Pith. sign in

REVIEW 2 major objections 5 minor 35 references

Synthetic speech from multiple TTS systems can train Luxembourgish spoken QA without large human-recorded question corpora.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 07:11 UTC pith:4KAOICC4

load-bearing objection Solid low-resource SQA recipe with a useful MOS-vs-task finding; the multi-source win is partly confounded with data scale, which the paper itself flags. the 2 major comments →

arxiv 2607.02763 v1 pith:4KAOICC4 submitted 2026-07-02 cs.CL

LuxSQA: Ask Me in Luxembourgish with TTS-Augmented Spoken Question Answering

Nina Hosseini-Kivanani , Marco Matassoni , Alessio Brutti This is my paper
classification cs.CL
keywords Luxembourgishspoken question answeringtext-to-speech augmentationspeech-LLMlow-resource languagesSLAM-style projectorTTS quality vs downstream performance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Spoken question answering has mostly been built for high-resource languages with carefully recorded speech. This paper asks whether machine translation plus text-to-speech can create usable training data for Luxembourgish SQA when almost no human-recorded QA pairs exist. The authors translate English questions, synthesize them with several TTS engines, pair them with textual answers, and train a compact connector that maps a frozen speech encoder into frozen multilingual language models. On a real two-speaker Luxembourgish test set, multi-source and voice-design synthetic data give the strongest results, while automatic MOS-style TTS scores do not rank systems the same way. The practical point is that low-resource speech-LLM work can start from text QA and carefully chosen synthetic speech, provided the synthetic data is judged by downstream QA success rather than by how natural it sounds.

Core claim

Multi-source and voice-design-based TTS training configurations produce the strongest Luxembourgish spoken QA on real speakers, and no-reference TTS quality scores do not monotonically predict that performance, so synthetic speech has to be treated as task-specific training data rather than as audio judged only for naturalness.

What carries the argument

A SLAM-style speech-to-answer pipeline: frozen Whisper encoder, learned linear projector, and LoRA adapters on frozen multilingual LLMs, trained on translated questions synthesized by multiple TTS systems (including a ~230k multi-source mix) and supervised directly with textual answers.

Load-bearing premise

The claim that synthetic multi-source training transfers to real Luxembourgish SQA rests on two native speakers, a 300-question test set, machine-translated questions, and automatic semantic and LLM-judge metrics being representative enough to rank TTS sources and show that synthetic data can stand in for missing human QA recordings.

What would settle it

Train the same projector setups on the same synthetic mixes, then evaluate on a larger multi-speaker, naturally recorded Luxembourgish SQA set with human answer judgments; if multi-source and voice-design systems no longer lead, or if MOS-ranked TTS suddenly matches QA ranking, the central transfer claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Low-resource SQA systems can be bootstrapped from existing text QA plus multilingual TTS without waiting for large human-recorded question corpora.
  • TTS selection for training data should be driven by downstream QA metrics, not by NISQA or DNSMOS alone.
  • Multi-source and controllable voice-design synthesis improve robustness across real speakers more than single-voice cloning or high-MOS single engines.
  • Parameter-efficient speech-to-LLM connectors remain a practical route when both the encoder and the LLM stay frozen and only the projector and LoRA are trained.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same translate-then-synthesize recipe could be tried for other low-resource languages that already have text QA but lack spoken QA recordings, with the ranking of TTS sources re-checked per language.
  • If automatic judges and embedding similarity systematically over-credit paraphrases that human raters would reject, the reported gap between speech and text baselines may shrink or reverse under human evaluation.
  • Scaling synthetic speaker diversity beyond the engines used here is a natural next control: if gains plateau once voice variety saturates, the multi-source benefit is mainly anti-overfitting rather than raw data volume.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes LuxSQA, a workflow for low-resource Luxembourgish spoken question answering that builds training data from existing text QA resources (mainly SQuAD) via machine translation and multiple TTS systems, then trains a parameter-efficient SLAM-style model (frozen Whisper-large-v3-turbo encoder, linear projector, LoRA on frozen EuroLLM-9B or Apertus-8B). Single-source synthetic corpora (~48k questions each from MMS-TTS, Qwen-CV/VD, Omni-C/VD) are compared to a 4TTS multi-source mix (~230k) and to ASR-initialized variants. Evaluation on LLAMA-LB-Test (300 human-translated questions recorded by two native speakers) with Semantic Similarity and dual LLM-judge metrics finds that multi-source and voice-design configurations perform best, while no-reference TTS quality scores (NISQA, DNSMOS) do not monotonically predict downstream QA. The authors conclude that synthetic speech should be treated as task-specific training data rather than judged only by naturalness.

