Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Voice Adaptation for Swiss German

T0 review · 3 major / 5 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 This paper establishes that a voice-adaptation model for Swiss German can be built from roughly 5,000 hours of automatically transcribed podcast audio, and that human listeners rate its output close to reference-quality speech.

desk verdict A genuinely useful Swiss German TTS corpus and recipe, but the dialect-control claim rests on a circular metric and Zurich DID drops to 0.071. read the letter →

arxiv 2505.22054 v1 pith:YERTPMHE submitted 2025-05-28 cs.CL cs.SDeess.AS

Samuel Stucki , Jan Deriu , Mark Cieliebak This is my paper
classification cs.CLcs.SDeess.AS
keywords voiceadaptationSwissGermantext-to-speechdialectsynthesislow-resourcelanguageXTTSv2weaklylabeledcorpusspeech
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 tries to establish that voice adaptation—cloning a reference speaker's voice while speaking a chosen dialect—is feasible for Swiss German, a spoken-only language family with no standard orthography. The authors build a roughly 5,000-hour corpus from Swiss podcasts by automatically transcribing Swiss German audio into Standard German text and tagging each segment with one of seven dialect regions, then fine-tune the multilingual XTTSv2 text-to-speech model on it. In human evaluations, the best model reaches comparative MOS of -0.28 (near the reference) in the long scenario and speaker-similarity MOS of 3.8, and automated back-translation and dialect-recognition scores support the same conclusion. If correct, this is a step toward bringing voice-cloning technology to an underrepresented language without expensive manual data collection.

What carries the argument

The load-bearing mechanism is a fully automated data pipeline that turns raw Swiss podcast audio into weakly labeled training material: speaker diarization and voice activity detection split the audio into single-speaker segments, a large automatic speech recognition model (Whisper-Large-V3) produces Standard German transcripts, and a phoneme n-gram Naive Bayes classifier assigns each speaker one of seven Swiss German dialect regions. This corpus, SRG, is then used to fine-tune XTTSv2, a multilingual zero-shot text-to-speech model that takes a text prompt and a short reference voice as inputs. The same phoneme-based classifier is reused as the dialect-recognition evaluation metric (DID), which is why it both enables and constrains the dialect-rendering claim.

What would settle it

Ask native Swiss German speakers from each region to listen to a blinded set of generated utterances and name the dialect; if Zurich and Central Switzerland utterances are confused at rates near the classifier's own confusion (Zurich DID as low as 0.071–0.457 in the paper's long scenario), the claim that the model correctly renders the desired dialect would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Standard German text-to-speech model can be adapted to render Swiss German in a user-selected dialect while preserving a reference voice, using only automatically produced training data. The supporting result is that fine-tuning XTTSv2, first on the noisy pseudo-labeled SRG corpus and then on the clean STT4SG-350 corpus, produces a model whose output is rated close to reference quality by human listeners, with the strongest combination showing CMOS -0.29 and SMOS 3.81 in the long scenario and intelligibility 4.51 in the short scenario. The authors also claim that dialect correctness is high across regions, with the known exception of Zurich being under-recognized because the dialect classifier itself confuses Zurich with Central Switzerland.

Load-bearing premise

The whole result rests on the automatic dialect classifier's labels being right, but the same classifier is known to confuse Zurich with Central Switzerland, so a non-trivial share of the training and evaluation labels could be wrong without being detected.

Editorial extensions

If this is right

  • A voice-adaptation system for Swiss German can be trained from automatically processed podcasts, without manual parallel data, and can render the requested dialect for most regions.
  • The two-stage recipe—pretrain on the large noisy SRG corpus, then fine-tune on clean STT4SG-350—gives the best intelligibility and back-translation scores on short utterances, at a small cost in speaker similarity.
  • Dialect coverage is uneven: Valais, with under 1% of the data, is the weakest dialect, while Central Switzerland benefits from the abundance of Zurich data.
  • Zurich dialect rendering is systematically underestimated in automated dialect recognition because the evaluation classifier itself confuses Zurich with Central Switzerland, so the true Zurich accuracy remains uncertain.
  • Human raters judged the SRG-trained models as close to or better than the baseline on naturalness (CMOS) and speaker similarity (SMOS), especially for longer 10–15 second utterances.

