Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

CoLMbo: Speaker Language Model for Descriptive Profiling

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CoLMbo claims a 24.85M-parameter model can generate accurate, context-rich speaker descriptions from a fixed speaker embedding without having seen the speaker in training.

desk verdict CoLMbo is a reasonable prefix-mapping approach to speaker profiling, but its headline accuracy numbers rest on unverifiable label sources and an internal fine-tuning contradiction. read the letter →

arxiv 2506.09375 v2 pith:REFRDJ3A submitted 2025-06-11 cs.CL cs.SDeess.AS

Massa Baali , Shuo Han , Syed Abdul Hannan , Purusottam Samal , Karanveer Singh , Soham Deshmukh , Rita Singh , Bhiksha Raj This is my paper
classification cs.CLcs.SDeess.AS
keywords speakerlanguagemodelprofilingembeddingprompt-basedconditioningzero-shotrecognitionattributepredictionaudiocaptioningprefixtuning
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 claims that a small generative model, CoLMbo, can write rich natural-language descriptions of a speaker from nothing but a fixed voice embedding, and that it predicts attributes such as gender, age, dialect, and ethnicity with near-perfect accuracy without ever seeing the target speaker during training. The authors argue that this overcomes the rigidity of conventional speaker-profiling classifiers, which are locked to predefined categories and cannot produce free-form descriptions. A trainable mapper turns the embedding into a forty-token audio prefix, a frozen language model turns the prefix into text, and a speaker-classification side loss forces the prefix to stay speaker-specific. If the claims hold, a 24.85M-parameter system can outperform a 7.7B general audio-language model on profiling tasks and adapt to new attributes simply by adding training captions.

What carries the argument

The load-bearing mechanism is prefix conditioning with a speaker-clustering auxiliary loss. The audio prefix is generated by a mapper adapted from CLIPCap, an image-captioning prefix-mapping design: a fully-connected layer expands the 1024-dimensional PDAF embedding to 30,720 values, reshaped into 40 vectors of 768 dimensions, then passed through eight transformer layers. Sending the first prefix vector through a linear speaker classifier adds a speaker-identification cross-entropy term to the next-token prediction loss, which the authors argue makes the prefix cluster by speaker even when the description itself does not mention identity. The frozen text decoder keeps the model small and lets new speaker attributes be added by augmenting the caption data rather than changing the architecture.

What would settle it

Release the TEARS caption dataset with the exact evaluation splits and the provenance of every age and ethnicity label; then rerun the evaluation using only TIMIT's documented dialect-region and sex labels as ground truth. If near-perfect accuracy does not survive on publicly documented labels, the central claim of generalizable zero-shot attribute prediction fails.

Watch

Extended reading notes

Core claim

CoLMbo's central claim is that speaker profiling can be re-framed as a conditional text-generation task. A PDAF speaker encoder extracts a 1024-dimensional embedding from a Mel-spectrogram; a lightweight mapper (MLP or small transformer) projects this embedding into a sequence of 40 prefix vectors; the prefix is concatenated with a ten-token text prompt; and a frozen GPT-2 generates the description. The first prefix vector also feeds a linear speaker classifier whose cross-entropy loss is added to the captioning loss, which the authors say forces the prefix to preserve speaker identity across recordings. The model is trained on TEARS, a dataset the authors build by asking Llama3 to write natural-language captions from the metadata of EARS and TIMIT, and is then evaluated zero-shot on VoxCeleb1-O, EARS, and TIMIT, reporting accuracy as high as 97.87 percent for age, 88.58 percent for dialect, and 86.42 percent for ethnicity, with BERT score agreement above 0.97 on several cells. The paper's stated result is that a 24.85M-parameter model outperforms Qwen Audio, a 7.7B-parameter model, across these attributes while also producing interpretable attention maps over the prefix.

Load-bearing premise

The evaluation depends on ground-truth age and ethnicity labels for VoxCeleb and TIMIT that are never sourced; TIMIT only documents dialect region and sex, and VoxCeleb does not distribute age or ethnicity labels, so if those labels were constructed by the authors, every accuracy table is unverifiable.

Editorial extensions

If this is right

  • If the accuracy numbers are reproducible, a 24.85M-parameter model can perform descriptive speaker profiling that previously required billion-parameter audio-language models.
  • The same architecture can absorb new speaker traits (e.g., emotional state or vocal health) by adding captioned audio, with no architectural change.
  • Zero-shot generalization across VoxCeleb, EARS, and TIMIT implies the learned prefix mapper, not dataset-specific classifiers, carries the profiling knowledge.
  • The attention maps over the prefix offer a debugging path for attribute-specific behavior, since different attribute values activate different prefix regions.

