Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Towards Improved Speech Recognition through Optimized Synthetic Data Generation

T0 review · 3 major / 5 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 Optimized synthetic speech brings ASR to 20.5% word error on real audio

desk verdict Useful, honest ASR-synthetic-data study for Québec French, but the headline claim is only partially supported because finetuning and filtering are never ablated in the final ASR evaluation. read the letter →

arxiv 2508.21631 v1 pith:JUTVWQOG submitted 2025-08-29 eess.AS

Yanis Perrin , Gilles Boulianne This is my paper · ORCID
classification eess.AS
keywords syntheticspeechASRtrainingQuébecFrenchvoicecloningTTSdatafilteringWhisperverifierlow-resourcerecognitionspontaneous
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that synthetic speech can substitute for most transcribed audio when training a speech recognizer, if synthetic generation is deliberately optimized. Using a Québec French conversational domain, it fine-tunes a voice-cloning text-to-speech model on out-of-domain audio, filters generated utterances with a Whisper-based generator-verifier loop, and tunes generation temperature for training value rather than audio quality. The result is a from-scratch transformer that reaches 20.5% WER on real test audio when trained on 710 hours of synthetic plus 60 hours of real speech, approaching the 14.2% achieved with the full real corpus. The central claim is that data-generation improvements, not just synthetic-data quantity, drive these gains.

What carries the argument

The generator-verifier filter: a TTS model (XTTS-V2) generates each utterance; a pretrained ASR (Whisper large-v3) transcribes it; if the WER against the input text exceeds a threshold (0.20), the sample is regenerated, up to ten attempts. This loop is the mechanism that removes hallucinations, disfluencies, and content mismatches from the synthetic corpus. The other two mechanisms are fine-tuning XTTS-V2 on out-of-domain Québec French (12 hours of Charb) and setting generation temperature to 0.65, which accepts lower per-utterance quality in exchange for prosodic and acoustic diversity that trains a better recognizer.

What would settle it

Train the same from-scratch ASR on two equal-size synthetic corpora, one passed through the Whisper generator-verifier filter and one left unfiltered. If the unfiltered corpus gives equal or lower WER on real test audio, the filter's rejection criterion is removing useful training signal rather than artifacts. A complementary check: run Whisper on the rejected utterances and measure whether their transcriptions are systematically disfluent or dialect-heavy; if so, the filter selects against the spontaneous patterns the target ASR needs.

Watch

Extended reading notes

Core claim

The central claim is that optimizing the TTS pipeline improves the ASR model trained on its output: fine-tuning XTTS-V2 on out-of-domain Québec French speech, filtering each generated utterance by asking Whisper large-v3 to transcribe it and rejecting outputs above a WER threshold, and generating at temperature 0.65 instead of 0.1 all contribute to a better final recognizer. The strongest evidence is the hybrid experiments: with only 10 hours of real audio, adding 710 hours of filtered synthetic audio yields 22.8% test WER versus 25.2% with 350 hours, and with 60 real hours the 710/60 mix reaches 20.5%. The paper also shows that purely synthetic training beats a 10-hour real baseline, and th

Load-bearing premise

The filtering step assumes that a strong ASR's inability to transcribe a synthetic utterance marks it as useless for ASR training; if Whisper instead rejects the very disfluencies and dialectal patterns the target model must learn, the filter could be discarding the most instructive examples.

Editorial extensions

If this is right

  • In data-scarce domains with 10 hours or less of real audio, a large filtered synthetic corpus gives large WER reductions compared to using only the real audio.
  • The benefit of synthetic data shrinks as real data grows: moving from 10 to 60 real hours narrows the gain from adding 710 synthetic hours to under one point absolute.
  • LLM-generated transcripts are far from the target text distribution when used alone, but adding a small real-audio seed makes them competitive with text from the actual domain.
  • For pretrained models, hybrid synthetic-plus-real fine-tuning improves over synthetic alone, yet a small amount of real data still outperforms a much larger synthetic set, suggesting synthetic data can partly interfere with pretrained representations.

