pith. sign in

arxiv: 2509.14804 · v1 · submitted 2025-09-18 · 💻 cs.SD · eess.AS

Towards Building Speech Large Language Models for Multitask Understanding in Low-Resource Languages

Mingchen Shao , Bingshen Mu , Chengyou Wang , Hai Li , Ying Yan , Zhonghua Fu , Lei Xie This is my paper

Pith reviewed 2026-05-18 16:22 UTC · model grok-4.3

classification 💻 cs.SD eess.AS
keywords speech large language modelslow-resource languagesThaiself-supervised speech encoderspeech-text alignmentspoken language understandingdata generation pipeline
0
0 comments X

The pith

A Thai-specific speech encoder trained on 36,000 hours, an efficient U-Align method, and a pipeline generating over 1,000 hours of Thai data together enable effective multitask speech understanding in a low-resource language.

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

Speech large language models perform well on multitask understanding for high-resource languages but degrade sharply for Thai because general encoders underperform, full ASR alignment is expensive, and paired data is scarce. The paper introduces XLSR-Thai by continuing self-supervised training of a standard SSL model on a large Thai speech collection. It pairs this with U-Align, which aligns speech and text for multiple tasks more efficiently than ASR methods, and Thai-SUP, which synthesizes a large spoken language understanding dataset from high-resource sources. These three pieces allow construction of a Thai SLLM that handles several understanding tasks. A reader would care because the same pattern could extend speech AI capabilities to many other languages that lack abundant labeled resources.

Core claim

The authors claim that XLSR-Thai, created by continued self-supervised pretraining on 36,000 hours of Thai speech, combined with the U-Align alignment procedure and the Thai-SUP data generation pipeline that produces more than 1,000 hours of Thai spoken language understanding examples, produces an SLLM capable of multitask understanding in Thai, as verified across multiple experiments.

What carries the argument

XLSR-Thai encoder together with U-Align for speech-text alignment and the Thai-SUP synthesis pipeline

If this is right

  • The resulting model supports multiple spoken language understanding tasks in Thai without the full retraining cost of ASR-based alignment.
  • Training becomes more computationally efficient than conventional approaches that require updating the entire SLLM for alignment.
  • Open-sourcing XLSR-Thai and the Thai-SUP dataset allows other groups to reproduce or extend the work for Thai and similar languages.
  • Performance on Thai tasks improves over models that rely on general encoders or limited real paired data.

Where Pith is reading between the lines

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

  • The same continued-pretraining and synthesis strategy could be tested on other low-resource languages that possess large unlabeled speech collections but few understanding labels.
  • The generated data pipeline might be adapted to create additional task types or to support languages with even smaller unlabeled speech resources.
  • Combining the resulting Thai SLLM with larger base LLMs or different adapter designs could be explored as a next step to increase capability further.

Load-bearing premise

The synthetic Thai-SUP dataset supplies enough quality and task variety that training on it does not introduce biases or noise that would reduce real performance.

What would settle it

A controlled experiment that trains an SLLM with the proposed components and finds no gain on Thai multitask benchmarks relative to a standard Whisper-based baseline would disprove the effectiveness of the methods.

read the original abstract

Speech large language models (SLLMs) built on speech encoders, adapters, and LLMs demonstrate remarkable multitask understanding performance in high-resource languages such as English and Chinese. However, their effectiveness substantially degrades in low-resource languages such as Thai. This limitation arises from three factors: (1) existing commonly used speech encoders, like the Whisper family, underperform in low-resource languages and lack support for broader spoken language understanding tasks; (2) the ASR-based alignment paradigm requires training the entire SLLM, leading to high computational cost; (3) paired speech-text data in low-resource languages is scarce. To overcome these challenges in the low-resource language Thai, we introduce XLSR-Thai, the first self-supervised learning (SSL) speech encoder for Thai. It is obtained by continuously training the standard SSL XLSR model on 36,000 hours of Thai speech data. Furthermore, we propose U-Align, a speech-text alignment method that is more resource-efficient and multitask-effective than typical ASR-based alignment. Finally, we present Thai-SUP, a pipeline for generating Thai spoken language understanding data from high-resource languages, yielding the first Thai spoken language understanding dataset of over 1,000 hours. Multiple experiments demonstrate the effectiveness of our methods in building a Thai multitask-understanding SLLM. We open-source XLSR-Thai and Thai-SUP to facilitate future research.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper addresses challenges in building speech large language models (SLLMs) for multitask understanding in low-resource languages like Thai. It introduces XLSR-Thai, a self-supervised speech encoder obtained by continued pretraining of XLSR on 36,000 hours of Thai speech data; U-Align, a resource-efficient speech-text alignment method that avoids full SLLM retraining; and Thai-SUP, a pipeline that generates over 1,000 hours of Thai spoken language understanding data by transforming high-resource language sources. The central claim is that multiple experiments demonstrate the effectiveness of these components in constructing a Thai multitask-understanding SLLM, with the resources open-sourced to support future work.