Reading between the lines

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

  • Because the same classifier supplies both the labels that condition the model and the DID score that evaluates it, an independent human dialect-judgment test would be needed to confirm the dialect-rendering claim.
  • The same automated pipeline should transfer to other dialect continua or low-resource languages wherever a strong speech recognizer and a small seed corpus for dialect classification exist.
  • Since the transcripts are translations rather than faithful transcriptions, the back-translation WER/BLEU scores mix rendering errors with transcription-level differences; a separate human intelligibility check already partly covers this.
  • The strong long-scenario result for SRG+STT4SG hints that matching training segment length to deployment length may matter more than additional clean data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper addresses voice adaptation (text-to-speech with a reference voice) for Swiss German, a low-resource dialect continuum. The authors build the SRG corpus from Swiss public-broadcast podcasts: voice activity detection and diarization produce roughly 1.8 million segments, WhisperX provides Standard German transcripts, and a phoneme n-gram Naive Bayes classifier (Section 2.4) assigns one of seven Swiss German dialect regions or Standard German to each speaker. They then fine-tune XTTS-v2 on this data, with and without a final fine-tuning pass on the STT4SG-350 corpus, yielding three models. Evaluations on the STT4SG-350 test set cover automatic back-translation WER/BLEU, ECAPA2 speaker similarity, dialect recognition (DID), and human SMOS/CMOS/intelligibility ratings. The best model, SRG+STT4SG++, reaches WER 0.328/BLEU 0.607 in the Short scenario and 0.156/0.786 in the Long scenario, with human intelligibility up to 4.51; the paper concludes that translating Standard German text to Swiss German speech is feasible and that the desired dialect is correctly rendered.

Significance. If the results hold, this is a useful step for a genuinely under-resourced language group: it demonstrates that weakly supervised podcast data can improve an open TTS backbone, and it reports human ratings from native raters, which is stronger evidence than purely automatic metrics. The dataset-construction pipeline and the consistent relative ordering of the three models (SRG+STT4SG++ best on intelligibility and back-translation, SRG+STT4SG best on speaker similarity in the Long scenario) are concrete contributions. The main weakness is that the dialect-rendering claim relies on a DID metric computed with the same classifier that produced the training labels, and no independent dialect check is offered; for Zurich the internal metric collapses. The central feasibility claim (intelligible, speaker-similar synthetic Swiss German) is nevertheless supported by non-circular human and back-translation evidence.

major comments (3)
  1. [Section 4.1 / Tables 3-4] The DID column is measured with "our phoneme-based classification pipeline (c.f. 2.4)"—the same classifier that assigned every dialect label in the SRG training corpus. Because the model is conditioned on those labels, near-ceiling DID scores for most regions may simply show that the model reproduces the classifier's dialect prototypes, including its systematic errors. The paper's own numbers show the metric is not trustworthy for Zurich: SRG+STT4SG++ achieves DID 0.457 (Short) and 0.071 (Long), and even the Baseline reaches only 0.313/0.200. Since Section 4 explicitly states that human raters did not judge dialect accuracy, there is no independent evidence for the abstract's claim that the model "can correctly render the desired dialect." I ask for an external dialect classifier or an expert listening test, or a substantial narrowing of the claim.
  2. [Section 2.4] The same classifier is applied out-of-distribution to label the SRG corpus. It is trained on STT4SG-350 plus 30 hours of Common Voice and evaluated on 30-second concatenated samples (macro F1 0.88), but it is then applied to merged same-speaker segments from spontaneous podcasts, for which no error analysis is reported. The paper concedes that the classifier "often confuses" Zurich and Central Switzerland. Because these labels are the only dialect supervision for the 4,979-hour corpus, the Zurich split (23.67% of the data, Table 1) is potentially contaminated with Central Swiss or mixed-dialect speech. The manuscript should quantify this label noise—for instance, with a manual audit of a sample of speaker-level labels or a confusion analysis on held-out podcast audio—and discuss how such noise propagates to the Zurich DID collapse.
  3. [Section 4.3 / Abstract] The statement "dialect correctness is generally high, except for Zurich misclassification due to classifier limitations" treats the DID scores as ground truth. Given the circularity described above, the correct summary is that dialect correctness is not independently measured for any region, and for Zurich even the internal metric fails. The conclusion and abstract should be revised to reflect this evidential gap, or new independent evidence should be supplied.
minor comments (5)
  1. [Section 1 / References] References [7] and [9] are both the AutoPrep paper, while the sentence in Section 1 that cites "Emilia [7]" points to the wrong reference; the Emilia entry appears to be missing.
  2. [Section 4] The evaluation description says 1400 generated utterances (200 per dialect region), but the tables contain eight regions (seven Swiss German regions plus German), which would imply 1600 utterances; please clarify the dialect-region count and the sampling procedure.
  3. [Table 5] The significance markers * and † are not tied to a stated statistical test; please specify the test used and the significance threshold.
  4. [Figure 1 / Table 1] Figure 1 reports 1.8M segments and 4.7k hours, while Table 1 reports 1,764K samples and 4,979.1 hours; please reconcile these numbers or explain the difference.
  5. [Sections 1 and 4.1] Calling a CMOS of -0.28 "almost human quality" is an interpretive leap, since CMOS is a relative comparison against a reference recording; please justify or soften this phrasing.