Reading between the lines

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

  • If the generator-verifier filter is doing what the paper thinks, then Whisper's own recognition biases become a bottleneck: speech patterns Whisper transcribes poorly, such as hesitations, dialectal variants, and disfluencies, would be systematically removed, making the synthetic corpus cleaner but less representative of spontaneous conversation.
  • The temperature sweet spot at 0.65 hints at an inverse relationship between per-utterance naturalness and corpus-level diversity; an adaptive temperature schedule per speaker or per utterance could yield both.
  • The method generalizes into a recipe for other low-resource dialects: fine-tune a voice-cloning TTS on any out-of-domain speech of the same language, filter with a strong multilingual ASR, then mix with a small real seed.
  • The LLM-text result suggests text-domain mismatch can be largely overcome by a real-audio seed, which could be tested further by combining domain-adaptive text generation with larger synthetic corpora.
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. The paper investigates whether synthetic speech generated by a voice-cloning TTS model can be optimized to train ASR systems for spontaneous, domain-specific Québec French speech. The authors finetune XTTS-V2 on out-of-domain Quebec French data, filter synthetic samples using a Whisper large-v3 generator-verifier scheme, vary generation temperature, and synthesize text with LLMs when real transcripts are unavailable. They evaluate two ASR setups: a from-scratch transformer trained on synthetic, hybrid, and real corpora, and LoRA finetuning of Whisper-medium. The central claim is that improving data generation leads to large improvements in the final ASR system; the strongest supported evidence is the temperature experiment (Table 10) and hybrid composition results (Table 4), where a from-scratch transformer reaches 20.5% test WER with 710h synthetic + 60h real audio versus 14.2% with all real data.

Significance. The paper addresses an important practical problem: training ASR under data confidentiality constraints using synthetic data. The dataset (CommissionsQC, Quebec French spontaneous speech) is nontrivial and underrepresented, and the from-scratch transformer results are internally consistent and clearly reported. The temperature experiment is a valuable controlled demonstration that generation parameters affect final ASR utility in a way that can oppose proxy quality metrics. The hybrid-data scaling behavior (Table 4) and the LLM-text simulation (Table 5) are also informative. If the missing ablations are added, the paper would provide a credible recipe for synthetic-data ASR training in low-resource domains. However, as it stands, the headline claim is only partially supported because two of the three main optimization components—finetuning and filtering—are validated solely on proxy metrics that the paper itself shows can be inversely related to final ASR performance.

major comments (3)
  1. [Section 3.2, Table 2, Figure 2] The generator-verifier filtering is a load-bearing component of the optimized pipeline, yet it is evaluated only on Eval-ASR WER and utMOS. No experiment trains the final ASR on filtered versus unfiltered synthetic data. This matters because Section 3.3 and Table 10 demonstrate that proxy metrics (%WERR, utMOS) can improve while final ASR performance degrades: T=0.1 yields better proxy quality but the ASR trained on it performs markedly worse on real speech than T=0.65. Since all Table 4 results use the fully filtered pipeline, the filter's contribution to the reported WER is never measured. The concern is not stylistic: if the Whisper verifier preferentially rejects disfluent or spontaneous patterns that the ASR needs to learn, the filter could remove the most informative examples while improving proxy scores. I request a matched comparison of final ASR WER with and without G-V filterin
  2. [Section 3.1, Table 7] The finetuning of XTTS-V2 is selected on the basis of Eval-ASR WER and utMOS (Table 7). This hyperparameter choice is never connected to the final ASR. Given the temperature result shows that better proxy quality does not guarantee better ASR training data, the same concern applies: the finetuned TTS might improve naturalness while reducing acoustic diversity in ways that hurt downstream ASR. A minimal experiment would train the from-scratch transformer on synthetic data generated by the initial XTTS versus the finetuned XTTS (keeping all other pipeline components fixed). Without this, the contribution of finetuning to the headline claim is unsupported.
  3. [Section 4.3, Table 6; Section 3.2] The Whisper finetuning experiments introduce a matched-filter circularity: the verifier used to filter synthetic data is Whisper large-v3, and the model being finetuned is Whisper-medium. If the verifier's acceptance criteria are correlated with Whisper-family recognition behavior, the filtered data may be biased toward what Whisper already transcribes well, inflating the apparent benefit of synthetic data in Table 6. The from-scratch transformer results (Table 4) are not affected by this, but the paper's broader claim about 'different model architectures' should be qualified. I recommend either finetuning a non-Whisper model on the filtered data, or reporting the finetuning results with a verifier that is not Whisper-based.
minor comments (5)
  1. [Table 1 and throughout] The table header uses '%WERR' while the body lists WER percentages; %WERR is used inconsistently (e.g., Table 3). Please clarify whether values are WER or relative WER reduction, and use a consistent notation.
  2. [Table 10] The columns labeled 'dev F', 'devM', 'testF', 'testM' are ambiguous: 'dev F' likely means female development set, but 'devM' could be male or mixed. Please clarify the split naming, and note in the caption which subsets correspond to the aggregate dev/test values.
  3. [Section 4.1] When describing the synthetic corpus generation, the constraint 'each text/audio pair appears only once' is stated, but the matching of random Charb voices to Bast texts is only partially described. Please specify whether the random selection of reference utterances is seeded, as this affects reproducibility.
  4. [Section 6, Limitations] The limitation paragraph appropriately notes that only XTTS was used and that Voicebox was unavailable. It would also be useful to acknowledge directly that finetuning and filtering were not validated on the final ASR metric, since this is a key limitation given the paper's title and abstract.
  5. [Appendix A] The LLM prompts are provided in French, which is helpful, but the paper does not state the model version or decoding parameters for GPT-4o-mini beyond 'optimized for faster inference'. A brief note on temperature or max tokens would aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central synthetic-data ASR results are evaluated on real speech, and the paper explicitly shows that its quality proxy can diverge from final ASR performance.

