pith. machine review for the scientific record. sign in

arxiv: 2604.09121 · v3 · submitted 2026-04-10 · 💻 cs.CL · cs.AI· cs.SD

Recognition: no theorem link

Interactive ASR: Towards Human-Like Interaction and Semantic Coherence Evaluation for Agentic Speech Recognition

Peng Wang , Yanqiao Zhu , Zixuan Jiang , Qinyuan Chen , Xingjian Zhao , Xipeng Qiu , Wupeng Wang , Zhifu Gao
show 3 more authors
Xiangang Li Kai Yu Xie Chen
Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:17 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.SD
keywords automatic speech recognitionsemantic evaluationinteractive correctionLLM-as-a-Judgeagentic ASRmulti-turn interactionsemantic fidelity
0
0 comments X

The pith

An agentic ASR framework uses LLM judges for semantic evaluation and multi-turn LLM agents for interactive refinement to improve meaning-level accuracy.

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

The paper seeks to replace token-level word error rate with semantic coherence as the primary measure of ASR quality and to add human-like back-and-forth correction that current systems lack. It builds an LLM-as-a-Judge to score whether an utterance preserves intended meaning and an LLM-driven agent that conducts multi-turn dialogues to fix errors iteratively. A sympathetic reader would care because everyday speech recognition often produces fluent but semantically distorted output and because people naturally clarify rather than restart conversations. Experiments across English, Chinese, and code-switching benchmarks show gains in both automated semantic scores and human judgments of fidelity and correction success.

Core claim

The paper establishes that an LLM prompted as a semantic judge can evaluate recognition quality beyond word matches, and that an LLM agent simulating multi-turn human interaction can iteratively refine ASR outputs through semantic feedback, producing higher semantic fidelity on GigaSpeech, WenetSpeech, and ASRU 2019 code-switching sets as measured by both objective and subjective criteria.

What carries the argument

LLM-as-a-Judge semantic evaluation metric together with an LLM-driven multi-turn agent that issues clarification requests and applies corrections to the ASR transcript.

If this is right

  • ASR quality can be assessed at the level of sentence meaning rather than individual word matches.
  • Systems can recover from recognition errors through dialogue instead of requiring a complete re-recognition pass.
  • The approach scales to multiple languages and code-switching without retraining the base ASR model.
  • Both automated semantic metrics and human listeners register measurable gains in fidelity and correction capability.

Where Pith is reading between the lines

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

  • Real voice interfaces could adopt similar agents to request user clarifications in live settings rather than waiting for full utterances.
  • Post-processing agent layers might reduce reliance on ever-larger training corpora by correcting errors after initial decoding.
  • The same judge-plus-agent pattern could apply to other modalities such as video captioning or live translation where semantic drift also occurs.

Load-bearing premise

An LLM prompted as a judge can reliably judge semantic correctness of ASR outputs without injecting its own biases or hallucinations, and the simulated multi-turn interactions do not degrade final accuracy.

What would settle it

A side-by-side human rating study in which raters consistently assign different semantic-correctness scores to the same ASR outputs than the LLM judge, or a blind test showing that adding interactive turns raises rather than lowers final word or semantic error rates.

Figures

Figures reproduced from arXiv: 2604.09121 by Kai Yu, Peng Wang, Qinyuan Chen, Wupeng Wang, Xiangang Li, Xie Chen, Xingjian Zhao, Xipeng Qiu, Yanqiao Zhu, Zhifu Gao, Zixuan Jiang.

