languages with a regional broadcast archive and no written standard; a testable prediction is that gains will track how much spontaneous, multi-speaker, noisy speech the new corpus contains relative to read speech.
  • The 2:1 male-to-female ratio and the predominance of adult presenters mean the model may underperform on children, elderly speakers, and rare dialects; checking per-demographic WER on a targeted test set would reveal this boundary, which the paper itself flags in its risk discussion.
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 / 8 minor

Summary. The paper introduces SRB-300, a 303-hour corpus of Swiss German broadcast audio collected from 39 radio and TV stations, consisting of 28-second speech chunks with metadata and Standard German reference transcripts. The authors fine-tune several OpenAI Whisper models (small, medium, large-v2, large-v3, large-v3-turbo) and XLS-R 1B on SRB-300, and compare them to zero-shot Whisper and to models fine-tuned on the sentence-level STT4SG-350 corpus. Evaluation uses WER and BLEU on a held-out test set that includes six stations absent from training. The paper reports relative WER reductions of 19--33% and BLEU improvements of 8--40% over zero-shot, with the best model, ZHAW large-v3, reaching 17.1% WER and 74.8 BLEU. It also reports that fine-tuning on STT4SG-350 alone degrades SRB-300 performance, and that concatenated long-audio test results are similar to the segment-level results.

Significance. If the evaluation benchmark is valid, the paper makes a useful contribution: a realistic long-form speech corpus for a low-resource dialect-to-standard language pair, and evidence that fine-tuning Whisper on in-domain spontaneous speech yields substantial gains over zero-shot and over sentence-level fine-tuning. Strengths include the station-excluded test partition, the long-audio concatenation experiment, and the comparison across multiple Whisper sizes and XLS-R. The central quantitative claims, however, rest on reference transcripts that are themselves derived from Whisper large-v3 pre-transcription followed by limited manual correction, creating a risk that the reported gains partly reflect alignment with Whisper's own error patterns rather than genuine transcription and translation quality. The corpus is not publicly released, which limits reproducibility and reuse.

major comments (3)
  1. [Section 4.1 (Manual Correction) and Appendix D, Table 12] The SRB-300 references are Whisper large-v3 pre-transcriptions that humans correct only where 'clearly wrong'. The correction rules instruct annotators to leave repetitions and filler words 'as is', and footnote 16 notes that Whisper generally does not transcribe fillers; consequently, fillers are absent from the gold standard by construction. Because every fine-tuned model is initialized from the Whisper family and optimized to predict these Whisper-derived references, the WER/BLEU improvements in Table 4 may partly reward models that reproduce Whisper's systematic omissions and translation choices rather than transcription quality per se. This is load-bearing for the main claim. I ask the authors to validate on a subset with references produced from scratch by independent transcribers without any pre-transcription, or at least to quantify filler recall and hallucination rates on a human-audited sample and to report WER/BLEU on that independently annotated subset.
  2. [Section 4.2 (Dataset Partitioning)] The paper states that samples from the same speaker may appear in both training and test sets because the metadata lack unique speaker IDs. This speaker-level leakage can inflate the reported gains, and the six held-out stations do not eliminate it: a speaker appearing on station 271 in the test set could also appear on station 201 in the training set. The station-level results in Figure 4 provide partial evidence that unseen stations also improve, but I ask for a quantitative bound: report WER and BLEU restricted to the six stations excluded from training, and, if feasible, cluster samples by speaker or use per-broadcast bootstrap confidence intervals to assess the impact of speaker overlap.
  3. [Section 6, Table 4] No confidence intervals or significance tests are reported for the WER and BLEU differences, and the test samples are not independent because they are grouped into broadcasts. Differences such as ZHAW large-v2 (18.4 WER) versus ZHAW large-v3 (17.1 WER), or ZHAW medium (19.7) versus ZHAW medium 3 (19.4), may not be meaningful without an error estimate. Please report per-broadcast bootstrap confidence intervals or paired significance tests, and clarify whether any model selection (epochs, learning rate) used only the validation set, since Section 5 mentions 4--12 epochs but does not state the early-stopping or model-selection protocol.
minor comments (8)
  1. [Throughout] The captions of Tables 1, 2, 3, and others contain the typo 'T able' instead of 'Table'; please correct throughout.
  2. [Section 5] The learning rate notation '5 × e−6' and '1×e−6' should be typeset as 5 × 10^-6 and 1 × 10^-6, and missing spaces should be added.
  3. [Section 6] Minor formatting issues include 'evaluate-library' and 'NLTK[2]implementation', which should be 'evaluate library' and 'NLTK [2] implementation' (similarly for other hyphenated names).
  4. [References] Reference 2 contains 'iSBN' and references 19 and 24 contain 'iSSN'; these should be 'ISBN' and 'ISSN'.
  5. [Section 4.1] The statement that 'ß' is replaced with 'ss' because it does not exist in Swiss German could be clarified: if the target transcripts are Standard German, Swiss orthography conventions should be stated explicitly, since this choice affects BLEU references.
  6. [Section 6] The phrase '50% worse performance in BLEU score' is ambiguous; it would be clearer to say 'a relative decrease of about 50% in BLEU score'.
  7. [Table 1] Including the SRB-300 row (XLS-R 1B, this work) in a table titled 'reported results of the fine-tuned SOTA models on the test sets of the most common scientific corpora' is confusing because SRB-300 is introduced in this paper; consider moving that row to Table 4 or clearly labeling it as a new evaluation.
  8. [Figure 4] The y-axis starting at 40 visually exaggerates the improvements; consider starting at 0 or using a broken axis to avoid misleading readers.