Significance. If the empirical results hold under rigorous validation, the work provides a practical pathway for adapting SLLMs to low-resource languages where standard encoders like Whisper underperform. The open-sourcing of XLSR-Thai and the Thai-SUP dataset constitutes a concrete contribution that could enable reproducible extensions to other languages. The U-Align approach offers a potential efficiency gain over ASR-based alignment, which is relevant for computational constraints in low-resource settings. Significance is tempered by the need to confirm that synthetic data does not introduce biases that undermine generalization claims.

major comments (2)
  1. [Abstract] Abstract: The statement that 'Multiple experiments demonstrate the effectiveness of our methods' is load-bearing for the central claim yet provides no quantitative results, baselines, error bars, ablation studies, or specific metrics (e.g., accuracy on multitask benchmarks). This omission prevents assessment of whether gains exceed those from existing Whisper-style encoders on native Thai data.
  2. [Thai-SUP pipeline] Thai-SUP pipeline (described in the methods section following the introduction of XLSR-Thai and U-Align): The claim that the generated >1,000-hour dataset supports effective multitask training rests on the unverified assumption that transforming high-resource language data preserves Thai-specific acoustic, prosodic, and semantic distributions. Without reported validation such as phonetic coverage statistics, human quality ratings, or performance comparison against real Thai speech, the multitask generalization results risk reflecting overfitting to synthetic artifacts rather than true low-resource capability.
minor comments (2)
  1. [Abstract] The abstract and introduction would benefit from explicit citation of the exact multitask tasks (e.g., ASR, intent classification, emotion recognition) evaluated in the experiments.
  2. [U-Align method] Notation for the alignment loss in the U-Align description should be clarified with respect to standard contrastive or reconstruction objectives to avoid ambiguity with prior adapter-based methods.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and constructive suggestions. We have carefully considered the major comments and provide point-by-point responses below. We will make revisions to the manuscript to address the concerns raised, particularly by enhancing the abstract and adding validation for the Thai-SUP pipeline.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The statement that 'Multiple experiments demonstrate the effectiveness of our methods' is load-bearing for the central claim yet provides no quantitative results, baselines, error bars, ablation studies, or specific metrics (e.g., accuracy on multitask benchmarks). This omission prevents assessment of whether gains exceed those from existing Whisper-style encoders on native Thai data.

    Authors: We acknowledge that the abstract, as currently written, does not include specific quantitative results. To better support the central claim and allow readers to assess the effectiveness immediately, we will revise the abstract to incorporate key metrics from our experiments, such as accuracy improvements on multitask benchmarks relative to Whisper-based baselines. The full set of results, including ablations and error bars, remains detailed in the body of the paper. revision: yes

  2. Referee: [Thai-SUP pipeline] Thai-SUP pipeline (described in the methods section following the introduction of XLSR-Thai and U-Align): The claim that the generated >1,000-hour dataset supports effective multitask training rests on the unverified assumption that transforming high-resource language data preserves Thai-specific acoustic, prosodic, and semantic distributions. Without reported validation such as phonetic coverage statistics, human quality ratings, or performance comparison against real Thai speech, the multitask generalization results risk reflecting overfitting to synthetic artifacts rather than true low-resource capability.