full rationale

The paper's main claim is that optimizing synthetic data generation improves a final ASR system trained on that data. The evaluation chain is not circular: synthetic data are generated with XTTS, filtered with a Whisper-based verifier, and used to train either a from-scratch transformer or a finetuned Whisper-medium, with final WER measured on real Bast and Charb speech. The optimization in Section 3 uses proxy metrics (Eval ASR WER and utMOS), but Section 3.3 directly contradicts a circular interpretation: lower temperature improves the proxy (lower Eval ASR WER, higher utMOS) yet yields a worse final ASR (Table 10), and the authors adopt temperature 0.65 based on final ASR performance rather than on the proxy. This shows the proxy is not definitionally identical to the final objective. The G-V filter uses Whisper large-v3 to reject synthetic utterances that Whisper cannot transcribe; this creates a potential matched-filter bias in the Whisper-medium finetuning experiment, but the paper does not claim that filtering alone improves the final ASR, and no equation or definition makes the final WER equal to the filter's criterion. The central from-scratch transformer results (Table 4) are self-contained comparisons against a real-data baseline. Hyperparameters and thresholds are chosen on synthetic development/test sets, not fitted to the final real test labels. Self-citations are limited to the dataset reference (Serrand et al., 2025), which is an external resource, not a load-bearing theoretical premise. The acknowledged limitation in Section 6 is about TTS model diversity, not about a circular argument. Overall, the derivation chain is not circular; any missing control (e.g., isolating the G-V filter's effect on the final ASR) is an experimental omission, not a circular reduction.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central pipeline depends on several empirically tuned hyperparameters (TTS learning rate and steps, G-V threshold, temperature, duration bounds) and on domain assumptions about the validity of Whisper as a verifier, the representativeness of CommissionsQC, and the Eval-ASR proxy. No new entities are introduced.

free parameters (7)
  • G-V filtering WER threshold = 0.20
    Selected based on dev/test WER improvements (Table 2); threshold of 0.20 gives the largest reported reduction and appears to be the chosen operating point.
  • Generation temperature = 0.65
    Temperature 0.65 chosen for training sets because it gave better downstream ASR than 0.1 despite lower quality metrics (Section 3.3, Table 10).
  • TTS finetuning learning rate = 5e-6
    Selected via hyperparameter sweep (Table 7) based on lowest WER and highest utMOS on both dev and test; test set was used in selection.
  • TTS finetuning steps = 60k
    Selected alongside learning rate in the same sweep; 60k steps at 5e-6 was optimal by reported criteria.
  • Duration filtering ratio bounds = 0.7/1.5 or 0.8/1.2
    Empirically chosen thresholds for duration ratio filter; only marginal WER improvements (Table 8).
  • Extreme duration/rate trimming in TTS finetuning = 5-15s; 10% tails removed
    Preprocessing choices for finetuning data, reported in Section 3.1.
  • Whisper model size for verifier = large-v3
    Chosen because medium size had hallucinations on QF; model choice affects filtering behavior.
assumptions (6)
  • domain assumption XTTS-V2 finetuning recipe (Coqui) produces a functional QF TTS
    Section 3.1; the paper relies on the public recipe's default hyperparameters as a valid baseline.
  • domain assumption Whisper large-v3 transcriptions are a reliable ground truth for synthetic speech verification
    Section 3.2; the filter assumes low Whisper WER implies clean, content-faithful synthetic audio.
  • domain assumption CommissionsQC corpora are representative of spontaneous QF conversational speech
    Section 2.1, citing Serrand et al. 2025; the evaluation and training rely on the corpus's acoustic and dialectal properties.
  • domain assumption The Eval ASR (trained on real Bast) is a valid proxy for synthetic data quality
    Section 3; used to optimize TTS, but never validated against downstream ASR for finetuning/filtering steps.
  • domain assumption Standard ESPnet transformer baseline with BPE tokenization transfers to QF
    Section 4.1; training config adapted from LibriSpeech_100 recipe.
  • domain assumption No leakage between synthetic train and real test through shared transcription text
    Section 4.1; synthetic train uses Bast train transcripts, real test is Bast test, but the same TTS was tuned on test transcripts, creating indirect leakage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Improved Speech Recognition through Optimized Synthetic Data Generation." pith.science (2026). https://pith.science/paper/JUTVWQOG