Reading between the lines

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

  • The authors leave implicit that the reported accuracies stand on ground-truth age and ethnicity labels that the paper never sources; until those labels are released or traced, the numbers should be treated as conditional on the authors' label construction.
  • The authors do not discuss that, because the training captions are generated by an LLM from dataset metadata, CoLMbo may be learning the LLM's stereotypes about voice and demographics rather than acoustic evidence; a testable extension is to compare predictions on speakers whose self-reported attributes contradict the LLM's priors.
  • A consequence the authors do not develop is that the same prefix-clustering design could enable privacy-invasive inference of sensitive attributes from voice without consent, which motivates controlled disclosure of the caption data.
  • An immediate testable extension would be to evaluate the model on a dataset with self-reported demographic labels, which would separate acoustic inference from caption-pipeline bias.
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

5 major / 5 minor

Summary. The paper introduces CoLMbo, a 24.85M-parameter speaker language model that maps a fixed speaker embedding to a soft prefix for a small causal LM (GPT-2), trained on the TEARS dataset of Llama3-generated captions derived from EARS and TIMIT metadata. The authors report attribute prediction accuracies for age, gender, ethnicity, and dialect on VoxCeleb1-O, VoxCeleb2, EARS, and TIMIT, claiming near-perfect performance and zero-shot generalization, and they ablate the audio mapper, speaker loss, and LM fine-tuning.

Significance. If the reported results are correct, CoLMbo would be a substantial practical contribution: a small, computationally cheap model that generates descriptive speaker profiles and predicts demographic attributes in an adaptable, prompt-conditioned manner. The architecture is clearly described in terms of mapper and prefix lengths, the training objective is explicit, and the paper includes an ablation study and a public code link. However, the central empirical claims are currently not verifiable because the ground-truth labels for age and ethnicity on VoxCeleb and TIMIT are not sourced, and the zero-shot framing overstates the evidence. The contribution is therefore promising but not yet demonstrated.

major comments (5)
  1. [Tables II–III and Section IV] The accuracy numbers for age, gender, ethnicity, and dialect in Tables II and III require ground-truth labels whose provenance is never stated. TIMIT provides sex and dialect-region labels but no ethnicity labels, and VoxCeleb does not distribute age or ethnicity labels for VoxCeleb1-O/VoxCeleb2. The paper neither cites an external label source nor describes an annotation protocol (annotation counts, annotator qualifications, inter-annotator agreement). If the labels were derived from the Llama3-generated TEARS captions, the evaluation is circular, because CoLMbo was trained on those captions. This is load-bearing for the central claim of near-perfect attribute prediction, and it cannot be checked or reproduced as written.
  2. [Section V and Section VI.B] The paper repeatedly claims zero-shot evaluation on all datasets, but Section V explicitly splits TIMIT and EARS into training and test sets that are used to train CoLMbo. Evaluation on TIMIT and EARS is therefore in-domain, not zero-shot. Only VoxCeleb can be considered unseen. The zero-shot claim in the abstract and Section VI.B overstates the generalization evidence and should be corrected.
  3. [Tables II and III] The configuration labeled 'MLP ✓ ✓' in Table III reports identical accuracy and BERTScore values on VoxCeleb2 (Age 91.80, Gender 82.75, Ethnicity 88.38) as the CoLMbo row in Table II reports on VoxCeleb1-O. Since these are different test sets, identical values are implausible and suggest an error in dataset labels or copied entries. The authors must clarify whether the two tables refer to the same evaluation set.
  4. [Section III.C, Section V, Table III] The manuscript is internally inconsistent about whether the language model is frozen: Section III.C states the causal LM 'remains frozen during both training and inference,' while Section V describes 'full fine-tuning of all components for an additional 30 epochs' and Table III ablates 'Finetune LM.' This discrepancy affects the reproducibility of the architecture and the interpretation of the ablations.
  5. [Section VI, Tables II and III] The evaluation protocol for attribute accuracy is not described: the paper does not state how prompts were constructed for each dataset, how categorical predictions were extracted from generated text (e.g., exact match, string matching, LLM-as-judge), or why BERTScore is reported for classification tasks. Without this protocol, the comparison against Qwen Audio and the ablation numbers cannot be reproduced.
minor comments (5)
  1. [Eq. (3)] The text says 'N represents the number of text tokens,' but in Eq. (3) N is the sample index and the text-token index is l; this should be corrected to avoid confusion.
  2. [Figure 2] The caption 'Female African MandarinAmerican British Male Caucasian Asian' is missing spaces and punctuation, making it difficult to read; please format the attribute labels clearly.
  3. [References] Reference [26] is cited as 'Llama3' but points to the original Llama paper by Touvron et al.; the authors should cite the actual Llama3 release or correct the reference.
  4. [Abstract and Section VI] The phrase 'near-perfect accuracy' overstates the reported numbers, since gender and ethnicity accuracies are in the 82–88% range; a more measured characterization would be more appropriate.
  5. [Table III] The column header semantics are unclear: the rows list 'MLP' or 'Transformer' followed by check marks, but it is not clear which check mark corresponds to 'Speaker Loss' and which to 'Finetune LM'; please make the table self-explanatory.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: CoLMbo's architecture, training targets, and evaluations are externally grounded where the paper states its data sources, and no load-bearing claim reduces to a fitted constant or a self-citation chain.