Circularity Check

1 steps flagged · score 6.0 of 10

Whisper-derived SRB-300 references make the reported WER/BLEU gains partly a measure of fit to Whisper's own output, not independent real-world accuracy.

  1. fitted input called prediction [Section 4.1 (Pre-Transcription and Manual Correction); Appendix D Table 12, footnote 16]
    "The approximately 300 hours of speech from the extraction step were automatically pre-transcribed using OpenAI's Whisper large-v3 model. ... repetitions, filler words [footnote 16: Whisper generally does not transcribe fillers [15]] leave as is, as long as it is also contained in the audio."

    The SRB-300 ground truth is generated by Whisper large-v3 pre-transcription followed by manual correction. The correction rules fix only what is 'clearly wrong' and explicitly tell annotators to leave fillers as is, but the footnote states that Whisper generally does not transcribe fillers. Therefore, fillers are absent from the pre-transcription and never reinserted by annotators, so the reference transcripts systematically inherit Whisper's deletion behavior and any other Whisper biases that survive the bounded correction process. All fine-tuned models are Whisper variants trained to maximize likelihood on these references, and WER/BLEU are computed against the same Whisper-derived references after applying Whisper's text normalizer.

full rationale

The evaluation loop is not completely closed: human annotators did review and correct the Whisper pre-transcriptions, so SRB-300 is not literally Whisper output. However, the correction rules are deliberately bounded by what the pre-transcription contains. The filler-word rule is the concrete exhibit: annotators may leave a filler only if it is already in the audio transcript, while the paper's own footnote confirms Whisper generally omits fillers, so those omissions are cemented into the gold standard. Fine-tuning Whisper models on these references and then scoring them against the same reference type rewards exactly the deletions, translation choices, and other systematic behaviors of Whisper large-v3. The external citation for Whisper's filler behavior is independent evidence, so this is not a self-citation problem; it is a ground-truth-construction problem. The speaker overlap noted in Section 4.2 is a separate leakage risk and is not counted as circularity here. Because manual correction does provide some independent linguistic decisions (spelling, clearly wrong content), the central result is not forced by definition alone, but the headline improvements substantially reduce to agreement with a Whisper-derived labeling pipeline. Score 6 reflects this partial circularity.

