Pith. sign in

REVIEW 3 major objections 5 minor 34 references

UniSLU: Unified Spoken Language Understanding from Heterogeneous Cross-Task Datasets

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single generative decoder can transcribe speech, extract named entities, and classify sentiment in one pass, outperforming separate task models on the SLUE benchmark.

desk verdict A clean, useful template for unified SLU, but the headline gain over GenSLU is confounded by data volume and the dynamic loss, so the claim needs a fair baseline before it sticks. read the letter →

arxiv 2507.12951 v1 pith:M4DC7VMP submitted 2025-07-17 eess.AS cs.AIcs.CLcs.MMcs.SD

Zhichao Sheng , Shilin Zhou , Chen Gong , Zhenghua Li This is my paper · ORCID
classification eess.AScs.AIcs.CLcs.MMcs.SD
keywords spokenlanguageunderstandingautomaticspeechrecognitionnamedentitysentimentanalysisunifiedgenerativemodelmultitasklearningheterogeneousdatasetsspeech-to-textsequencegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that three spoken-language tasks—transcribing speech, extracting named entities, and classifying sentiment—can be solved by one generative model instead of separate systems. The key claim is a shared output template, audio transcript followed by a delimiter, a task token, and the structured result, which lets a single decoder learn from datasets that are only partially annotated per task. Fine-tuning a pretrained speech encoder-decoder on the two SLUE datasets with this template raises the overall SLU score to 68.97 for the small model and 73.81 for the medium model, above the best separate generative baselines at 66.17 and 71.65. If correct, heterogeneous speech corpora become usable for joint training without needing multi-task aligned annotations.

What carries the argument

The load-bearing object is the unified sequence template with four named segments: the ASR transcription, the '[T/L]' end-of-transcript delimiter, a task-control token ('[NER]' or '[SA]'), and the task-specific output (tagged entities or sentiment label). The template converts all three tasks into one text-generation problem, so the decoder autoregressively produces the transcript, transitions at the delimiter, and then generates task output conditioned on the transcript. The second mechanism is the dynamic loss, which assigns each task a weight equal to the other task's token length divided by total sequence length, counteracting the tendency for long ASR transcripts to dominate the cross-entropy objective.

What would settle it

Probe the fine-tuned model on utterances whose transcripts naturally contain the literal strings '[T/L]', '[NER]', or '[SA]', or compare free decoding against forced decoding with oracle control tokens; if task outputs collapse when the delimiter is removed or when these strings appear naturally, the claim that the template provides clean task boundaries would be refuted.

Watch

Extended reading notes

Core claim

UniSLU establishes that automatic speech recognition, spoken named-entity recognition, and spoken sentiment analysis can be jointly modeled by formatting every instance as '[ASR Transcript][T/L][Task-Control Token][Task-Specific Outputs]', for example 'This council made great progress[T/L][NER][ORG]council[/ORG]', and fine-tuning a single pretrained encoder-decoder on mixed SLUE-VoxPopuli and SLUE-VoxCeleb data. The model generates the transcript first, emits the '[T/L]' delimiter, then continues under the inserted '[NER]' or '[SA]' control token to produce the task output in the same decoding pass. A length-proportional dynamic loss reweights the ASR portion so that the shorter NER and SA outputs are not dominated during training. The paper reports that UniSLU yields a higher SLUE SCORE than separate-model and separate-generative baselines, with the gains attributed to shared representations and the joint use of both heterogeneous datasets.

Load-bearing premise

The decoder can learn, from roughly 10,777 training sentences, to emit the exact unified sequence—transcript, '[T/L]', task token, structured output—with the inserted tokens acting as reliable boundaries rather than being confused with ordinary lexical content.

Editorial extensions

If this is right

  • A single fine-tuned decoder can serve ASR, NER, and SA simultaneously, removing separate classification heads and task-specific modules.
  • Datasets annotated for only one task, such as ASR plus NER or ASR plus SA, can be combined in one training run, reducing dependence on rare multi-task aligned corpora.
  • Scaling to a larger pretrained backbone increases the gains, suggesting the unified formulation benefits from model capacity.
  • Plugging a large language model decoder into the framework boosts NER and SA further, reaching a SLUE SCORE of 72.75 with a 7B decoder, while ASR worsens slightly.
  • The dynamic length weighting is most valuable for smaller models and long utterances, preventing ASR from crowding out NER and SA learning.