Significance. If the ranking of multi-source and voice-design TTS holds under better-controlled ablations, the work offers a practical, reproducible recipe for building SQA systems in low-resource languages without large human-recorded QA corpora. Strengths include a clean SLAM-style architecture that reuses strong frozen components, explicit comparison of several modern TTS engines (including voice-design modes), dual semantics-aware metrics that mitigate exact-match brittleness for a morphologically rich language, and an honest demonstration that MOS-like scores diverge from task utility. The Luxembourgish case study and the methodological emphasis on what works versus what does not are useful for other low-resource speech-LLM settings.

major comments (2)
  1. [II-C, II-E, III-B, Fig. 2, IV-A] Sections II-C, II-E, III-B and Fig. 2: the central claim that multi-source and voice-design configurations yield the strongest SQA performance confounds source/voice diversity with corpus size. Single-source sets are ~48k questions while the 4TTS mix is ~230k; Limitations (IV-A) explicitly note that data scale is not isolated. Without equal-size multi-source vs single-source (or multi-voice vs single-voice) ablations, the practical recipe 'prefer multi-source / voice-design over MOS' is not established; only a size-plus-source effect is shown. This is load-bearing for the abstract and conclusions.
  2. [II-C, II-F, III-B, IV-A] Sections II-C, II-F and III-B: transfer claims rest on a 300-item LLAMA-LB-Test set recorded by only two native speakers, scored by automatic SemSim and two LLM judges (no human evaluation of answers). The paper itself flags limited speaker diversity and LLM-judge limitations. With only two evaluation speakers, rankings of TTS sources and the claim that synthetic data compensates for missing human QA recordings remain fragile; at minimum the manuscript should report per-speaker variance more carefully and temper the transfer language until broader human-recorded evaluation is available.
minor comments (5)
  1. [II-D, Table II] Table II caption and surrounding text: Omni-VD is described as the only system for which NISQA was computed, yet the table reports NISQA for all five systems; reconcile the prose with the table.
  2. [Fig. 2, III] Fig. 2 is dense (many configurations, two metrics, two speakers, text baselines). Adding a compact numerical table of the same scores would improve readability and citation of exact numbers.
  3. [II-E] Clarify whether the ~230k 4TTS mix contains unique questions or repeated questions synthesized by different engines; this affects interpretation of the size confound.
  4. [II-B, III-A] Minor typos and inconsistencies: 'wo complementary' (should be 'two'), 'lB-adapted' / 'generic_lb', and occasional capitalization of V oiceDesign / CustomV oice.
  5. [II-C] The proprietary in-domain development set is mentioned but not quantified or released; a short description of its size and role would help reproducibility claims.

Circularity Check

0 steps flagged

No circularity: empirical TTS-config comparisons on held-out real-speaker SQA; outcomes not forced by construction or self-citation.

full rationale

LuxSQA is a data-centric empirical study. The load-bearing claims (multi-source / voice-design synthetic training yields strongest SQA on LLAMA-LB-Test Speakers 1–2; no-reference TTS MOS/DNSMOS do not monotonically predict downstream SemSim or LLM-Judge scores) are measured outcomes of training projector+LoRA models on different synthetic corpora and evaluating on held-out human-recorded Luxembourgish questions. Nothing in the architecture, loss, or metrics forces those rankings by definition: the same SLAM-style pipeline is trained under each TTS condition and scored independently. TTS quality scores (Table II) are reported separately and explicitly shown to diverge from QA rankings (Section III-C), so the paper does not redefine MOS as QA success. Self-citations (LuxASR as an ASR resource; SLAM-style projector recipe) supply components and prior methodology; they do not define or uniquely force the measured SQA ranking. Corpus-size confounding (48k vs ~230k) is a validity limitation the paper itself flags (IV-A), not circularity. No self-definitional equations, fitted-input-as-prediction, uniqueness import, or ansatz smuggling that collapses the central claim into its inputs. Score 0 is the correct honest finding.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