Circularity Check

1 steps flagged · score 6.0 of 10

Dialect correctness is measured by the same classifier that created the training labels, so the central claim 'can correctly render the desired dialect' is largely a self-consistency result rather than an independent verification.

  1. self definitional [Section 2.4 and Section 4.1, Tables 3-4]
    "To control the dialect during the generation, we automatically annotate the dialect of each sample. For this, we use the approach described in [19], which is based on a pipeline that converts audio to phonemes using [20] and then trains a simple Naive Bayes model on top of phoneme n-grams. ... Dialect Recognition Accuracy (DID): Measured using our phoneme-based classification pipeline (c.f. 2.4)."

    The same phoneme n-gram Naive Bayes classifier assigns the dialect labels that condition the XTTSv2 model during training (Section 2.4) and also produces the DID scores reported as dialect correctness (Section 4.1). Thus the DID evaluation measures how well generated audio matches the classifier's own label distribution, not an independent ground truth. The paper concedes the classifier 'often confuses the two geographically close regions, Zurich and Central Switzerland', and the numbers show it: Zurich DID collapses to 0.412/0.457 (Short, SRG models) and 0.071 (Long, SRG+STT4SG++).

full rationale

The central dialect claim rests on the DID column of Tables 3 and 4, which is generated by the same phoneme Naive Bayes classifier (Section 2.4) that assigned every dialect label in the SRG training corpus. The model is conditioned on those labels, and the evaluation asks whether the generated audio receives the same label from the same classifier. High DID scores are therefore partly a measure of self-consistency with the labeling instrument rather than an independent confirmation of dialect correctness. The paper's own caveat that the classifier 'often confuses' Zurich and Central Switzerland, together with the Zurich DID collapse to 0.071-0.457, demonstrates that the metric is sensitive to the instrument's biases without correcting them. Because human raters were not asked to judge dialect accuracy (the paper states such evaluation is 'often unreliable'), no independent dialect check is present. The WER, BLEU, SMOS, CMOS and intelligibility results are independent external evaluations of intelligibility and voice quality, so the paper is not wholly circular; however, the abstract's 'correctly render the desired dialect' claim is load-bearing and reduces, to a significant degree, to the classifier's own labels.

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

No new physical or theoretical entities are postulated; the contributions are a dataset and trained models. The ledger is dominated by hand-chosen pipeline parameters and by three automated stages (diarization, transcription, dialect labeling) whose outputs are treated as ground truth. The most fragile of these is the dialect classifier, because it both labels the training data and serves as the evaluation instrument for dialect rendering.