Figure 1
Figure 1. Figure 1: Traditional vs. Interactive ASR paradigms. (Top) Tra￾ditional systems struggle with named entities (e.g., ”Night” vs. ”Knight”); (Bottom), our proposed Interactive ASR can take user’s spoken language instructions (e.g., ”starts with a K”) as the feedback to update and correct ASR results. However, despite the rapid progress in ASR, several impor￾tant aspects remain underexplored. The first concerns the eva… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Interactive ASR framework. An LLM Intent Router classifies the base ASR hypothesis (𝐻𝑡) using the previous transcript (𝑌𝑡−1). New utterances bypass correction and are output directly, while corrective instructions trigger an LLM Reasoning Corrector to refine 𝑌𝑡−1 via a three-step CoT process (Locate, Reason, Surgical Replacement). modal Interfaces [20], requiring users to manually select al… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Automated Simulation Framework. An LLM-as-a-judge first evaluates the semantic coherence of the initial ASR hypothesis. Upon detecting an error, the User Simu￾lator generates a spoken correction instruction via an LLM and TTS. The Interactive ASR then processes this feedback to rea￾son and update the hypothesis, forming an automated correction process. 4.1. User Simulator The User Simulator… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of 𝑆 2𝐸 𝑅 (top) and CER/WER/MER (bot￾tom) reduction trends across three datasets. tively) not as practical operational targets, but to establish the system’s theoretical upper bound. By this 10th loop, our system achieves near-perfect performance. Qualitative analysis reveals that the very few bad cases at this stage primarily stem from cas￾cading ASR errors during the interaction. When the base… view at source ↗
read the original abstract

Recent years have witnessed remarkable progress in automatic speech recognition (ASR), driven by advances in model architectures and large-scale training data. However, two important aspects remain underexplored. First, Word Error Rate (WER), the dominant evaluation metric for decades, treats all words equally and often fails to reflect the semantic correctness of an utterance at the sentence level. Second, interactive correction-an essential component of human communication-has rarely been systematically studied in ASR research. In this paper, we integrate these two perspectives under an agentic framework for interactive ASR. We propose leveraging LLM-as-a-Judge as a semantic-aware evaluation metric to assess recognition quality beyond token-level accuracy. Furthermore, we design an LLM-driven agent framework to simulate human-like multi-turn interaction, enabling iterative refinement of recognition outputs through semantic feedback. Extensive experiments are conducted on standard benchmarks, including GigaSpeech (English), WenetSpeech (Chinese), the ASRU 2019 code-switching test set. Both objective and subjective evaluations demonstrate the effectiveness of the proposed framework in improving semantic fidelity and interactive correction capability. We will release the code to facilitate future research in interactive and agentic ASR.

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

3 major / 2 minor

Summary. The paper proposes an agentic framework for interactive ASR. It uses an LLM-as-a-Judge to provide semantic-aware evaluation of ASR outputs beyond token-level WER and designs an LLM-driven multi-turn agent to simulate human-like interactive corrections via semantic feedback. Experiments are conducted on GigaSpeech (English), WenetSpeech (Chinese), and the ASRU 2019 code-switching set, with claims that both objective and subjective evaluations show improved semantic fidelity and correction capability. Code release is promised.

Significance. If the LLM judge proves reliable and the agent yields genuine gains without accuracy degradation, the work could meaningfully advance ASR beyond WER-centric evaluation and static decoding toward more interactive, semantically grounded systems. The idea of agentic interaction is timely. However, the absence of any reported validation, baselines, or quantitative evidence in the provided description substantially weakens the potential impact at present.

major comments (3)
  1. The abstract asserts effectiveness based on experiments on GigaSpeech, WenetSpeech, and ASRU 2019 but supplies no prompting details, baselines, quantitative deltas, statistical significance, or controls for LLM variability, leaving the central claim of improved semantic fidelity and interactive correction unsupported by visible evidence.
  2. No validation of the LLM-as-Judge metric against human semantic judgments (e.g., correlation coefficients, inter-annotator agreement, or ablation on prompt sensitivity) is described. This is load-bearing because the entire evaluation framework and the agent's feedback loop depend on the judge being free of systematic bias or hallucination when scoring meaning preservation.
  3. The multi-turn agent framework is presented without analysis of error propagation across turns, comparison to non-agentic correction baselines, or measurement of whether iterative refinement preserves or degrades overall WER. These omissions directly affect the claim that the approach enables human-like interaction without accuracy loss.
minor comments (2)
  1. Clarify the exact definition of 'semantic coherence' used by the LLM judge and how it differs from existing semantic similarity metrics in the ASR literature.
  2. The abstract mentions 'standard benchmarks' but does not list exact test-set sizes or splits; include these in the experimental section for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key areas where additional empirical detail and validation will strengthen the paper. We have revised the manuscript to address each major comment directly by expanding the experimental reporting, adding validation studies, and including targeted analyses of the agent framework.