full rationale

I walked the claimed derivation chain for CoLMbo: audio is parameterized by a fixed PDAF speaker encoder, mapped through a trainable ClipCap-style prefix mapper, and used to condition a frozen GPT-2 to generate speaker descriptions. The training targets come from Llama3-generated captions derived from EARS and TIMIT metadata, which is an external grounding step rather than a self-referential one. The speaker classification auxiliary loss is trained against speaker identities in EARS/TIMIT, again external supervision. The reported attribute accuracies on EARS and TIMIT can in principle be checked against the stated dataset metadata, and the VoxCeleb evaluation is presented as zero-shot with respect to a model trained on EARS/TIMIT, so the central empirical claim is not defined in terms of its own output. The paper does cite prior work by overlapping authors (PDAF encoder [21], prompting strategy [27], Pengi [6]), but none of these citations is load-bearing in a circular way: PDAF is a fixed external feature extractor, the prompting strategy merely produced training captions, and Pengi is a baseline comparison. The most serious weakness is that the provenance of VoxCeleb age and ethnicity ground-truth labels is never given, and TIMIT does not natively provide ethnicity labels; that is a correctness and reproducibility concern about unverifiable numbers, not a circularity in the derivation chain, because the paper does not state that those labels were derived from CoLMbo's own predictions or from the Llama3 captions used in training. Under the hard rules, I cannot flag a circular step without a quoted reduction, and no such reduction appears in the manuscript. Therefore the appropriate finding is no significant circularity, score 0.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central result rests on a frozen PDAF embedding retaining demographic attributes, synthetic captions standing in for ground truth, and unverified attribute labels for evaluation. Alpha, prefix lengths, mapper sizes, and augmentation ranges are hand-chosen, and some are load-bearing.

free parameters (8)
  • loss weight alpha = 0.3
    Chosen by hand in Eq. 2; balances captioning and speaker-id losses. Model sensitivity to it is not reported.
  • audio prefix length = 40 tokens
    Selected in Sections III-B and V; determines how much speaker information is passed to the language model.
  • text prompt prefix length = 10 tokens
    Fixed in Section V; a short prompt budget constrains the kinds of questions the model can answer.
  • mapper hidden dimension = 30720
    Taken from ClipCap [22]; an ad hoc architectural choice.
  • mapper transformer layers = 8
    Architecture choice for transformer mapper; the MLP variant works as well in ablation.
  • training schedule = 100 epochs frozen LM + 30 epochs full fine-tune
    Chosen in Section V; no sensitivity analysis is provided.
  • augmentation parameters = reverb 0.5-2s, noise 10-20 dB SNR, drop 500-2000 Hz, time cut 100-500ms
    Heuristic ranges; augmentations change age accuracy by 32.6 points on VoxCeleb1-O, so these choices are load-bearing.
  • learning rate and batch size = 1e-4, batch 64
    Standard optimizer settings; not swept.
assumptions (4)
  • domain assumption PDAF embeddings contain recoverable demographic information (gender, age, dialect, ethnicity) despite PDAF being trained only for speaker verification.
    The entire mapping depends on attribute info surviving in the fixed embedding; no analysis or experiment in the paper demonstrates this.
  • domain assumption Llama3-generated captions from metadata are a valid ground truth for speaker descriptions.
    TEARS captions are synthesized by an LLM using prompt strategies from [7,27]; no human validation or caption-quality statistics are reported.
  • domain assumption VoxCeleb and TIMIT provide reliable age and ethnicity labels for evaluation.
    The paper never cites VoxCeleb or states its label sources; TIMIT does not provide ethnicity labels, so this assumption is doubtful.
  • domain assumption A linear speaker classifier on the first prefix vector enforces speaker-level clustering that improves attribute prediction.
    Ablation shows removal of the speaker loss hurts several tasks, but the mechanism is asserted, not measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoLMbo: Speaker Language Model for Descriptive Profiling." pith.science (2026). https://pith.science/paper/REFRDJ3A