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

The central claims depend on corpus construction and evaluation assumptions rather than on theoretical derivations. The dataset assumes station location is a proxy for speaker dialect; the evaluation assumes the human-corrected Whisper pre-transcriptions form an unbiased gold standard; and the metrics assume the task is best treated as translation into Standard German. Fine-tuning hyperparameters are chosen by hand and reported. No new theoretical entities are introduced.

free parameters (3)
  • initial learning rate (small/medium) = 5e-6
    Hyperparameter chosen by hand; likely affects final WER but not optimized against the test set.
  • initial learning rate (large) = 1e-6
    Lower learning rate for large models, chosen by hand.
  • number of epochs = 4 to 12
    Range stated without selection criterion; may affect results.
assumptions (3)
  • domain assumption The dialect spoken by a presenter aligns with the broadcast station's location (Section 4.1, Figure 2).
    The paper uses this to claim all major Swiss dialects are represented; the authors acknowledge it is an assumption.
  • domain assumption Whisper's pre-transcriptions, after manual correction, are accurate Standard German references (Section 4.1).
    If the manual correction fails to remove Whisper's systematic biases, the gold standard is biased toward Whisper's output; this is the load-bearing premise for the evaluation.
  • domain assumption WER and BLEU are appropriate metrics for Swiss German to Standard German transcription (Section 1).
    The task is treated as translation, and BLEU is used; this is a modeling choice that affects the reported numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing STT for Low-Resource Real-World Speech." pith.science (2026). https://pith.science/paper/TOTUV6A3

@misc{pith2026250608836,
  author       = {Pith},
  title        = {Pith review of: Advancing STT for Low-Resource Real-World Speech},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TOTUV6A3}},
  note         = {Machine review of arXiv:2506.08836}
}
read the original abstract

Swiss German is a low-resource language represented by diverse dialects that differ significantly from Standard German and from each other, lacking a standardized written form. As a result, transcribing Swiss German involves translating into Standard German. Existing datasets have been collected in controlled environments, yielding effective speech-to-text (STT) models, but these models struggle with spontaneous conversational speech. This paper, therefore, introduces the new SRB-300 dataset, a 300-hour annotated speech corpus featuring real-world long-audio recordings from 39 Swiss German radio and TV stations. It captures spontaneous speech across all major Swiss dialects recorded in various realistic environments and overcomes the limitation of prior sentence-level corpora. We fine-tuned multiple OpenAI Whisper models on the SRB-300 dataset, achieving notable enhancements over previous zero-shot performance metrics. Improvements in word error rate (WER) ranged from 19% to 33%, while BLEU scores increased between 8% and 40%. The best fine-tuned model, large-v3, achieved a WER of 17.1% and a BLEU score of 74.8. This advancement is crucial for developing effective and robust STT systems for Swiss German and other low-resource languages in real-world contexts.

Figures

Figures reproduced from arXiv: 2506.08836 by the authors.

Figure 1
Figure 1. Distribution of sample duration over all datasets in SRB-300. The durations range from 3.4 s to 30.0 s, with an average of 25.4 s. Each of the three datasets (ref. Section 4.2) has a similar distribution [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Distribution of samples over Switzerlands German-speaking area 4.2 Dataset Partitioning The SRB-300 speech corpus is divided into three datasets: training (76%), val￾idation (10%), and test (14%), as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Distribution of the total sample duration per dialect region 5 Experiments We used different sizes of OpenAI’s Whisper8 model for our fine-tuning experi￾ments. Whisper is an open-source multilingual transformer [29] model designed for automatic speech recognition, translation, and language detection [24]. It was trained on an extensive multilingual corpus collected through web crawling. Thanks to its internal segmen… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Improvement of mean BLEU score on the SRB-300 test set for all samples of each broadcasting station for the ZHAW large-v3 model, which was fine-tuned on SRB-300. The broadcast stations not in the training set are marked with a star. The y-axis begins at 40 for better r…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 15 canonical work pages