Reading between the lines

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

  • An implication the paper leaves implicit is that the control tokens could be evaluated as explicit segment boundaries; a direct error analysis around '[T/L]' and the task tokens would clarify whether the delimiter genuinely separates transcript and task output or merely co-occurs with them.
  • The same template mechanism could plausibly extend to other spoken-language tasks such as speaker-change detection, intent classification, or emotion tagging, reusing the unified generative framework without architectural changes.
  • Because the dynamic loss reweights purely by token length, alternative weighting schemes based on uncertainty or task difficulty might deliver similar or better balance; the paper does not explore them.
  • If the template proves robust, partially labeled speech corpora across languages could be pooled for joint training without requiring transcription-level alignment, which would broaden the data available for SLU systems.
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 proposes UniSLU, a unified generative framework that jointly models ASR, spoken NER, and spoken SA within a single Whisper-based encoder-decoder. The key idea is a unified output template, "[ASR Transcript][T/L][Task-Control Token][Task-Specific Outputs]", where [T/L] separates the transcript, and [NER] or [SA] selects the downstream task. The model is fine-tuned on two heterogeneous SLUE datasets (VoxPopuli for NER, VoxCeleb for SA) using a dynamic length-weighted cross-entropy loss. Experiments on SLUE test sets report SLUE SCORE improvements over separate generative baselines (GenSLU) and over SLUE's separate classification baselines, with additional experiments on ablations, sequence-length robustness, and integration with Qwen2.5 LLM decoders. The central claim is that a single model can exploit partially annotated, heterogeneous speech datasets to improve overall SLU performance.

Significance. If the claimed gains hold, the paper offers a simple and practical way to unify ASR, NER, and SA in one generative model, avoiding the need for costly multi-task aligned data while still leveraging cross-task interaction. The dynamic loss is a sensible mechanism to counter ASR's usually longer output dominating the loss. The paper also reports three-seed averaging and plans code/model release, which are good reproducibility habits. The main scientific value is the demonstration that heterogeneous datasets with different task annotations can be combined in a single template, a useful direction for the SLU community. However, the headline comparison is currently confounded, and the absence of variance reporting makes the small residual gains difficult to evaluate, which tempers the significance until these concerns are addressed.