@misc{pith2026250609375,
  author       = {Pith},
  title        = {Pith review of: CoLMbo: Speaker Language Model for Descriptive Profiling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/REFRDJ3A}},
  note         = {Machine review of arXiv:2506.09375}
}
read the original abstract

Speaker recognition systems are often limited to classification tasks and struggle to generate detailed speaker characteristics or provide context-rich descriptions. These models primarily extract embeddings for speaker identification but fail to capture demographic attributes such as dialect, gender, and age in a structured manner. This paper introduces CoLMbo, a Speaker Language Model (SLM) that addresses these limitations by integrating a speaker encoder with prompt-based conditioning. This allows for the creation of detailed captions based on speaker embeddings. CoLMbo utilizes user-defined prompts to adapt dynamically to new speaker characteristics and provides customized descriptions, including regional dialect variations and age-related traits. This innovative approach not only enhances traditional speaker profiling but also excels in zero-shot scenarios across diverse datasets, marking a significant advancement in the field of speaker recognition.

Figures

Figures reproduced from arXiv: 2506.09375 by the authors.

Figure 1
Figure 1. depicts the architecture of the CoLMbo system. The modules of the model are described below. A. Audio parameterization Input audio are first parameterized into a Mel￾frequency spectro- gram. The spectrogram is input into a speaker encoder which derives a speaker embedding from it. In our work, we use the PDAF encoder [21], which extracts a 1024-dimensional speech embedding vector from each recording. The speaker enc… view at source ↗
Figure 2
Figure 2. CoLMbo Attention maps showing distinct activation patterns for speaker attributes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustrative examples of prompt-response for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of CoLMbo and Other ALMs in Generating Speaker Features on Unseen Speakers in a [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Stacked confusion plots for CoLMbo on TEARS, revealing bias in dialect and ethnicity predictions due [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Radar plot of CoLMbo’s accuracy for speaker [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Large Audio Language Models for Spoofing-Aware Speaker Verification

    cs.SD 2026-07 conditional novelty 6.0 of 10

    Adapted LALMs can reach competitive spoofing-aware speaker verification (89.3% accuracy, 0.19 min a-DCF on an ASVspoof5 subset), though zero-shot performance is near chance.

Reference graph

Works this paper leans on

29 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    Singh, Profiling humans from their voice

    R. Singh, Profiling humans from their voice . Springer, 2019, vol. 41

    work page 2019

  2. [2]

    SEGAA: A Unified Approach to Predicting Age, Gender, and Emotion in Speech

    A. R, I. Sigicharla, C. Periwal, M. K, N. D. P. S, S. Tiwari, and S. Arora, “Segaa: A unified approach to predicting age, gender, and emotion in speech,” 2024. [Online]. Available: https://arxiv.org/abs/2403.00887

    work page Pith review arXiv 2024

  3. [3]

    Prediction of age from speech features using a multi-layer perceptron model,

    S. Ravishankar, K. PrasannaKumarM., V . V . Patage, S. Tiwari, and S. Goyal, “Prediction of age from speech features using a multi-layer perceptron model,” 2020 11th International Conference on Computing, Communication and Networking Technologies (ICCCNT), pp. 1–6, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:224779262

    work page 2020

  4. [4]

    Salmonn: Towards generic hearing abilities for large language models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, M. Zejun, and C. Zhang, “Salmonn: Towards generic hearing abilities for large language models,” in The Twelfth Interna- tional Conference on Learning Representations

    work page

  5. [5]

    Qwen-audio: Advancing universal audio under- standing via unified large-scale audio-language models,

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

    arXiv 2023

  6. [6]

    Pengi: An audio language model for audio tasks,

    S. Deshmukh, B. Elizalde, R. Singh, and H. Wang, “Pengi: An audio language model for audio tasks,” Advances in Neural Information Processing Systems , vol. 36, pp. 18 090–18 108, 2023

    2023

  7. [7]

    Lis- ten, think, and understand,

    Y . Gong, H. Luo, A. H. Liu, L. Karlinsky, and J. Glass, “Lis- ten, think, and understand,” arXiv preprint arXiv:2305.10790 , 2023

    arXiv 2023

  8. [8]

    Gama: A large audio-language model with advanced audio under- standing and complex reasoning abilities,

    S. Ghosh, S. Kumar, A. Seth, C. K. R. Evuru, U. Tyagi, S. Sakshi, O. Nieto, R. Duraiswami, and D. Manocha, “Gama: A large audio-language model with advanced audio under- standing and complex reasoning abilities,” arXiv preprint arXiv:2406.11768, 2024

    arXiv 2024

Show all 29 references
  1. [9]