Empirical ML paper; load-bearing content is experimental protocol and frozen pretrained components rather than new formal axioms. Free parameters are standard training hyperparameters. Domain assumptions cover transfer of MT+TTS to real speech and adequacy of automatic QA metrics. No new physical or mathematical entities are postulated.

free parameters (3)
  • LoRA rank r and alpha = r=16, α=16
    Set to r=16, α=16 by hand; controls capacity of the only trainable LLM adapters and therefore how much the model can adapt to synthetic Luxembourgish speech.
  • speech embedding downsample factor k = 5
    Fixed at k=5 to match speech frames to LLM token space; choice affects information retained from Whisper encoder.
  • learning rate, warmup, epochs, batch size = 1e-4 / 1000 / 5 / 16
    Adam 1e-4, 1000-step warmup, 5 epochs, batch 16; standard but un-ablated choices that can change absolute SQA scores.
axioms (4)
  • domain assumption Frozen Whisper-large-v3-turbo encoder plus frozen multilingual LLM (EuroLLM-9B or Apertus-8B) with a linear projector and LoRA can map spoken questions to text answers when trained on paired synthetic speech and text answers.
    Core SLAM-style modeling assumption stated in Sections I and II-E; not proved, only trained and measured.
  • domain assumption Machine translation (TranslateGemma 12B) of English SQuAD questions into Luxembourgish preserves question meaning well enough for SQA training.
    Data pipeline in II-C; translation quality is not independently validated against human Luxembourgish questions beyond the final test set.
  • domain assumption Semantic Similarity (multilingual MiniLM cosine) and averaged GPT-4o/Gemini-2.5-Flash judge scores with a fixed partial-credit rubric are adequate proxies for Luxembourgish answer quality without human ratings.
    Metrics section II-F; authors note LLM-judge limitations but still rest rankings on these scores.
  • ad hoc to paper Two real native speakers on a 300-question set are sufficient to rank TTS training sources for transfer to real Luxembourgish speech.
    Evaluation design II-C/III; acknowledged as limited in IV-A but used as the sole real-speech test.

pith-pipeline@v1.1.0-grok45 · 16818 in / 3146 out tokens · 27869 ms · 2026-07-12T07:11:31.798400+00:00 · methodology

0 comments
read the original abstract

Spoken Question Answering (SQA) remains largely focused on high-resource languages and carefully recorded speech, limiting the reach of speech-LLM methods in low-resource settings. This paper investigates whether text-to-speech (TTS) can provide task-specific training data for Luxembourgish SQA without requiring a large human-recorded QA corpus. Starting from existing text-based QA resources, we translate questions into Luxembourgish, synthesize spoken questions with multiple TTS systems, and pair them with textual answers. We train a parameter-efficient SLAM-style architecture that connects a frozen Whisper encoder to frozen multilingual LLM backends through a learned projector and LoRA adapters. We compare MMS-TTS, Qwen3-TTS, and OmniVoice variants, including single-source corpora of about 48k questions and a 4TTS multi-source mix of approximately 230k questions. Evaluation on LLAMA-LB-Test with two real Luxembourgish speaker conditions shows that multi-source and voice-design-based synthetic training configurations yield the strongest SQA performance. The results also show that no-reference TTS quality scores do not monotonically predict downstream QA performance, indicating that synthetic speech must be evaluated as task-specific training data rather than only as natural-sounding audio.

Figures

Figures reproduced from arXiv: 2607.02763 by Alessio Brutti, Marco Matassoni, Nina Hosseini-Kivanani.