read point-by-point responses

  1. Referee: The abstract asserts effectiveness based on experiments on GigaSpeech, WenetSpeech, and ASRU 2019 but supplies no prompting details, baselines, quantitative deltas, statistical significance, or controls for LLM variability, leaving the central claim of improved semantic fidelity and interactive correction unsupported by visible evidence.

    Authors: We agree the abstract is necessarily concise and omits these specifics. The full manuscript already contains the experimental details in Sections 3 and 4, including the LLM prompts, baseline systems (standard ASR and non-interactive correction), quantitative deltas on both WER and semantic metrics, and significance testing. To improve visibility, we have revised the abstract to reference the key gains and added an explicit summary paragraph plus controls for LLM variability (results averaged over multiple runs with varied seeds) in the experimental section. revision: yes

  2. Referee: No validation of the LLM-as-Judge metric against human semantic judgments (e.g., correlation coefficients, inter-annotator agreement, or ablation on prompt sensitivity) is described. This is load-bearing because the entire evaluation framework and the agent's feedback loop depend on the judge being free of systematic bias or hallucination when scoring meaning preservation.

    Authors: This is a fair and important observation. We have added a dedicated validation subsection that reports Pearson and Spearman correlations with human semantic judgments, inter-annotator agreement statistics, and an ablation on prompt variations to demonstrate robustness. The revised manuscript now explicitly addresses potential biases and how they were mitigated. revision: yes

  3. Referee: The multi-turn agent framework is presented without analysis of error propagation across turns, comparison to non-agentic correction baselines, or measurement of whether iterative refinement preserves or degrades overall WER. These omissions directly affect the claim that the approach enables human-like interaction without accuracy loss.

    Authors: We accept that these analyses are necessary to support the interaction claims. The revision includes new experiments that track WER and semantic scores across turns to quantify error propagation, direct comparisons to non-agentic single-turn baselines, and explicit measurements confirming that iterative refinement preserves (and in some cases slightly improves) WER while raising semantic fidelity. These results appear in updated tables and figures. revision: yes

Circularity Check

0 steps flagged

No circularity: framework and evaluations are externally grounded

full rationale

The paper introduces an LLM-as-Judge metric and multi-turn agent framework for interactive ASR, with effectiveness claims resting on described experiments across GigaSpeech, WenetSpeech, and code-switching benchmarks rather than any internal derivation. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text that would reduce the central claims to tautological inputs by construction. The methodology applies external LLMs for judgment and correction without redefining success metrics in terms of themselves, making the work self-contained against the reported objective and subjective results.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The central claims rest on unverified assumptions about LLM reliability for semantic judgment and interaction simulation; no free parameters are explicitly fitted in the abstract, but the framework implicitly depends on LLM capabilities treated as given.

axioms (2)
  • domain assumption LLMs can serve as reliable judges of semantic correctness in ASR transcripts
    Invoked as the basis for the proposed evaluation metric replacing or supplementing WER.
  • domain assumption LLM agents can simulate effective human-like multi-turn correction without introducing new errors
    Underpins the interactive refinement component of the agent framework.
invented entities (2)
  • LLM-as-a-Judge no independent evidence
    purpose: Semantic-aware evaluation of recognition quality beyond token accuracy
    New role assigned to existing LLMs for this ASR task.
  • LLM-driven agent framework no independent evidence
    purpose: Simulate multi-turn interaction for iterative refinement via semantic feedback
    Proposed architecture for interactive ASR.