free parameters (5)
  • Learning rate 6e-5 = 6e-5
    Deviation from XTTS-v2's 5e-5, chosen 'due to internal tests and listening to the generated audio files' (Section 3); the reported results depend on this hand-tuned value.
  • Segment length cap = 15 seconds
    All speech longer than 15s is cut into uniform 15s chunks (Section 2.2); the paper notes this dominates the segment-length distribution and shapes the Short-scenario results.
  • Diarization speaker range = 2 to 6 speakers
    pyannote is configured with an assumed speaker range of 2-6 per episode (Section 2.2); over-segmentation is accepted as a trade-off.
  • Minimum segment duration = 2 seconds
    Segments below 2s are removed (Section 2.2); this filters some short utterances from the training data.
  • Dialect classifier training recipe = 30h per dialect plus 30h German Common Voice
    The Naive Bayes phoneme n-gram model is trained on phonemized STT4SG-350 (30h per dialect) enriched with German Common Voice (Section 2.4); label quality for the whole corpus is determined by this choice.
assumptions (5)
  • domain assumption Whisper-Large-V3 produces adequate Standard German translations of Swiss German podcast speech
    The entire pseudolabel corpus rests on these transcripts; validation is only 100 manually transcribed samples (median WER 21.1%, BLEU 64.08) at a scale of 1.7M segments.
  • domain assumption pyannote diarization yields single-speaker segments that are good enough for TTS training
    DER of 14.15% measured on a single 42.63-minute episode (Section 2.2) is extrapolated to the full corpus.
  • domain assumption The seven-region dialect taxonomy is the right conditioning granularity
    Adopted from references [10, 11, 19] without argument; real dialectal variation exists within each region, and region-level conditioning is the design choice.
  • domain assumption XTTS-v2 can absorb a new dialect condition from weakly labeled data
    The training section assumes the pretrained model can be extended to Swiss German dialect regions with about 5,000 hours of noisy data; the paper's results are the evidence for this capability rather than an independent proof.
  • ad hoc to paper CMOS near zero can be interpreted as 'almost human quality'
    The phrase 'almost human quality' (Abstract) is attached to CMOS -0.28, but the measured -0.29 +/- 0.74 (Table 5) has an interval overlapping zero.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Voice Adaptation for Swiss German." pith.science (2026). https://pith.science/paper/YERTPMHE

@misc{pith2026250522054,
  author       = {Pith},
  title        = {Pith review of: Voice Adaptation for Swiss German},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YERTPMHE}},
  note         = {Machine review of arXiv:2505.22054}
}
read the original abstract

This work investigates the performance of Voice Adaptation models for Swiss German dialects, i.e., translating Standard German text to Swiss German dialect speech. For this, we preprocess a large dataset of Swiss podcasts, which we automatically transcribe and annotate with dialect classes, yielding approximately 5000 hours of weakly labeled training material. We fine-tune the XTTSv2 model on this dataset and show that it achieves good scores in human and automated evaluations and can correctly render the desired dialect. Our work shows a step towards adapting Voice Cloning technology to underrepresented languages. The resulting model achieves CMOS scores of up to -0.28 and SMOS scores of 3.8.

Figures

Figures reproduced from arXiv: 2505.22054 by the authors.

Figure 1
Figure 1. Data Processing Pipeline for pseudo labeled data. 2.1. Download We manually selected 25 podcasts out of 34, filtering out those programs that mainly contain music. Each podcast contains multiple episodes. Thus, each podcast’s audio range varies be￾tween 9 hours and 1661 hours of raw audio. We manually clas￾sified the podcasts as being mainly Standard German speech (13), mainly Swiss German speech (6), or mixed (6) b… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 18 canonical work pages

  1. [1]

    It is now possible to clone a voice across languages with less than a minute of audio required [3, 4]

    Introduction V oice Adaptation has seen an unprecedented increase in perfor- mance and utility in recent years [1, 2]. It is now possible to clone a voice across languages with less than a minute of audio required [3, 4]. The two factors that led to these advancements are using large amounts of data and computing power, which are well-leveraged using the ...

    work page

  2. [2]

    Voice Adaptation for Swiss German

    Data Collection Pipeline In this section, we describe the creation of the pseudo-labeled dataset, which we ref to as SRG-corpus. We leverage data from the Swiss Broadcasting Cooperation (SRG) using their API 2, where we access available podcasts 3. These podcasts cover various topics, from politics to entertainment and science. Fig- ure 1 shows the proces...

    work page Pith review arXiv 2025

  3. [3]

    For our first model, we fine- tuned XTTS-v2 using the SRG and STT4SG-350 data mix