Figure 1
Figure 1. Figure 1: LuxSQA training pipeline with a frozen speech encoder, a trainable [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Speaker-conditioned SQA performance on LLAMA-LB-Test. Blue circles denote Eval-S1, orange triangles denote Eval-S2, and dashed vertical [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 6 linked inside Pith

  1. [1]

    End- to-end Spoken Conversational Question Answering: Task, Dataset and Model,

    C. You, N. Chen, F. Liu, S. Ge, X. Wu, and Y . Zou, “End- to-end Spoken Conversational Question Answering: Task, Dataset and Model,” inFindings of the Association for Computational Linguistics: NAACL 2022, M. Carpuat, M.-C. de Marneffe, and I. V . Meza Ruiz, Eds. Seattle, United States: Association for Computational Linguistics, Jul. 2022, pp. 1219–1232. ...

    2022

  2. [2]

    On the Impact of Speech Recognition Errors in Passage Retrieval for Spoken Question Answering,

    G. Sidiropoulos, S. Vakulenko, and E. Kanoulas, “On the Impact of Speech Recognition Errors in Passage Retrieval for Spoken Question Answering,” inProceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM ’22). ACM, 2022, pp. 17–21

    2022

  3. [3]

    SLM: Bridge the thin gap between speech and text foundation models,

    M. Wang, W. Han, I. Shafran, Z. Wu, C.-C. Chiu, Y . Cao, N. Chen, Y . Zhang, H. Soltau, P. K. Rubensteinet al., “SLM: Bridge the thin gap between speech and text foundation models,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  4. [4]

    SLAM-ASR: An Embarrassingly Sim- ple Approach for LLM with Strong ASR Capacity,

    Z. Ma, G. Yang, Y . Yang, Z. Gao, J. Wang, Z. Du, F. Yu, Q. Chen, S. Zheng, S. Zhanget al., “SLAM-ASR: An Embarrassingly Sim- ple Approach for LLM with Strong ASR Capacity,”arXiv preprint arXiv:2402.08846, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Speech Recognition Meets Large Language Model: Benchmark- ing, Models, and Exploration,

    ——, “Speech Recognition Meets Large Language Model: Benchmark- ing, Models, and Exploration,”Proc. AAAI, 2025

    2025

  6. [6]

    SLAM-LLM: A Modular, Open-Source Multimodal Large Language Model Framework and Best Practice for Speech, Language, Audio and Music Processing,

    Z. Ma, G. Yang, W. Chen, Z. Gao, Y . Du, X. Li, Z. Zheng, H. Zhu, J. Zhuo, Z. Songet al., “SLAM-LLM: A Modular, Open-Source Multimodal Large Language Model Framework and Best Practice for Speech, Language, Audio and Music Processing,”IEEE Journal of Selected Topics in Signal Processing, 2026

    2026

  7. [7]

    Performance evaluation of slam-asr: The good, the bad, the ugly, and the way forward,

    S. Kumar, I. Thorbecke, S. Burdisso, E. Villatoro-Tello, M. KE, K. Ha- cio˘glu, P. Rangappa, P. Motlicek, A. Ganapathiraju, and A. Stolcke, “Performance evaluation of slam-asr: The good, the bad, the ugly, and the way forward,” in2025 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 2025, pp. 1–5

    2025

  8. [8]

    Speech LLMs in Low-Resource Scenarios: Data V olume Requirements and the Impact of Pretraining on High-Resource Languages,

    S. Fong, M. Matassoni, and A. Brutti, “Speech LLMs in Low-Resource Scenarios: Data V olume Requirements and the Impact of Pretraining on High-Resource Languages,” inProc. Interspeech 2025, 2025, pp. 2003– 2007

    2025

  9. [9]

    Towards data distil- lation for end-to-end spoken conversational question answering,

    C. You, N. Chen, F. Liu, D. Yang, and Y . Zou, “Towards data distil- lation for end-to-end spoken conversational question answering,”arXiv preprint arXiv:2010.08923, 2020

    Pith/arXiv arXiv 2010

  10. [10]

    Dealing with data scarcity in spoken question answering,

    M. ¨U. Menevs ¸e, Y . Manav, E. Arisoy, and A. ¨Ozg¨ur, “Dealing with data scarcity in spoken question answering,” inProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024, pp. 4449–4455

    2024

  11. [11]

    A framework for automatic generation of spoken question- answering data,

    ——, “A framework for automatic generation of spoken question- answering data,” inFindings of the Association for Computational Linguistics: EMNLP 2022, 2022, pp. 4659–4666