@misc{pith2026250821631,
  author       = {Pith},
  title        = {Pith review of: Towards Improved Speech Recognition through Optimized Synthetic Data Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUTVWQOG}},
  note         = {Machine review of arXiv:2508.21631}
}
read the original abstract

Supervised training of speech recognition models requires access to transcribed audio data, which often is not possible due to confidentiality issues. Our approach to this problem is to generate synthetic audio from a text-only corpus using a state-of-the-art text-to-speech model with voice cloning capabilities. Our goal is to achieve automatic speech recognition (ASR) performance comparable to models trained on real data. We explore ways to optimize synthetic data generation through finetuning, filtering and evaluation, and its use for training an end-to-end encoder-decoder ASR model. Experiments were conducted using two datasets of spontaneous, conversational speech in Qu\'ebec French. We show that improving data generation leads to large improvements in the final ASR system trained on synthetic data.

Figures

Figures reproduced from arXiv: 2508.21631 by the authors.

Figure 1
Figure 1. Top: Phase 1. Optimization of synthetic quality. Bottom: Phase 2. Speech recognition evaluation. Red indicates target domain data, blue out-of-domain data, and green, LLM-generated data. 2.1. Datasets While most ASR research on synthetic speech uses widely-available English-language bench￾marks such as LibriSpeech (Panayotov et al., 2015), our work focuses on QF speech. We con￾duct our experiments on CommissionsQC (… view at source ↗
Figure 2
Figure 2. Generator-Verifier based filtering [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Synthetic text generation pipeline We initially used GPT-4o to generate a base prompt, tailored to our predefined requirements. This prompt was then provided to GPT-4o-mini to produce the first batch of synthetic texts. Upon analyzing these initial outputs, we identified limi￾tations in relevance, which led us to develop five targeted sub-prompts to guide the generation more effectively. The process was restarted us… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. When Synthetic Speech Is All You Have: Better Call GRPO

    cs.CL 2026-07 conditional novelty 6.0 of 10

    On synthetic banking speech alone, GRPO cuts ASR WER 40% relative to SFT (36.71%→22.09%) by improving stopping calibration and attention anchoring to audio.

  2. How to Leverage Synthetic Speech for LLM-Based ASR Systems?

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    Layer-wise pooling plus RIR-augmented synthetic speech matches a 100%-real ASR baseline with only 25% real data (13.6 h) and beats it at higher real fractions.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages · cited by 2 Pith papers

  1. [1]

    Introduction Automatic Speech Recognition (ASR) systems have achieved remarkable performance in recent years, but this success has largely been contingent on access to vast amounts of transcribed audio data for training (Radford et al., 2023a; Zhang et al., 2023; Pratap et al., 2023). In specialized domains such as banking, healthcare and call centers, th...

    arXiv 2023

  2. [2]

    The smaller one is con- sidered as the target domain and the largest as out of domain

    Methodology To address this challenge, we use two real-world datasets containing spontaneous speech in QF , each providing both manual transcriptions and ref- erence audio recordings. The smaller one is con- sidered as the target domain and the largest as out of domain. Figure 1 illustrates the main phases of our methodology. In phase 1, the synthetic spe...

    work page 2022

  3. [3]

    In our experiments, the Bast corpus is used as the target domain for synthetic replication due to its relatively limited size, which enables more 1https://github.com/coqui-ai/TTS

    Optimization of synthetic quality Our objective is to create synthetic utterances that closely approximate the acoustic and linguistic characteristics of our authentic QF speech cor- pora. In our experiments, the Bast corpus is used as the target domain for synthetic replication due to its relatively limited size, which enables more 1https://github.com/co...