major comments (3)
  1. [Table 3 and Section 4.1] The central comparison between UniSLU and GenSLU is confounded by two simultaneous changes: GenSLU is trained per dataset (VoxPopuli for NER, VoxCeleb for SA) and does not use the dynamic loss of Eqs. (4)-(5), while UniSLU trains on both datasets and uses dynamic loss. The numbers in Table 3 show how much this matters: on Whisper-small, GenSLU scores 66.17, UniSLU without dynamic loss scores 66.57, and full UniSLU scores 68.97. Thus the unified representation and joint data alone contribute only +0.40 SLUE, while dynamic loss contributes +2.40. On Whisper-medium the pattern reverses (dynamic loss adds only +0.32 while joint modeling adds +1.84). Because GenSLU was not rerun with dynamic loss or on combined data, the paper's attributions to 'jointly train on both datasets' and 'multitask interaction' (Section 4.1) are not controlled. Please add GenSLU with dynamic loss and with combined data, and ablations of UniSLU on each single dataset, to isolate the effect of the unified template.
  2. [Section 4, Implementation Details] The paper states that each model is run three times with different random seeds and the average is reported, but no standard deviations, confidence intervals, or significance tests are given anywhere in the results. Several reported differences are small (e.g., the +0.40 SLUE between UniSLU without dynamic loss and GenSLU on Whisper-small), and without variance information it is impossible to tell whether the unified-template effect is real or within seed noise. Please report per-seed results or standard deviations for the key comparisons in Tables 3 and 4, and run a paired significance test (e.g., bootstrap or t-test over the three seeds) for the headline comparisons.
  3. [Section 3.2 and Table 3] The proposed mechanism relies on the decoder correctly learning to emit the template tokens [T/L], [NER], and [SA] and using them to transition to task-specific outputs. However, the paper evaluates only end-to-end task metrics (WER, Micro-F1, Macro-F1, SLUE SCORE) and never analyzes whether the generated sequences actually conform to the template, where errors occur relative to the control tokens, or how often the model confuses control tokens with lexical content. This is load-bearing because the claimed benefit of multitask interaction depends on the template functioning as a clean task boundary. Please add an error analysis, such as the success rate of generating [T/L], [NER], and [SA] in the correct positions, and an example-level breakdown of errors before vs. after the delimiter.
minor comments (5)
  1. [Equation (5)] The symbol 'Lentaski' appears to be a typo for 'Len_taski'; please fix and also clarify that the length measurement is per training sample at the token level, as mentioned in the text but not in the equation notation.
  2. [Table 3] The column headers are confusing: 'ASR-WER' has VP and VC subcolumns, and 'NER-F1' and 'SA-F1' each also have two subcolumns. Please make explicit that WER is reported for both VP and VC, while NER-F1 is only on VP and SA-F1 is only on VC, e.g., by using 'WER-VP', 'WER-VC', 'NER-F1 (VP)', 'SA-F1 (VC)'.
  3. [Section 4.4] The sentence 'Among these five experiments, only "UniSLU (Whisper-small)" is not pre-trained on LS960' is inaccurate or at least unclear, because the LLM-based variants also use an adapter pretrained on LibriSpeech-960. Please rephrase to specify that UniSLU (Whisper-small) is the only model that did not undergo the additional adapter pretraining on LS960, or clarify that the comparison is about the adapter, not the audio encoder.
  4. [Figure 3] The x-axis is labeled 'Seq. Length' but the caption does not state that lengths are measured in words of the ASR transcript; please add this detail to the caption or the axis label.
  5. [References] Several references contain formatting artifacts (e.g., 'V oxPopuli', 'V oxCeleb' with nonbreaking thin spaces, and 'Arxiv' and 'Arxiv' with inconsistent capitalization). Please normalize the reference list and check all URLs.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the evaluation is against held-out SLUE test sets with fixed public checkpoints, and the only self-citation is not load-bearing.

full rationale

The paper's derivation chain is: define a unified template in Sec. 3.1, train a generative Whisper-based model in Sec. 3.2 with the loss in Eqs. (4)-(5), and evaluate on the public SLUE-VoxPopuli and SLUE-VoxCeleb test sets using the external SLUE SCORE, WER, and F1 metrics. No fitted parameter is renamed as a prediction: Eqs. (4)-(5) only reweight the cross-entropy terms and do not enter the test metrics. The comparison against GenSLU is an internal baseline, but GenSLU is itself run on the same held-out test sets with fixed public Whisper checkpoints; the fact that GenSLU lacks the dynamic loss and joint data is an experimental confound, not a circular reduction. The ablation row 'w/o Dynamic Loss' in Table 3 even shows that the template plus joint data alone yields 66.57 versus GenSLU's 66.17 on Whisper-small, so the headlined gain is not forced by the template definition. The only self-citation found is Zhou et al. (2024), co-authored by current authors, cited in the Introduction to define NER as a task; it is not used to justify the unified representation, the training objective, or the benchmark claims, and it does not invoke any uniqueness theorem. Thus the central empirical claim is externally testable and not circular; at most there is a minor, non-load-bearing self-citation.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central result rests on the learnability of the unified generative template, the transferability of Whisper features, and the validity of the composite SLUE SCORE as the main measure. The main free parameters are training hyperparameters chosen without a reported search. The special tokens are invented vocabulary entries with no independent evidence. The dynamic loss formula is an ad hoc weighting, listed as an axiom rather than a free parameter because it is a design rule applied to each sample.