  1. [1]

    ss" instead of

    Babu, A., Wang, C., Tjandra, A., Lakhotia, K., Xu, Q., Goyal, N., Singh, K., Von Platen, P., Saraf, Y., Pino, J., Baevski, A., Conneau, A., Auli, M.: XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale. In: Interspeech 2022. pp. 2278–2282. ISCA (Sep 2022). https://doi.org/ 10.21437/Interspeech.2022-143, https://www.isca-archive.or...

    doi:10.21437/interspeech.2022-143 2022

  2. [2]

    O’Reilly Media Inc

    Bird, S., Klein, E., Loper, E.: Natural Language Processing with Python. O’Reilly Media Inc. (Jun 2009), iSBN: 9780596516499

    work page 2009

  3. [3]

    https://doi.org/10.48550/arXiv.1604.06174, http: //arxiv.org/abs/1604.06174, arXiv:1604.06174 [cs] version: 2

    Chen, T., Xu, B., Zhang, C., Guestrin, C.: Training Deep Nets with Sublin- ear Memory Cost (Apr 2016). https://doi.org/10.48550/arXiv.1604.06174, http: //arxiv.org/abs/1604.06174, arXiv:1604.06174 [cs] version: 2

    arXiv doi:10.48550/arxiv.1604.06174 2016

  4. [4]

    Deschamps-Berger, T.: ASR-whisper-finetuning (Oct 2024), https://github.com/ Theodb/ASR-whisper-finetuning

    work page 2024

  5. [5]

    Do, A., Brown, O., Wang, Z., Mathew, N., Liu, Z., Ahmed, J., Yu, C.: Using fine-tuning and min lookahead beam search to improve Whisper (Sep 2023), http: //arxiv.org/abs/2309.10299, arXiv:2309.10299 [cs, eess]

    work page Pith review arXiv 2023

  6. [6]

    Dogan-Schönberger, P., Mäder, J., Hofmann, T.: SwissDial: Parallel Multidialectal Corpus of Spoken Swiss German (Mar 2021), http://arxiv.org/abs/2103.11401, arXiv:2103.11401 [cs]

    work page Pith review arXiv 2021

  7. [7]

    In: Scherrer, Y., Jauhiainen, T., Ljubešić, N., Zampieri, M., Nakov, P., Tiedemann, J

    Dolev, E., Lutz, C., Aepli, N.: Does Whisper Understand Swiss German? An Automatic, Qualitative, and Human Evaluation. In: Scherrer, Y., Jauhiainen, T., Ljubešić, N., Zampieri, M., Nakov, P., Tiedemann, J. (eds.) Proceedings of the Eleventh Workshop on NLP for Similar Languages, Varieties, and Di- alects (VarDial 2024). pp. 28–40. Association for Computat...

    doi:10.18653/v1/2024.vardial-1.3 2024

  8. [8]

    In: ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Ferraz, T.P., Zanon Boito, M., Brun, C., Nikoulina, V.: Multilingual Distilwhis- per: Efficient Distillation of Multi-Task Speech Models Via Language-Specific Experts. In: ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 10716–10720 (Apr 2024). https://doi. org/10.1109/ICASSP48485.2024.10447520, htt...

    arXiv 2024

Show all 30 references
  1. [9]

    Gandhi, S.: Fine-Tune Whisper For Multilingual ASR with Hugging Face Trans- formers (Nov 2022), https://huggingface.co/blog/fine-tune-whisper

    2022

  2. [10]

    In: Interspeech 2020

    Gulati, A., Qin, J., Chiu, C.C., Parmar, N., Zhang, Y., Yu, J., Han, W., Wang, S., Zhang, Z., Wu, Y., Pang, R.: Conformer: Convolution-augmented Trans- former for Speech Recognition. In: Interspeech 2020. pp. 5036–5040. ISCA (Oct Advancing STT for Low-Resource Real-World Speec...