    work page 2024

  4. [4]

    best of the bad

    Speech recognition evaluation In order to evaluate the generated synthetic data usefulness to train a speech recognition model, we generate multiple training datasets varying in size and composition. 4.1. Training data generation We create purely synthetic datasets of 90 and 360 hours to assess the effect of synthetic dataset size. For evaluating the impa...

    work page 2018

  5. [5]

    We proposed methods to improve synthetic speech quality and demonstrated that optimizing the generation pro- cesses enhanced synthetic data quality

    Conclusion We explored the effectiveness of generating syn- thetic speech datasets for ASR training, using data from two Québec commissions, across multiple scenarios of varying data availability, including a simulation with no initial real data. We proposed methods to improve synthetic speech quality and demonstrated that optimizing the generation pro- c...

    work page

  6. [6]

    Limitations In this study, we used XTTS as our only synthesis model, which may have limited the diversity of the resulting generated data. We considered conduct- ing experiments using the model proposed in (Le et al., 2023) due to its promising performance on LibriSpeech, but its unavailability to the research community precluded its use in our study. Inv...

    work page 2023

  7. [7]

    Ethical Considerations The approach we propose involves cloning voices from audio samples, raising ethical considerations regarding data privacy and consent. We acknowl- edge the importance of ensuring informed consent from all individuals whose voice samples are used for model training, particularly given the potential applications in sensitive domains d...

    work page

  8. [8]

    Acknowledgements We acknowledge the support of the Natural Sci- ences and Engineering Research Council of Canada (NSERC) for this work, and would also like to thank Ministry of Economy and Innovation (MEI) of the Government of Québec for its contin- ued support

    work page

Show all 45 references
  1. [9]

    Bibliographical References Alëna Aksënova, Zhehuai Chen, Chung-Cheng Chiu, Daan van Esch, Pavel Golik, et al

  2. [10]

    Mon- sieur

    Respectez les conventions d’écriture : • Évitez les abréviations (ex. “Mon- sieur” au lieu de “M.”) • Écrivez les nombres en toutes let- tres • Maintenez une mise en page claire Exemple de format à respecter : Président : “Quelle est votre opinion sur l’indépendance ju- diciai...

  3. [15]

    Créez des dialogues réalistes et cen- trés sur le sujet principal, permettant de couvrir différentes dimensions du processus de nomination judiciaire

  4. [16]

    Intégrez des détails concrets dans les échanges, tels que des dates précises, années, numéros d’articles de loi, références de documents dé- posés, noms de villes, lieux, partis politiques, montants financiers, et procédures spécifiques

  5. [17]

    Variez les intervenants, incluant : • Le président de la commission • Des experts juridiques • Des témoins • Des citoyens ou parties prenantes

  6. [18]

    Indiquez seulement le prénom et le nom de chaque intervenant, suivi de leur texte dans la transcription

  7. [19]

    Assurez-vous que les échanges in- cluent aussi des réponses très cour- tes et naturelles de quelques mots, pour refléter des dialogues réalistes

  8. [20]

    Gardez les conversations perti- nentes au thème principal, tout en explorant des sous-thèmes comme : • Critères de qualification et éthique • Influence politique et externe • Transparence et responsabilité

  9. [21]

    Structurez le texte pour qu’il reflète un échange continu sans introduc- tions ni conclusions formelles

  10. [22]

    Adaptez le registre de langue selon les participants : • Un ton formel pour le président et les experts • Un langage plus spontané pour les témoins et citoyens

  11. [23]

    Intégrez des variations émotion- nelles et subtilités culturelles pour rendre les échanges vivants, tou- jours sous forme de dialogue

  12. [25]

    Continuez les dialogues en gardant bien à l’esprit que le sujet principal est le processus de nomination des juges au Québec

  13. [26]

    Maintenez le flux de la conversation, en développant progressivement les idées et les arguments de manière approfondie avec un vocabulaire ap- proprié à une commission d’enquête

  14. [27]

    Continuez de structurer le texte pour qu’il reflète un échange continu sans introductions ni conclusions, typique d’une commission d’enquête

  15. [28]

    Produisez une variété de dialogues, alternant entre des échanges courts et d’autres plus approfondis

  16. [29]

    Explorez de nouveaux aspects du su- jet si le précédent est épuisé, et intro- duisez de nouveaux participants au besoin, sans forcer des conclusions précipitées. A.3. Development of Facts Objective: Promotes the inclusion of specific fac- tual elements such as dates, places, p...