free parameters (6)
  • Learning rate (Whisper fine-tuning) = 1e-5 (small), 2e-5 (medium)
    Chosen by hand; no search procedure reported.
  • Learning rate (LLM adapter fine-tuning) = 2e-4
    Used for adapter pretraining and LoRA fine-tuning; no ablation.
  • Fine-tuning epochs = 100 (Whisper), 50 (LLM)
    Hand-picked; no dev-stopping criterion was reported.
  • LoRA rank = 32
    Standard default; no rank ablation.
  • Beam size = 5 (Whisper), 1 (LLM)
    Inference-time choice; affects NER tag generation and transcript quality.
  • Sampling parameters for LLM = temperature 0.7, top-p 0.9, repetition penalty 1.2
    LLM decoding settings; not ablated.
assumptions (6)
  • standard math Cross-entropy loss and left-to-right autoregressive decoding are appropriate objectives for ASR, NER, and SA generation.
    Used in Eq. 4 without discussion; standard in sequence modeling.
  • domain assumption Whisper's pretrained encoder and decoder transfer to spoken NER and SA in a generative setting.
    Section 3.2 adopts Whisper modules directly; no analysis of why the pretrained decoder can generate entity or sentiment markup.
  • ad hoc to paper The unified template with [T/L], [NER], [SA] and entity tags is learnable from a small amount of data.
    Template introduced in Section 3.1; no tokenization or convergence analysis beyond final metrics.
  • ad hoc to paper The dynamic loss weighting (Eq. 5) balances tasks better than uniform loss.
    Heuristic in Eq. 4-5; ablation results are positive for Whisper-small (68.97 vs 66.57) but weak for Whisper-medium (73.81 vs 73.49).
  • domain assumption SLUE-VoxPopuli and SLUE-VoxCeleb test sets are representative and their label schemes are compatible in one output vocabulary.
    Evaluation relies entirely on these two datasets (Table 1); mixed entity and sentiment vocabulary is used without a leakage or ambiguity check.
  • domain assumption Fine-tuning the Whisper encoder helps across heterogeneous audio sources.
    Stated in Section 3.2 Training Process; ablation 'w/o Fine-tuning Encoder' shows this is true for Whisper-small but not for Whisper-medium ASR.
invented entities (1)
  • Special tokens [T/L], [NER], [SA] and entity tags such as [ORG]
    purpose: Separate the ASR transcript from downstream task outputs and identify the active SLU task within one generated sequence.
    These new vocabulary entries get their semantics only from the training template; the paper provides no external evidence (e.g., tokenization analysis or forced-decoding checks) that the tokens behave as task boundaries.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniSLU: Unified Spoken Language Understanding from Heterogeneous Cross-Task Datasets." pith.science (2026). https://pith.science/paper/M4DC7VMP

@misc{pith2026250712951,
  author       = {Pith},
  title        = {Pith review of: UniSLU: Unified Spoken Language Understanding from Heterogeneous Cross-Task Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M4DC7VMP}},
  note         = {Machine review of arXiv:2507.12951}
}
read the original abstract

Spoken Language Understanding (SLU) plays a crucial role in speech-centric multimedia applications, enabling machines to comprehend spoken language in scenarios such as meetings, interviews, and customer service interactions. SLU encompasses multiple tasks, including Automatic Speech Recognition (ASR), spoken Named Entity Recognition (NER), and spoken Sentiment Analysis (SA). However, existing methods often rely on separate model architectures for individual tasks such as spoken NER and SA, which increases system complexity, limits cross-task interaction, and fails to fully exploit heterogeneous datasets available across tasks. To address these limitations, we propose UniSLU, a unified framework that jointly models multiple SLU tasks within a single architecture. Specifically, we propose a unified representation for diverse SLU tasks, enabling full utilization of heterogeneous datasets across multiple tasks. Built upon this representation, we propose a unified generative method that jointly models ASR, spoken NER, and SA tasks, enhancing task interactions and enabling seamless integration with large language models to harness their powerful generative capabilities. Extensive experiments on public SLU datasets demonstrate the effectiveness of our approach, achieving superior SLU performance compared to several benchmark methods, making it well-suited for real-world speech-based multimedia scenarios. We will release all code and models at github to facilitate future research.

Figures

Figures reproduced from arXiv: 2507.12951 by the authors.