    Model Training As a baseline, we fine-tuned the pre-trained XTTS-v2 [4] on the STT4SG-350 data, which consists of 343 hours of training data, we refer to this model as Baseline. For our first model, we fine- tuned XTTS-v2 using the SRG and STT4SG-350 data mix. We Model Description Steps Baseline XTTSv2 fine-tuned on STT4SG. 170k SRG+STT4SG XTTSv2 fine-tun...

    work page

  4. [4]

    The learning rate was changed to 6e-5 from the original 5e-5 due to internal tests and listening to the generated audio files

    for the training setup, applying an AdamW optimizer with betas 0.9 and 0.96 and weight decay 0.01. The learning rate was changed to 6e-5 from the original 5e-5 due to internal tests and listening to the generated audio files. Weight decay was applied only to the weights, and the learning rate was decayed using MultiStepLR with a gamma of 0.5 using milesto...

    work page

  5. [5]

    The training and test sets were pre- partitioned by the dataset authors to ensure speaker indepen- dence, such that no speaker or sample appears in both splits

    Evaluation and Results We conducted two types of evaluations on the STT4SG-350 test set: automated and human. The training and test sets were pre- partitioned by the dataset authors to ensure speaker indepen- dence, such that no speaker or sample appears in both splits. The evaluation covers two scenarios. • Short: Texts from the STT4SG-350 test set corre...

    work page

  6. [6]

    We showed that translating Standard German text to Swiss German dialect speech is feasible and yields satisfactory results

    Conclusion This work presented the transfer of voice adaptation technolo- gies to the challenging scenario of Swiss German dialects. We showed that translating Standard German text to Swiss German dialect speech is feasible and yields satisfactory results. For this, we created the SRG corpus of approximately 5000 hours of Swiss German dialect speech and f...

    work page

  7. [7]

    Neural codec language mod- els are zero-shot text to speech synthesizers,

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al. , “Neural codec language mod- els are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111, 2023

    arXiv 2023

  8. [8]

    Vall-e 2: Neural codec language models are hu- man parity zero-shot text to speech synthesizers,

    S. Chen, S. Liu, L. Zhou, Y . Liu, X. Tan, J. Li, S. Zhao, Y . Qian, and F. Wei, “Vall-e 2: Neural codec language models are hu- man parity zero-shot text to speech synthesizers,” arXiv preprint arXiv:2406.05370, 2024

    arXiv 2024

Show all 28 references
  1. [9]

    Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,

    Z. Zhang, L. Zhou, C. Wang, S. Chen, Y . Wu, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Liet al., “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,”arXiv preprint arXiv:2303.03926, 2023

    2023 arXiv

  2. [10]

    XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morais, S. Olayemi, and J. We- ber, “XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model,” in Interspeech 2024, 2024, pp. 4978–4982

    2024

  3. [11]

    Atten- tion is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Atten- tion is all you need,” in Advances in Neural Information Processing Systems , I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Ga...

    2017

  4. [12]

    Libriheavy: A 50,000 hours asr corpus with punc- tuation casing and context,

    W. Kang, X. Yang, Z. Yao, F. Kuang, Y . Yang, L. Guo, L. Lin, and D. Povey, “Libriheavy: A 50,000 hours asr corpus with punc- tuation casing and context,” in ICASSP 2024 - 2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 10 991–10 995

    2024

  5. [14]

    Wenetspeech4tts: A 12,800-hour mandarin tts corpus for large speech generation model bench- mark,

    L. Ma, D. Guo, K. Song, Y . Jiang, S. Wang, L. Xue, W. Xu, H. Zhao, B. Zhang, and L. Xie, “Wenetspeech4tts: A 12,800-hour mandarin tts corpus for large speech generation model bench- mark,” in Interspeech 2024, 2024, pp. 1840–1844

    2024

  6. [15]

    Autoprep: An automatic preprocessing framework for in-the-wild speech data,

    J. Yu, H. Chen, Y . Bian, X. Li, Y . Luo, J. Tian, M. Liu, J. Jiang, and S. Wang, “Autoprep: An automatic preprocessing framework for in-the-wild speech data,” in ICASSP 2024 - 2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. ...