pith-pipeline@v0.9.0 · 5542 in / 1456 out tokens · 55331 ms · 2026-05-10T18:17:50.600075+00:00 · methodology

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

38 extracted references · 16 canonical work pages · 4 internal anchors

  1. [1]

    Interactive ASR: Towards Human-Like Interaction and Semantic Coherence Evaluation for Agentic Speech Recognition

    Introduction Automatic speech recognition (ASR) plays a pivotal role in human–computer interaction by enabling computers to under- stand users’ intent through speech. In recent years, ASR tech- nologies have achieved remarkable progress, driven by advances in both model architectures and large-scale training data. Ex- tensive research has explored a varie...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    To address this, several semantic-aware metrics have been proposed

    Related Works While WER has long served as the standard metric for ASR eval- uation, its inherent design of assigning equal weight to all words fails to capture critical semantic errors. To address this, several semantic-aware metrics have been proposed. Semantic WER

  3. [3]

    Moving to embedding-based evaluation, SemDist [18] utilized RoBERTa-based sentence embeddings to measure se- mantic similarity beyond literal overlap

    introduced dynamic weighting, which utilizes Named En- tity Recognition to extract keywords, assigning higher weights to critical entities and lower weights to filler words during error cal- culation. Moving to embedding-based evaluation, SemDist [18] utilized RoBERTa-based sentence embeddings to measure se- mantic similarity beyond literal overlap. Most ...

  4. [4]

    Unlike these continuous scoring metrics, our LLM-as-a-Judge adopts a binary functional criterion, acting as a strict gatekeeper to determine if the user’s intent is executable

    leveraged LLMs to assign graded penalties based on er- ror severity (e.g., ignoring colloquial variations while penalizing meaning changes). Unlike these continuous scoring metrics, our LLM-as-a-Judge adopts a binary functional criterion, acting as a strict gatekeeper to determine if the user’s intent is executable. Regarding error correction based on hum...

  5. [5]

    Proposed Paradigm Let 𝐼 denote the user speech and𝑌 the output transcript. Existing ASR systems operate under a single-pass decoding paradigm: 𝑌 = ASR(𝐼) (1) This formulation is static: once a transcription is produced, the system has no mechanism to incorporate subsequent user feed- back. To address this limitation, we propose the Interactive ASR framewo...

  6. [6]

    Automated Simulation Framework Large-scale human evaluation of interactive ASR in continuous scenarios is expensive and limits reproducibility. To address this and systematically evaluate the corrective capability of our sys- tem (the right branch in Figure 2), we design an automated sim- ulation framework tailored for single-utterance iterative correc- t...

  7. [7]

    We first outline the experimental setup in Section 5.1, detailing the diverse benchmarks and the foun- dational models

    Experiments In this section, we comprehensively evaluate the proposed Inter- active ASR framework. We first outline the experimental setup in Section 5.1, detailing the diverse benchmarks and the foun- dational models. Before analyzing the system performance, we conduct a Human-AI Alignment Study in Section 5.2 to es- tablish the credibility of 𝑆2𝐸 𝑅 by d...

  8. [8]

    We introduced 𝑆2𝐸 𝑅, a novel met- ric that leverages LLMs as judges to prioritize sentence-level semantic coherence

    Conclusion In this work, we addressed two critical limitations in traditional ASR: semantic-blind evaluation and the absence of interactive correction mechanisms. We introduced 𝑆2𝐸 𝑅, a novel met- ric that leverages LLMs as judges to prioritize sentence-level semantic coherence. Furthermore, we proposed an Interactive ASR framework that employs CoT reason...