    Authors: We appreciate this insightful comment regarding the Thai-SUP pipeline. Our current experiments show that the SLLM trained with Thai-SUP data outperforms baselines in multitask understanding, providing indirect evidence of its utility. However, we agree that explicit validation of the synthetic data's alignment with Thai acoustic and semantic characteristics would strengthen the claims. In the revised manuscript, we will include phonetic coverage statistics, results from human quality assessments on sampled data, and additional experiments comparing performance on synthetic versus available real Thai speech data to mitigate concerns about potential biases or overfitting to artifacts. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical pipeline is self-contained

full rationale

The paper describes an empirical engineering workflow: continuous pretraining of XLSR on 36k hours of Thai speech to create XLSR-Thai, introduction of the U-Align alignment procedure, and construction of the Thai-SUP synthetic dataset by transforming high-resource data. Effectiveness is shown via downstream experiments rather than any derivation chain. No equations, fitted parameters renamed as predictions, self-definitional steps, or load-bearing self-citations appear in the provided text. The central claim rests on experimental outcomes against external baselines and does not reduce to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on the effectiveness of continued pretraining, the superiority of U-Align over ASR alignment, and the fidelity of synthetic data generation. No free parameters are explicitly fitted in the abstract. No new physical entities are postulated.

axioms (2)
  • domain assumption Continued self-supervised training on 36,000 hours of Thai speech improves encoder performance on downstream Thai tasks.
    Invoked when introducing XLSR-Thai as the solution to encoder underperformance.
  • domain assumption Synthetic data generated by Thai-SUP preserves task semantics and acoustic characteristics sufficiently for multitask training.
    Central to the claim that the resulting dataset enables effective Thai SLLM training.

pith-pipeline@v0.9.0 · 5801 in / 1307 out tokens · 39683 ms · 2026-05-18T16:22:58.979627+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

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

  1. Ti-Audio: The First Multi-Dialectal End-to-End Speech LLM for Tibetan

    cs.SD 2026-04 unverdicted novelty 7.0

    Ti-Audio is the first multi-dialectal end-to-end Speech-LLM for Tibetan that achieves state-of-the-art performance on ASR and speech translation benchmarks via a Dynamic Q-Former Adapter and cross-dialect cooperation.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · cited by 1 Pith paper · 7 internal anchors

  1. [1]

    Towards Building Speech Large Language Models for Multitask Understanding in Low-Resource Languages

    INTRODUCTION Large language models (LLMs) have demonstrated exceptional ca- pabilities in numerous natural language processing tasks, including text understanding, generation, and reasoning [1, 2, 3]. This ca- pability has promoted considerable development in speech LLMs (SLLMs), which extend the LLMs to process speech input directly. In particular, SLLMs...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    To extract rich speech represen- tations and support multitask requirements, we continue pretraining a multilingual SSL XLSR model on readily available unlabeled speech

    PROPOSED METHODS To develop SLLMs with strong multitask understanding capability in low-resource languages, we propose a comprehensive solution and take Thai as a representative case. To extract rich speech represen- tations and support multitask requirements, we continue pretraining a multilingual SSL XLSR model on readily available unlabeled speech. We ...

    work page

  3. [3]

    Data Collection High-resource text understanding data

    work page

  4. [4]

    Data Augmentation Deepseek

    work page

  5. [5]

    Text Translation Gemini

    work page

  6. [6]

    Giga2 Test

    Text to Speech Low-resource speech understanding data Fig. 2:Thai-SUP pipeline.Thai-SUP generates low-resource Thai spoken language understanding data from high-resource English text corpora using LLM-based data augmentation, translation, and TTS. 2.2.2. Universal speech-text alignment Traditional ASR-based alignment methods fine-tune the entire SLLM to o...

    work page 2023

  7. [7]

    EXPERIMENTS 3.1. Experimental setup We continue pretraining XLSR on 16,000 hours of public Thai data, including GigaSpeech2 [22] and MSR-86K [23], and 20,000 hours of in-house unlabeled Thai to obtain XLSR-Thai. To verify encoder gains, we fine-tune ASR on GigaSpeech2, MSR-86K, and Common V oice [24] using either XLSR-Thai or the original XLSR and re- por...

    work page 2000

  8. [8]

    CONCLUSION In this work, we propose a comprehensive solution for building multitask understanding SLLMs for low-resource languages. We leverage easily accessible unlabeled data for continuously pretrain- ing XLSR, and introduce U-Align to achieve more resource-efficient and multitask-effective speech-text alignment, and develop the Thai- SUP pipeline to t...

    work page

  9. [9]

    GPT-4 Technical Report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al., “GPT-4 Technical Report,”arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Am- jad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al., “Llama 2: Open Foundation and Fine-Tuned Chat Models,”arXiv preprint arXiv:2307.09288, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    Qwen2 Technical Report