    2024

  7. [16]

    SDS-200: A Swiss German speech to Standard German text corpus,

    M. Pl ¨uss, M. H ¨urlimann, M. Cuny, A. St ¨ockli, N. Kapotis, J. Hartmann, M. A. Ulasik, C. Scheller, Y . Schraner, A. Jain, J. Deriu, M. Cieliebak, and M. V ogel, “SDS-200: A Swiss German speech to Standard German text corpus,” in Proceedings of the Thirteenth Language Resou...

    2022

  8. [17]

    STT4SG-350: A speech corpus for all Swiss German dialect regions,

    M. Pl ¨uss, J. Deriu, Y . Schraner, C. Paonessa, J. Hartmann, L. Schmidt, C. Scheller, M. H ¨urlimann, T. Samardˇzi´c, M. V ogel, and M. Cieliebak, “STT4SG-350: A speech corpus for all Swiss German dialect regions,” in Proceedings of the 61st Annual Meeting of the Association ...

    2023

  9. [18]

    Fine-tuning whisper on low-resource languages for real- world applications,

    V . Timmel, C. Paonessa, R. Kakooee, M. V ogel, and D. Perru- choud, “Fine-tuning whisper on low-resource languages for real- world applications,” arXiv preprint arXiv:2412.15726, 2024

    2024 arXiv

  10. [19]

    Swissdial: Parallel multidialectal corpus of spoken swiss german,

    P. Dogan-Sch ¨onberger, J. M ¨ader, and T. Hofmann, “Swissdial: Parallel multidialectal corpus of spoken swiss german,” arXiv preprint arXiv:2103.11401, 2021

    2021 arXiv

  11. [20]

    Natural tts synthesis by condi- tioning wavenet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, R. Skerrv-Ryan, R. A. Saurous, Y . Agiomvrgiannakis, and Y . Wu, “Natural tts synthesis by condi- tioning wavenet on mel spectrogram predictions,” in 2018 IEEE International Conferenc...

    2018

  12. [21]

    pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,

    H. Bredin, “pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,” in Proc. INTERSPEECH 2023 , 2023

    2023

  13. [22]

    Powerset multi-class cross entropy loss for neural speaker diarization,

    A. Plaquet and H. Bredin, “Powerset multi-class cross entropy loss for neural speaker diarization,” in Proc. INTERSPEECH 2023, 2023

    2023

  14. [23]

    ELAN (Version 6.8) [Computer software],

    “ELAN (Version 6.8) [Computer software],” Nijmegen, The Netherlands, 2024. [Online]. Available: https://archive.mpi.nl/tla/ elan

    2024

  15. [24]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” in Proceedings of the 40th International Conference on Machine Learning, ser. ICML’23. JMLR.org, 2023

    2023

  16. [25]

    Automatische erkennung schweizerdeutscher dialekte anhand von audiodaten via phonem- transkriptionen,

    L. Bolliger and S. Waldburger, “Automatische erkennung schweizerdeutscher dialekte anhand von audiodaten via phonem- transkriptionen,” ZHAW Z ¨urcher Hochschule f ¨ur Angewandte Wissenschaften, Tech. Rep., 2024

    2024

  17. [26]

    Simple and effective zero-shot cross-lingual phoneme recognition,

    Q. Xu, A. Baevski, and M. Auli, “Simple and effective zero-shot cross-lingual phoneme recognition,” arXiv preprint arXiv:2109.11680, 2021

    2021 arXiv

  18. [27]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the Twelfth Language Resources and Evaluation Conference, N. Calzolari, F. B ´echet, P. ...

    2020

  19. [28]

    Ecapa2: A hybrid neural net- work architecture and training strategy for robust speaker embed- dings,

    J. Thienpondt and K. Demuynck, “Ecapa2: A hybrid neural net- work architecture and training strategy for robust speaker embed- dings,” in 2023 IEEE Automatic Speech Recognition and Under- standing Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  20. [29]

    Dialect transfer for Swiss German speech translation,

    C. Paonessa, Y . Schraner, J. Deriu, M. H ¨urlimann, M. V ogel, and M. Cieliebak, “Dialect transfer for Swiss German speech translation,” in Findings of the Association for Computational Linguistics: EMNLP 2023 , H. Bouamor, J. Pino, and K. Bali, Eds. Singapore: Association fo...

    2023

Pith tools

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