  17. [30]

    Continuez le dialogue en vous con- centrant sur le développement de faits passés, en questionnant les té- moins et les experts à ce sujet

  18. [31]

    Concentrez-vous sur la génération de témoignages détaillés et de dia- logues investigatifs, en explorant des sujets spécifiques avec des faits con- crets

  19. [32]

    Formulez des questions précises sur des événements antérieurs liés au processus de nomination des juges, sollicitant des détails et clarifications

  20. [33]

    Maintenez un ton investigatif, s’assurant que la conversation demeure engagée et directement liée aux faits examinés. A.4. Familiar Language Objective: Guides the model to produce utter- ances in informal Quebec French, capturing re- gional linguistic nuances. Prompt Content: ...

  21. [34]

    Continuez le dialogue en intégrant des échanges plus spontanés et fam- iliers, surtout pour les témoins et les citoyens

  22. [35]

    Utilisez un langage qui reflète les nu- ances et le ton informel du français québécois, en vous éloignant d’un registre trop formel

  23. [36]

    Encouragez les participants à répon- dre avec des expressions courantes et des réactions naturelles pour ren- dre les interactions plus authen- tiques

  24. [37]

    Maintenez le focus général sur le processus de nomination des juges, mais avec une approche plus décon- tractée de la conversation. A.5. Short Answers Objective: Constrains the model to generate concise responses, preventing overly long or ver- bose texts. Prompt Content: Inst...

  25. [38]

    Continuez le dialogue en encour- ageant des réponses très courtes et des échanges rapides entre les par- ticipants incluant des réponses sim- ples de quelques mots

  26. [39]

    Introduisez de nouvelles voix ou per- spectives si nécessaire, mais main- tenez un échange très rapide et con- cis

  27. [40]

    Assurez-vous que le ton et le style restent conformes au français québé- cois

  28. [41]

    Assurez-vous que les interactions restent centrées sur le processus de nomination des juges au Québec, tout en intégrant divers points de vue. A.6. Exploration of Unrelated Topics Objective: Orient the model to explore other more distant subjects but which remain linked to the...

  29. [42]

    Poursuivez le dialogue en explorant des sujets plus éloignés mais tou- jours liés au processus de nomina- tion des juges

  30. [43]

    Continuez de structurer le texte pour qu’il reflète un échange continu sans introductions ni conclusions

  31. [44]

    Produisez cette fois des échanges plus long tout en alternant avec quelques échanges courts

  32. [45]

    Maintenez un style engageant et in- formatif tout en explorant ces nou- velles dimensions. B. Tables %WERR utMOS LR Steps dev test dev test None None 30.8 29.4 2.04 2.01 1e-06 15k 29.7 28.5 2.15 2.11 1e-06 60k 24.1 24.0 2.20 2.14 1e-06 82k 23.1 22.6 2.23 2.15 5e-06 15k 24.2 25...

  33. [2016]

    ArXiv:1508.07909

    Neural Machine Translation of Rare Words with Subword Units. ArXiv:1508.07909. Coralie Serrand, Amira Morsli, and Gilles Bou- lianne. 2025. CommissionsQC: a Québec French speech corpus for automatic speech recognition. In Proc. Interspeech, pages 3918– 3922. Changhan Wang, Mor...

    2025 arXiv

  34. [2017]

    We used a warm-up strategy with 9,000 steps to gradually increase the learning rate to 4e-03

    with weight decay regularization (coefficient: 1e-06). We used a warm-up strategy with 9,000 steps to gradually increase the learning rate to 4e-03. The loss function used a weighted combi- nation of CTC loss (0.3) and label smoothed cross- entropy loss (0.7) ((Boyer et al., 2...

    2021

  35. [2021]

    VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation. In Proc. ACL, pages 993–1003. Shinji Watanabe, Takaaki Hori, Shigeki Karita, Tomoki Hayashi, Jiro Nishitoba, et al. 2018. ES- Pnet: End-to-End Speech P...

    2018 arXiv

  36. [2022]

    ArXiv:2205.08014 [eess]

    Accented Speech Recognition: Bench- marking, Pre-training, and Diverse Data. ArXiv:2205.08014 [eess]. Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, et al. 2020. Common voice: A massively-multilingual speech corpus. In Proc. LREC, pages 4218–4222....

    2020 arXiv

  37. [2023]

    Journal of Machine Learning Re- search, 25(97):1–52

    Scaling Speech Technology to 1,000+ Languages. Journal of Machine Learning Re- search, 25(97):1–52. Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. MLS: A large-scale multilingual dataset for speech research. In Proc. Interspeech, pages...

    2020

Pith tools

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