    2020 doi

  3. [11]

    https://doi.org/10.48550/arXiv

    Hsu, M.H., Huang, K.P., Lee, H.y.: Meta-Whisper: Speech-Based Meta-ICL for ASR on Low-Resource Languages (Sep 2024). https://doi.org/10.48550/arXiv. 2409.10429, http://arxiv.org/abs/2409.10429, arXiv:2409.10429 [eess]

    2024 doi arXiv

  4. [12]

    EURASIP Journal on Audio, Speech, and Music Process- ing 2024(1), 29 (Jun 2024)

    Liu, Y., Yang, X., Qu, D.: Exploration of Whisper fine-tuning strategies for low-resource ASR. EURASIP Journal on Audio, Speech, and Music Process- ing 2024(1), 29 (Jun 2024). https://doi.org/10.1186/s13636-024-00349-3, https: //doi.org/10.1186/s13636-024-00349-3

    2024 doi

  5. [13]

    Lodagala, V.: Fine-tuning and evaluating Whisper models for Automatic Speech Recognition (May 2023), https://github.com/vasistalodagala/whisper-finetune

    2023

  6. [14]

    In: 7th Inter- national Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

    Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: 7th Inter- national Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net (2019), https://openreview.net/forum?id= Bkg6RiCqY7

    2019

  7. [15]

    In: 9th Workshop on Speech and Language Tech- nology in Education (SLaTE)

    Ma, R., Qian, M., Gales, M., Knill, K.M.: Adapting an ASR Foundation Model for Spoken Language Assessment. In: 9th Workshop on Speech and Language Tech- nology in Education (SLaTE). pp. 104–108. ISCA (Aug 2023). https://doi.org/ 10.21437/SLaTE.2023-20, https://www.isca-archive...

    2023 doi

  8. [16]

    In: Bouamor, H., Pino, J., Bali, K

    Paonessa, C., Schraner, Y., Deriu, J., Hürlimann, M., Vogel, M., Cieliebak, M.: Dialect Transfer for Swiss German Speech Translation. In: Bouamor, H., Pino, J., Bali, K. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 15240–15254. Association ...

    2023 doi

  9. [17]

    In: Isabelle, P., Charniak, E., Lin, D

    Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a Method for Automatic Evaluation of Machine Translation. In: Isabelle, P., Charniak, E., Lin, D. (eds.) Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. pp. 311–318. Association for C...

    2002

  10. [18]

    https://doi.org/10.48550/arXiv.2411.04573, http://arxiv.org/abs/ 2411.04573, arXiv:2411.04573 [cs]

    Pillai, L.G., Manohar, K., Raju, B.K., Sherly, E.: Multistage Fine-tuning Strategies for Automatic Speech Recognition in Low-resource Languages (Nov 2024). https://doi.org/10.48550/arXiv.2411.04573, http://arxiv.org/abs/ 2411.04573, arXiv:2411.04573 [cs]

    2024 doi arXiv

  11. [19]

    In: Interspeech 2024

    Piñeiro-Martín, A., García-Mateo, C., Docio-Fernandez, L., López-Pérez, M.D.C., Rehm, G.: Weighted Cross-entropy for Low-Resource Languages in Multilingual Speech Recognition. In: Interspeech 2024. pp. 1235–1239. ISCA, Kos, Greece (Sep 2024). https://doi.org/10.21437/Interspee...

    2024 doi

  12. [20]

    In: Rogers, A., Boyd-Graber, J., Okazaki, N

    Plüss, M., Deriu, J., Schraner, Y., Paonessa, C., Hartmann, J., Schmidt, L., Scheller, C., Hürlimann, M., Samardžić, T., Vogel, M., Cieliebak, M.: STT4SG- 350: A Speech Corpus for All Swiss German Dialect Regions. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings...