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al., “Qwen2 Technical Report,”arXiv preprint arXiv:2407.10671, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Qwen2-Audio Technical Report

    Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhi- fang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al., “Qwen2-Audio Technical Report,”arXiv preprint arXiv:2407.10759, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    Kimi-Audio Technical Report

    Ding Ding, Zeqian Ju, Yichong Leng, Songxiang Liu, Tong Liu, Zeyu Shang, Kai Shen, Wei Song, Xu Tan, Heyi Tang, et al., “Kimi-Audio Technical Report,”arXiv preprint arXiv:2504.18425, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Baichuan-audio: A unified frame- work for end-to-end speech interaction

    Tianpeng Li, Jun Liu, Tao Zhang, Yuanbo Fang, Da Pan, Min- grui Wang, Zheng Liang, Zehuan Li, Mingan Lin, Guosheng Dong, et al., “Baichuan-Audio: A Unified Framework for End- to-End Speech Interaction,”arXiv preprint arXiv:2502.17239, 2025

    work page arXiv 2025

  15. [15]

    Enhancing Generalization of Speech Large Language Models with Multi-Task Behav- ior Imitation and Speech-Text Interleaving,

    Jingran Xie, Xiang Li, Hui Wang, Yue Yu, Yang Xiang, Xixin Wu, and Zhiyong Wu, “Enhancing Generalization of Speech Large Language Models with Multi-Task Behav- ior Imitation and Speech-Text Interleaving,”arXiv preprint arXiv:2505.18644, 2025

    work page arXiv 2025

  16. [16]

    V oxtral.arXiv preprint arXiv:2507.13264, 2025

    Alexander H. Liu, Andy Ehrenberg, Andy Lo, Cl ´ement De- noix, Corentin Barreau, Guillaume Lample, et al., “V oxtral,” arXiv preprint arXiv:2507.13264, 2025

    work page arXiv 2025

  17. [17]

    Speech Discrete Tokens or Continuous Features? A Comparative Analysis for Spo- ken Language Understanding in SpeechLLMs,

    Dingdong Wang, Junan Li, Mingyu Cui, Dongchao Yang, Xueyuan Chen, and Helen Meng, “Speech Discrete Tokens or Continuous Features? A Comparative Analysis for Spo- ken Language Understanding in SpeechLLMs,”arXiv preprint arXiv:2508.17863, 2025

    work page arXiv 2025

  18. [18]

    GLM-4-Voice: Towards Intelligent and Human-Like End-to-End Spoken Chatbot

    Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang, “GLM- 4-V oice: Towards Intelligent and Human-Like End-to-End Spoken Chatbot,”arXiv preprint arXiv:2412.02612, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

    Freeze-omni: A smart and low latency speech-to-speech dialogue model with frozen llm

    Xiong Wang, Yangze Li, Chaoyou Fu, Yunhang Shen, Lei Xie, Ke Li, Xing Sun, and Long Ma, “Freeze-Omni: A Smart and Low Latency Speech-to-speech Dialogue Model with Frozen LLM,”arXiv preprint arXiv:2411.00774, 2024

    work page arXiv 2024

  20. [20]

    TASTE: Text-Aligned Speech To- kenization and Embedding for Spoken Language Modeling,

    Liang-Hsuan Tseng, Yi-Chang Chen, Kuan-Yi Lee, Da-Shan Shiu, and Hung yi Lee, “TASTE: Text-Aligned Speech To- kenization and Embedding for Spoken Language Modeling,” arXiv preprint arXiv:2504.07053, 2025

    work page arXiv 2025

  21. [21]

    SALMONN: Towards Generic Hearing Abilities for Large Language Mod- els,

    Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang, “SALMONN: Towards Generic Hearing Abilities for Large Language Mod- els,” inProc. ICLR, 2024

    work page 2024

  22. [22]

    Unveiling the Potential of LLM-Based ASR on Chinese Open-Source Datasets,

    Xuelong Geng, Tianyi Xu, Kun Wei, Bingshen Mu, Hongfei Xue, He Wang, Yangze Li, Pengcheng Guo, Yuhang Dai, Longhao Li, et al., “Unveiling the Potential of LLM-Based ASR on Chinese Open-Source Datasets,” inProc. ISCSLP , 2024, pp. 26–30

    work page 2024

  23. [23]