Figure 1
Figure 1. Illustration of real-world SLU application [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The architecture of our proposed UniSLU framework. Whisper (Radford et al., 2023), the audio signal is transformed into a log-mel spectrogram representa￾tion to extract audio features before being fed into the encoder. Formally, given an input speech signal X of length t, the audio feature extractor produces: F = MelSpec(X), F ∈ Rt ′×d (1) where t ′ is the downsampled sequence length, and d is the feature dimension.… view at source ↗
Figure 3
Figure 3. Comparison of Micro-F1, Macro-F1 and SLUE SCORE metrics across different sequence length categories [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, and Furu Wei. 2022. Speecht5: Unified-modal encoder-decoder pre-training for spoken language processing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL...

    work page 2022

  2. [2]

    Gil Ayache, Menachem Pirchi, Aviv Navon, Aviv Shamsian, Gill Hetz, and Joseph Keshet. 2024. Whisperner: Unified open named entity and speech recognition. Arxiv, abs/2409.08107

    arXiv 2024

  3. [3]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems

    work page 2020

  4. [4]

    Dario Bertero, Farhad Bin Siddique, Chien - Sheng Wu, Yan Wan, Ricky Ho Yin Chan, and Pascale Fung. 2016. Real-time speech emotion and sentiment recognition for interactive dialogue systems. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016 , pages 1042--1047. The ...

    work page 2016

  5. [5]

    Boli Chen, Guangwei Xu, Xiaobin Wang, Pengjun Xie, Meishan Zhang, and Fei Huang. 2022. AISHELL-NER: named entity recognition from chinese speech. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022 , pages 8352--8356. IEEE

    work page 2022

  6. [6]

    Luyao Cheng, Siqi Zheng, Qinglin Zhang, Hui Wang, Yafeng Chen, and Qian Chen. 2023. Exploring speaker-related information in spoken language understanding for better speaker diarization. In Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023 , pages 14068--14077. Association for Computational Linguistics

    work page 2023

  7. [7]

    Yuya Chiba and Ryuichiro Higashinaka. 2025. Investigating the impact of incremental processing and voice activity projection on spoken dialogue systems. In Proceedings of the 31st International Conference on Computational Linguistics, pages 3687--3696, Abu Dhabi, UAE. Association for Computational Linguistics

    work page 2025

  8. [8]

    Zhao, Yanping Huang, Andrew M

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro - Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Y. Zhao, Yanp...

    work page 2024

Show all 34 references
  1. [9]

    Chenquan Dai, Xiaobin Zhuang, and Jiaxin Cai. 2023. A survey on deep learning for chinese medical named entity recognition. In Proceedings of the 9th International Conference on Computing and Artificial Intelligence, ICCAI 2023, Tianjin, China, March 17-20, 2023 , pages 472--476. ACM

    2023

  2. [10]

    Pavel Denisov and Ngoc Thang Vu. 2023. Leveraging multilingual self-supervised pretrained models for sequence-to-sequence end-to-end spoken language understanding. In IEEE Automatic Speech Recognition and Understanding Workshop, ASRU 2023, Taipei, Taiwan, December 16-20, 2023 ...

    2023

  3. [11]

    Zhihao Du, Shiliang Zhang, Siqi Zheng, and Zhi - Jie Yan. 2022. Speaker overlap-aware neural diarization for multi-party meeting analysis. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, De...

    2022

  4. [12]

    Sahar Ghannay, Antoine Caubri \` e re, Yannick Est \` e ve, Nathalie Camelin, Edwin Simonnet, Antoine Laurent, and Emmanuel Morin. 2018. End-to-end named entity and semantic concept extraction from speech. In 2018 IEEE Spoken Language Technology Workshop, SLT 2018, Athens, Gre...

    2018

  5. [13]

    Saki Imai, Tahiya Chowdhury, and Amanda J. Stent. 2025. Evaluating open-source ASR systems: Performance across diverse audio conditions and error correction methods. In Proceedings of the 31st International Conference on Computational Linguistics, pages 5027--5039

    2025

  6. [14]

    Madikeri, Juan Pablo Zuluaga - Gomez, Iuliia Thorbecke, Esa \' u Villatoro - Tello, Sergio Burdisso, Petr Motl \' cek, Karthik S, and Aravind Ganapathiraju

    Shashi Kumar, Srikanth R. Madikeri, Juan Pablo Zuluaga - Gomez, Iuliia Thorbecke, Esa \' u Villatoro - Tello, Sergio Burdisso, Petr Motl \' cek, Karthik S, and Aravind Ganapathiraju. 2024. Tokenverse: Towards unifying speech and NLP tasks via transducer-based ASR . In Proceedi...