    2022

  12. [12]

    HeySQuAD: A Spoken Question Answering Dataset,

    Y . Wu, S. Rallabandi, R. Srinivasamurthy, P. P. Dakle, A. Gon, and P. Raghavan, “HeySQuAD: A Spoken Question Answering Dataset,” 2023

    2023

  13. [13]

    LibriSQA: A novel dataset and framework for spoken question answering with large language models,

    Z. Zhao, Y . Jiang, H. Liu, Y . Wang, and Y . Wang, “LibriSQA: A novel dataset and framework for spoken question answering with large language models,”IEEE Transactions on Artificial Intelligence, 2024

    2024

  14. [14]

    Self-supervised dialogue learning for spoken conversational question answering,

    N. Chen, C. You, and Y . Zou, “Self-supervised dialogue learning for spoken conversational question answering,” inProceedings of INTER- SPEECH, 2021

    2021

  15. [15]

    SLAM-Omni: Timbre-Controllable V oice Interaction System with Single-Stage Training,

    W. Chen, Z. Ma, R. Yan, Y . Liang, X. Li, R. Xu, Z. Niu, Y . Zhu, Y . Yang, Z. Liu, K. Yu, Y . Hu, J. Li, Y . Lu, S. Liu, and X. Chen, “SLAM-Omni: Timbre-Controllable V oice Interaction System with Single-Stage Training,” inFindings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, Eds. Vienna,...

    2025

  16. [16]

    GSQA: An End-to-End Model for Generative Spoken Question Answering,

    M.-H. Shih, H.-L. Chung, Y .-C. Pai, M.-H. Hsu, G.-T. Lin, S.-W. Li, and H.-y. Lee, “GSQA: An End-to-End Model for Generative Spoken Question Answering,” inProc. Interspeech 2024, 2024, pp. 2970–2974

    2024

  17. [17]

    Spoken Question An- swering and Speech Continuation Using Spectrogram-Powered LLM,

    E. Nachmani, A. Levkovitch, R. Hirschet al., “Spoken Question An- swering and Speech Continuation Using Spectrogram-Powered LLM,” inProceedings of the International Conference on Learning Represen- tations (ICLR), 2024

    2024

  18. [18]

    Inserter: Speech instruction following with unsupervised interleaved pre-training,

    D. Wang, J. Xu, R. Chu, Z. Guo, X. Wang, J. Wu, D. Yang, S. Ji, and J. Lin, “Inserter: Speech instruction following with unsupervised interleaved pre-training,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025, pp. 18 024–18 046

    2025

  19. [19]

    Data Augmentation for Low-Resource Quechua ASR Improvement,

    R. Zevallos, N. Bel, G. C ´ambara, M. Farr ´us, and J. Luque, “Data Augmentation for Low-Resource Quechua ASR Improvement,” inProc. Interspeech 2022, 2022, pp. 3518–3522

    2022

  20. [20]

    TTS applied to the generation of datasets for automatic speech recognition,

    E. Casanova, S. Alu ´ısio, and M. A. Ponti, “TTS applied to the generation of datasets for automatic speech recognition,” inProceedings of the 16th International Conference on Computational Processing of Portuguese- Vol. 1, 2024, pp. 633–638

    2024

  21. [21]

    Improving data augmentation for low resource speech-to-text translation with diverse paraphrasing,

    C. Mi, L. Xie, and Y . Zhang, “Improving data augmentation for low resource speech-to-text translation with diverse paraphrasing,”Neural Networks, vol. 148, pp. 194–205, 2022

    2022

  22. [22]

    SpeechDPR: End-to-end spoken passage retrieval for open-domain spoken question answering,

    C.-J. Lin, G.-T. Lin, Y .-S. Chuang, W.-L. Wu, S.-W. Li, A. Mohamed, H.-y. Lee, and L.-s. Lee, “SpeechDPR: End-to-end spoken passage retrieval for open-domain spoken question answering,”arXiv preprint arXiv:2401.13463, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    LibriSQA: A Novel Dataset and Framework for Spoken Question Answering with Large Language Models,