    Efficient Scaling for LLM-based ASR,

    Bingshen Mu, Yiwen Shao, Kun Wei, Dong Yu, and Lei Xie, “Efficient Scaling for LLM-based ASR,”arXiv preprint arXiv:2508.04096, 2025

    work page arXiv 2025

  24. [24]

    Robust Speech Recognition via Large-Scale Weak Supervision,

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever, “Robust Speech Recognition via Large-Scale Weak Supervision,” inProc, ICML, 2023, pp. 28492–28518

    work page 2023

  25. [25]

    Weakly supervised data refinement and flexible sequence compression for efficient thai llm-based ASR,

    Mingchen Shao, Xinfa Zhu, Chengyou Wang, Bingshen Mu, Hai Li, Ying Yan, Junhui Liu, Danming Xie, and Lei Xie, “Weakly supervised data refinement and flexible sequence compression for efficient thai llm-based ASR,”arXiv preprint arXiv:2505.22063, 2025

    work page arXiv 2025

  26. [26]

    OSUM: Advancing Open Speech Understanding Models with Limited Resources in Academia,

    Xuelong Geng, Kun Wei, Qijie Shao, Shuiyun Liu, Zhennan Lin, Zhixian Zhao, Guojian Li, Wenjie Tian, Peikun Chen, Yangze Li, Pengcheng Guo, Mingchen Shao, Shuiyuan Wang, Yuang Cao, Chengyou Wang, Tianyi Xu, Yuhang Dai, Xinfa Zhu, Yue Li, Li Zhang, and Lei Xie, “OSUM: Advancing Open Speech Understanding Models with Limited Resources in Academia,”arXiv prepr...

    work page arXiv 2025

  27. [27]

    XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale,

    Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakho- tia, Qiantong Xu, Naman Goyal, Kritika Singh, et al., “XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale,” inProc. Interspeech, 2022, pp. 2278–2282

    work page 2022

  28. [28]

    Typhoon 2: A Family of Open Text and Multimodal Thai Large Lan- guage Models,

    Kunat Pipatanakul, Potsawee Manakul, Natapong Nitarach, Warit Sirichotedumrong, Surapon Nonesung, et al., “Typhoon 2: A Family of Open Text and Multimodal Thai Large Lan- guage Models,”arXiv preprint arXiv:2412.13702, 2024

    work page arXiv 2024

  29. [29]

    XEmoRAG: Cross-Lingual Emotion Transfer with Controllable Intensity Using Retrieval-Augmented Generation,

    Tianlun Zuo, Jingbin Hu, Yuke Li, Xinfa Zhu, Hai Li, Ying Yan, Junhui Liu, Danming Xie, and Lei Xie, “XEmoRAG: Cross-Lingual Emotion Transfer with Controllable Intensity Using Retrieval-Augmented Generation,”arXiv preprint arXiv:2508.07302, 2025

    work page arXiv 2025

  30. [30]

    GigaSpeech 2: An Evolving, Large- Scale and Multi-domain ASR Corpus for Low-Resource Lan- guages with Automated Crawling, Transcription and Refine- ment,

    Yifan Yang, Zheshu Song, Jianheng Zhuo, Mingyu Cui, Jin- peng Li, Bo Yang, Yexing Du, Ziyang Ma, Xunying Liu, Ziyuan Wang, et al., “GigaSpeech 2: An Evolving, Large- Scale and Multi-domain ASR Corpus for Low-Resource Lan- guages with Automated Crawling, Transcription and Refine- ment,”arXiv preprint arXiv:2406.11546, 2024

    work page arXiv 2024

  31. [31]

    MSR-86K: An Evolving, Multilingual Corpus with 86,300 Hours of Transcribed Au- dio for Speech Recognition Research,

    Song Li, Yongbin You, Xuezhi Wang, Zhengkun Tian, Ke Ding, and Guanglu Wan, “MSR-86K: An Evolving, Multilingual Corpus with 86,300 Hours of Transcribed Au- dio for Speech Recognition Research,”arXiv preprint arXiv:2406.18301, 2024

    work page arXiv 2024

  32. [32]

    Common V oice: A Massively-Multilingual Speech Corpus,

    Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saun- ders, Francis M. Tyers, and Gregor Weber, “Common V oice: A Massively-Multilingual Speech Corpus,” inProc. LREC, 2020, pp. 4218–4222

    work page 2020