    2023 doi

  13. [21]

    In: Calzolari, N., Béchet, F., Blache, P., Choukri, K., Cieri, C., Declerck, T., Goggi, S., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Odijk, J., Piperidis, S

    Plüss,M.,Hürlimann,M.,Cuny,M.,Stöckli,A.,Kapotis,N.,Hartmann,J.,Ulasik, M.A., Scheller, C., Schraner, Y., Jain, A., Deriu, J., Cieliebak, M., Vogel, M.: SDS- 200: A Swiss German Speech to Standard German Text Corpus. In: Calzolari, N., Béchet, F., Blache, P., Choukri, K., Cier...

    2022

  14. [22]

    In: Proceedings of the Swiss Text Analytics Conference 2021

    Plüss, M., Neukom, L., Scheller, C., Vogel, M.: Swiss Parliaments Corpus, an Automatically Aligned Swiss German Speech to Standard German Text Corpus. In: Proceedings of the Swiss Text Analytics Conference 2021. Online (Jun 2021), https://ceur-ws.org/Vol-2957/paper3.pdf

    2021

  15. [23]

    In: Interspeech 2024

    Qian, M., Tang, S., Ma, R., Knill, K.M., Gales, M.J.: Learn and Don’t For- get: Adding a New Language to ASR Foundation Models. In: Interspeech 2024. pp. 2544–2548. ISCA (Sep 2024). https://doi.org/10.21437/Interspeech.2024-1045, https://www.isca-archive.org/interspeech_2024/q...

    2024 doi

  16. [24]

    In: Proceedings of the 40th International Conference on Machine Learning

    Radford, A., Kim, J.W., Xu, T., Brockman, G., Mcleavey, C., Sutskever, I.: Ro- bust Speech Recognition via Large-Scale Weak Supervision. In: Proceedings of the 40th International Conference on Machine Learning. pp. 28492–28518. ICML’23, JMLR.org, Honolulu, Hawaii, USA (Jul 202...

    2023 doi

  17. [25]

    Schraner, Y., Scheller, C., Plüss, M., Vogel, M.: Swiss German Speech to Text system evaluation (Nov 2022), http://arxiv.org/abs/2207.00412, arXiv:2207.00412 [cs]

    2022 arXiv

  18. [26]

    In: Ghorbel, H., Sokhn, M., Cieliebak, M., Hür- limann, M., de Salis, E., Guerne, J

    Sicard, C., Gillioz, V., Pyszkowski, K.: Spaiche: Extending State-of-the-Art ASR Models to Swiss German Dialects. In: Ghorbel, H., Sokhn, M., Cieliebak, M., Hür- limann, M., de Salis, E., Guerne, J. (eds.) Proceedings of the 8th edition of the Swiss Text Analytics Conference. ...

    2023

  19. [27]

    Srivastav, V.: Faster Whisper Finetuning with LoRA powered by Hugging Face PEFT (Apr 2023), https://github.com/Vaibhavs10/fast-whisper-finetuning

    2023

  20. [28]

    https://doi.org/10.48550/arXiv.2412.15726, http://arxiv.org/abs/ 2412.15726, arXiv:2412.15726 [cs]

    Timmel, V., Paonessa, C., Kakooee, R., Vogel, M., Perruchoud, D.: Fine- tuning Whisper on Low-Resource Languages for Real-World Applications (Dec 2024). https://doi.org/10.48550/arXiv.2412.15726, http://arxiv.org/abs/ 2412.15726, arXiv:2412.15726 [cs]

    2024 doi arXiv

  21. [29]

    In: Advances in Neural In- formation Processing Systems

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Advances in Neural In- formation Processing Systems. vol. 30, pp. 6000–6010. Curran Associates, Inc. (2017). https://doi.org/10.5555/3295222....

    2017

  22. [30]

    In: Liu, Q., Schlangen, D

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., Davison, J., Shleifer, S., von Platen, P., Ma, C., Jernite, Y., Plu, J., Xu, C., Le Scao, T., Gugger, S., Drame, M., Lhoest, Q., Rush, A.: Transformers: State-of...

    2020

Pith tools

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