    Z. Zhao, Y . Jiang, H. Liu, Y . Wang, and Y . Wang, “LibriSQA: A Novel Dataset and Framework for Spoken Question Answering with Large Language Models,”IEEE Transactions on Artificial Intelligence, 2023

    2023

  24. [24]

    Scaling speech technology to 1,000+ languages,

    V . Pratap, A. Tjandra, B. Shi, P. Tomasello, A. Babu, S. Kundu, A. Elkahky, Z. Ni, A. Vyas, M. Fazel-Zarandi, A. Baevski, Y . Adi, X. Zhang, W.-N. Hsu, A. Conneau, and M. Auli, “Scaling speech technology to 1,000+ languages,”J. Mach. Learn. Res., vol. 25, no. 1, Jan. 2024

    2024

  25. [25]

    Qwen3-tts technical report,

    H. Hu, X. Zhu, T. He, D. Guo, B. Zhang, X. Wang, Z. Guo, Z. Jiang, H. Hao, Z. Guo, X. Zhang, P. Zhang, B. Yang, J. Xu, J. Zhou, and J. Lin, “Qwen3-tts technical report,”arXiv preprint arXiv:2601.15621, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    OmniV oice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models,

    H. Zhu, L. Ye, W. Kang, Z. Yao, L. Guo, F. Kuang, Z. Han, W. Zhuang, L. Lin, and D. Povey, “OmniV oice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models,”arXiv preprint arXiv:2604.00688, 2026

    Pith/arXiv arXiv 2026

  27. [27]

    LUX-ASR: Building an ASR system for the Luxembourgish language,

    P. Gilles, N. Hosseini-Kivanani, and L. E. A. Hillah, “LUX-ASR: Building an ASR system for the Luxembourgish language,” in2022 IEEE Spoken Language Technology Workshop (SLT) SLT 2022, 2023

    2022

  28. [28]

    HeySQuAD: A Spoken Question Answering Dataset,

    Y . Wu, S. Rallabandi, R. Srinivasamurthy, P. P. Dakle, A. Gon, and P. Raghavan, “HeySQuAD: A Spoken Question Answering Dataset,” arXiv preprint arXiv:2304.13689, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Deep Learning Based Assessment of Synthetic Speech Naturalness,

    G. Mittag and S. M ¨oller, “Deep Learning Based Assessment of Synthetic Speech Naturalness,”Interspeech 2020, 2020

    2020

  30. [30]

    DNSMOS: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. K. Reddy, V . Gopal, and R. Cutler, “DNSMOS: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6493–6497

    2021

  31. [31]

    EuroLLM: Multilingual Language Models for Europe,

    P. H. Martinset al., “EuroLLM: Multilingual Language Models for Europe,”Procedia Computer Science, vol. 255, pp. 53–62, 2025, pro- ceedings of the Second EuroHPC user day

    2025

  32. [32]

    Apertus: Democratizing open and compliant llms for global language environments,

    A. Hern ´andez-Cano, A. H ¨agele, A. H. Huang, A. Romanou, A.-J. Solergibert, B. Pasztor, B. Messmer, D. Garbaya, E. F. ˇDurech, I. Hakimi et al., “Apertus: Democratizing open and compliant llms for global language environments,” inProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2026, pp. 46...

    2026

  33. [33]

    On the Problem of Text- To-Speech Model Selection for Synthetic Data Generation in Automatic Speech Recognition,

    N. Rossenbach, S. Sakti, and R. Schl ¨uter, “On the Problem of Text- To-Speech Model Selection for Synthetic Data Generation in Automatic Speech Recognition,” inProc. SynData4GenAI 2024, 2024, pp. 21–25

    2024

  34. [34]

    Judging llm-as-a-judge with mt-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xinget al., “Judging llm-as-a-judge with mt-bench and chatbot arena,”Advances in neural information processing systems, vol. 36, pp. 46 595–46 623, 2023

    2023

  35. [35]

    G-eval: Nlg evaluation using gpt-4 with better human alignment,

    Y . Liu, D. Iter, Y . Xu, S. Wang, R. Xu, and C. Zhu, “G-eval: Nlg evaluation using gpt-4 with better human alignment,” inProceedings of the 2023 conference on empirical methods in natural language processing, 2023, pp. 2511–2522

    2023