  9. [9]

    Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning , 2006, pp. 369–376

    2006

  10. [10]

    Sequence transduction with recurrent neural networks.arXiv preprint arXiv:1211.3711,

    A. Graves, “Sequence transduction with recurrent neural net- works,” arXiv preprint arXiv:1211.3711, 2012

    work page arXiv 2012

  11. [11]

    Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,

    W . Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” in 2016 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2016, pp. 4960– 4964

    2016

  12. [12]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

    2023

  13. [13]

    On decoder-only architecture for speech- to-text and large language model integration,

    J. Wu, Y . Gaur, Z. Chen, L. Zhou, Y . Zhu, T. Wang, J. Li, S. Liu, B. Ren, L. Liu et al. , “On decoder-only architecture for speech- to-text and large language model integration,” in 2023 IEEE au- tomatic speech recognition and understanding workshop (ASRU) . IEEE, 2023, pp. 1–8

    2023

  14. [14]

    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. Rubenstein et al. , “Slm: Bridge the thin gap between speech and text foundation models,” in 2023 IEEE Automatic Speech Recognition and Understanding Work- shop (ASRU). IEEE, 2023, pp. 1–8

    2023

  15. [15]

    Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289, 2023

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “Salmonn: Towards generic hearing abilities for large language models,” arXiv preprint arXiv:2310.13289, 2023

    work page arXiv 2023

  16. [16]

    Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models

    Y . Chu, J. Xu, X. Zhou, Q. Y ang, S. Zhang, Z. Y an, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,”arXiv preprint arXiv:2311.07919, 2023

    work page internal anchor Pith review arXiv 2023

  17. [17]

    Prompt- ing large language models with speech recognition abilities,

    Y . Fathullah, C. Wu, E. Lakomkin, J. Jia, Y . Shangguan, K. Li, J. Guo, W. Xiong, J. Mahadeokar, O. Kalinli et al. , “Prompt- ing large language models with speech recognition abilities,” in ICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 13 351–13 355

    2024

  18. [18]

    An embarrassingly simple approach for LLM with strong ASR capacity,

    Z. Ma, G. Y ang, Y . Y ang, Z. Gao, J. Wang, Z. Du, F. Yu, Q. Chen, S. Zheng, S. Zhang et al., “An embarrassingly simple approach for llm with strong asr capacity,” arXiv preprint arXiv:2402.08846 , 2024

    work page arXiv 2024

  19. [19]

    Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition,

    Y . Bai, J. Chen, J. Chen, W. Chen, Z. Chen, C. Ding, L. Dong, Q. Dong, Y . Du, K. Gao et al., “Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition,” arXiv preprint arXiv:2407.04675, 2024

    work page arXiv 2024

  20. [20]

    Fireredasr: Open- source industrial-grade mandarin speech recognition models from encoder-decoder to llm integration,

    K.-T. Xu, F.-L. Xie, X. Tang, and Y . Hu, “Fireredasr: Open-source industrial-grade mandarin speech recognition mod- els from encoder-decoder to llm integration,” arXiv preprint arXiv:2501.14350, 2025

    work page arXiv 2025

  21. [21]

    Fun-ASR technical report,

    K. An, Y . Chen, Z. Chen, C. Deng, Z. Du, C. Gao, Z. Gao, B. Gong, X. Li, Y . Li et al. , “Fun-asr technical report,” arXiv preprint arXiv:2509.12508, 2025

    work page arXiv 2025

  22. [22]

    Qwen3-ASR Technical Report

    X. Shi, X. Wang, Z. Guo, Y . Wang, P . Zhang, X. Zhang, Z. Guo, H. Hao, Y . Xi, B. Y anget al., “Qwen3-asr technical report,” arXiv preprint arXiv:2601.21337, 2026

    work page internal anchor Pith review arXiv 2026

  23. [23]

    arXiv preprint arXiv:2106.02016 , year=

    S. Roy, “Semantic-wer: A unified metric for the evaluation of asr transcript for end usability,” arXiv preprint arXiv:2106.02016, 2021

    work page arXiv 2021

  24. [24]

    De- noising ger: A noise-robust generative error correction with llm for speech recognition,