    2024

  7. [15]

    Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, and Emmanuel Dupoux. 2021. On generative spoken language modeling from raw audio. Transactions of the Association for Compu...

    2021

  8. [16]

    Zhiyun Lu, Liangliang Cao, Yu Zhang, Chung - Cheng Chiu, and James Fan. 2020. Speech sentiment analysis via pre-trained features from end-to-end ASR models. In 2020 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2020, Barcelona, Spain, May 4-8...

    2020

  9. [17]

    Saturnino Luz, Matt-Mouley Bouamrane, and Masood Masoodian. 2006. Gathering a corpus of multimodal computer-mediated meetings. In Proceedings of the Fifth International Conference on Language Resources and Evaluation ( LREC `06) , Genoa, Italy. European Language Resources Asso...

    2006

  10. [18]

    Kentaro Mitsui, Koh Mitsuda, Toshiaki Wakatsuki, Yukiya Hono, and Kei Sawada. 2024. PSLM: parallel generation of text and speech with llms for low-latency spoken dialogue systems. In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, No...

    2024

  11. [19]

    Motoi Omachi, Yuya Fujita, Shinji Watanabe, and Matthew Wiesner. 2021. End-to-end ASR to jointly predict transcriptions and linguistic annotations. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2021

  12. [20]

    Soujanya Poria, Erik Cambria, Rajiv Bajpai, and Amir Hussain. 2017. A review of affective computing: From unimodal analysis to multimodal fusion. Inf. Fusion, 37:98--125

    2017

  13. [21]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , volume 202 of Proc...

    2023

  14. [22]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1--140:67

    2020

  15. [23]

    Suwon Shon, Ankita Pasad, Felix Wu, Pablo Brusco, Yoav Artzi, Karen Livescu, and Kyu Jeong Han. 2022. SLUE: new benchmark tasks for spoken language understanding evaluation on natural speech. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2...

    2022

  16. [24]

    Bin Wang, Meishan Zhang, Hao Fei, Yu Zhao, Bobo Li, Shengqiong Wu, Wei Ji, and Min Zhang. 2024. Speechee: A novel benchmark for speech event extraction. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 -...

    2024

  17. [25]

    Changhan Wang, Morgane Rivi \` e re, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Miguel Pino, and Emmanuel Dupoux. 2021. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation....

    2021

  18. [26]

    Yi Xu, Hai Zhao, and Zhuosheng Zhang. 2021. Topic-aware multi-turn dialogue modeling. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educ...

    2021

  19. [27]

    Hemant Yadav, Sreyan Ghosh, Yi Yu, and Rajiv Ratn Shah. 2020. End-to-end named entity recognition from english speech. In 21st Annual Conference of the International Speech Communication Association, Interspeech 2020, Virtual Event, Shanghai, China, October 25-29, 2020, pages ...

    2020

  20. [28]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  21. [29]

    Hao Yang, Min Zhang, Shimin Tao, Miaomiao Ma, and Ying Qin. 2023. Chinese ASR and NER improvement based on whisper fine-tuning. In 25th International Conference on Advanced Communication Technology, ICACT 2023, Pyeongchang, Korea, Republic of, February 19-22, 2023 , pages 213-...

    2023

  22. [30]

    Ziqi Yuan, Baozheng Zhang, Hua Xu, Zhiyun Liang, and Kai Gao. 2024. O pen VNA : A framework for analyzing the behavior of multimodal language understanding system under noisy scenarios. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics ...

    2024

  23. [31]

    Shilin Zhou, Zhenghua Li, Chen Gong, Lei Zhang, Yu Hong, and Min Zhang. 2024. C hinese spoken named entity recognition in real-world scenarios: Dataset and approaches. In Findings of the Association for Computational Linguistics: ACL 2024, pages 1872--1884. Association for Com...

    2024

  24. [32]

    Yicheng Zou, Lujun Zhao, Yangyang Kang, Jun Lin, Minlong Peng, Zhuoren Jiang, Changlong Sun, Qi Zhang, Xuanjing Huang, and Xiaozhong Liu. 2021. Topic-oriented spoken dialogue summarization for customer service with saliency-aware topic modeling. In Thirty-Fifth AAAI Conference...

    2021

  25. [33]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  26. [34]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.