    Y . Liu, M. Xu, Y . Chen, L. He, L. Fang, S. Fang, and L. Liu, “De- noising ger: A noise-robust generative error correction with llm for speech recognition,” in 2025 International Joint Conference on Neural Networks (IJCNN). IEEE, 2025, pp. 1–8

    2025

  25. [25]

    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. Xing et al., “Judging llm-as-a-judge with mt-bench and chatbot arena,” Advances in neural information pro- cessing systems, vol. 36, pp. 46 595–46 623, 2023

    2023

  26. [26]

    Semantic distance: A new metric for asr perfor- mance analysis towards spoken language understanding,

    S. Kim, A. Arora, D. Le, C.-F. Y eh, C. Fuegen, O. Kalinli, and M. L. Seltzer, “Semantic distance: A new metric for asr perfor- mance analysis towards spoken language understanding,” arXiv preprint arXiv:2104.02138, 2021

    work page arXiv 2021

  27. [27]

    Laser: An llm-based asr scoring and evaluation rubric,

    A. Parulekar and P . Jyothi, “Laser: An llm-based asr scoring and evaluation rubric,” in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025, pp. 24 773–24 782

    2025

  28. [28]

    Multimodal error correction for speech user interfaces,

    B. Suhm, B. Myers, and A. Waibel, “Multimodal error correction for speech user interfaces,”ACM transactions on computer-human interaction (TOCHI), vol. 8, no. 1, pp. 60–98, 2001

    2001

  29. [29]

    Efficient speech transcription through respeaking

    M. Sperber, G. Neubig, C. Fügen, S. Nakamura, and A. Waibel, “Efficient speech transcription through respeaking.” in Inter- speech, 2013, pp. 1087–1091

    2013

  30. [30]

    React: Synergizing reasoning and acting in language models,

    S. Y ao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,” in The eleventh international conference on learning rep- resentations, 2022

    2022

  31. [31]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al., “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural information pro- cessing systems, vol. 35, pp. 24 824–24 837, 2022

    2022

  32. [32]

    The asru 2019 mandarin-english code-switching speech recognition challenge: Open datasets, tracks, methods and results,

    X. Shi, Q. Feng, and L. Xie, “The asru 2019 mandarin-english code-switching speech recognition challenge: Open datasets, tracks, methods and results,” arXiv preprint arXiv:2007.05916 , 2020

    work page arXiv 2019

  33. [33]

    Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio,

    G. Chen, S. Chai, G. Wang, J. Du, W.-Q. Zhang, C. Weng, D. Su, D. Povey, J. Trmal, J. Zhang et al. , “Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio,” arXiv preprint arXiv:2106.06909, 2021

    work page arXiv 2021

  34. [34]

    Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,

    B. Zhang, H. Lv, P . Guo, Q. Shao, C. Y ang, L. Xie, X. Xu, H. Bu, X. Chen, C. Zeng et al., “Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 6182–6186

    2022

  35. [35]

    Qwen3 Technical Report

    A. Y ang, A. Li, B. Y ang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv et al., “Qwen3 technical report,” arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    Indextts: An industrial-level controllable and efficient zero-shot text-to-speech system,

    W . Deng, S. Zhou, J. Shu, J. Wang, and L. Wang, “Indextts: An industrial-level controllable and efficient zero-shot text-to-speech system,” arXiv preprint arXiv:2502.05512, 2025

    work page arXiv 2025

  37. [37]

    Zechner and K

    K. Zechner and K. Evanini, Eds., Automated Speaking As- sessment: Using Language Technologies to Score Spontaneous Speech, 1st ed. Routledge, 2019

    2019

  38. [38]

    Proceedings of the Royal Society of London 58, 240–242

    K. Pearson, “Vii. note on regression and inheritance in the case of two parents,” Proceedings of the Royal Society of London , vol. 58, no. 347-352, pp. 240–242, 12 1895. [Online]. Available: https://doi.org/10.1098/rspl.1895.0041

    work page doi:10.1098/rspl.1895.0041