    Speaker recognition for multi-speaker conversations using x-vectors,

    D. Snyder, D. Garcia-Romero, G. Sell, A. McCree, D. Povey, and S. Khudanpur, “Speaker recognition for multi-speaker conversations using x-vectors,” in ICASSP. IEEE, 2019, pp. 5796–5800

    2019

  2. [10]

    Front-end factor analysis for speaker verification,

    N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front-end factor analysis for speaker verification,” in IEEE Transactions on Audio, Speech, and Language Processing , vol. 19, no. 4, 2011, pp. 788–798

    2011

  3. [11]

    Deep neural networks for small foot- print text-dependent speaker verification,

    E. Variani, X. Lei, E. McDermott, I. L. Moreno, and J. Gonzalez-Dominguez, “Deep neural networks for small foot- print text-dependent speaker verification,” in ICASSP, 2014, pp. 4052–4056

    2014

  4. [12]

    Generalized end-to-end loss for speaker verification,

    L. Wan, Q. Wang, A. Papir, and I. L. Moreno, “Generalized end-to-end loss for speaker verification,” in ICASSP, 2018, pp. 4879–4883

    2018

  5. [13]

    Front-end factor analysis for speaker verification,

    N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front-end factor analysis for speaker verification,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 19, no. 4, pp. 788–798, 2010

    2010

  6. [14]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C. C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, and Y . Wu, “Conformer: Convolution-augmented transformer for speech recognition,” in Interspeech, 2020, pp. 5036–5040

    2020

  7. [15]

    Mfa-conformer: Multi-scale feature aggregation conformer for automatic speaker verification,

    Y . Zhang, Z. Lv, H. Wu, S. Zhang, P. Hu, Z. Wu, H. Y . Lee, and H. Meng, “Mfa-conformer: Multi-scale feature aggregation conformer for automatic speaker verification,” in Interspeech, 2022, pp. 70–74

    2022

  8. [16]

    A novel scheme for speaker recognition using a phonetically-aware deep neural network,

    Y . Lei, N. Scheffer, L. Ferrer, and M. McLaren, “A novel scheme for speaker recognition using a phonetically-aware deep neural network,” in ICASSP, 2014, pp. 1695–1699

    2014

  9. [17]

    Employing phonetic information in dnn speaker embeddings to improve speaker recognition performance,

    M. H. Rahman, I. Himawan, M. McLaren, C. Fookes, and S. Sridharan, “Employing phonetic information in dnn speaker embeddings to improve speaker recognition performance,” in Interspeech, 2018, pp. 3593–3597

    2018

  10. [18]

    Speaker em- bedding extraction with phonetic information,

    Y . Liu, L. He, J. Liu, and M. T. Johnson, “Speaker em- bedding extraction with phonetic information,” arXiv preprint arXiv:1804.04862, 2018

    2018 arXiv

  11. [19]

    Gender and age estimation methods based on speech using deep neural networks,

    D. Kwasny and D. Hemmerling, “Gender and age estimation methods based on speech using deep neural networks,” Sensors, vol. 21, no. 14, p. 4785, 2021. [Online]. Available: https://www.mdpi.com/1424-8220/21/14/4785

    2021

  12. [20]

    Explainable attribute- based speaker verification,

    X. Wu, C. Luu, P. Bell, and A. Rajan, “Explainable attribute- based speaker verification,” arXiv preprint arXiv:2405.19796 , 2024

    2024 arXiv

  13. [21]

    Pdaf: A phonetic debiasing attention framework for speaker verification,

    M. Baali, A. Aldoobi, H. Dhamyal, R. Singh, and B. Raj, “Pdaf: A phonetic debiasing attention framework for speaker verification,” in SLT, 2024

    2024

  14. [22]

    Clipcap: Clip prefix for image captioning,

    R. Mokady, A. Hertz, and A. H. Bermano, “Clipcap: Clip prefix for image captioning,” arXiv preprint arXiv:2111.09734, 2021

    2021 arXiv

  15. [23]

    Multimodal few-shot learning with frozen lan- guage models,

    M. Tsimpoukelli, J. L. Menick, S. Cabi, S. Eslami, O. Vinyals, and F. Hill, “Multimodal few-shot learning with frozen lan- guage models,” Advances in Neural Information Processing Systems, vol. 34, pp. 200–212, 2021

    2021

  16. [24]

    EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,

    J. Richter, Y .-C. Wu, S. Krenn, S. Welker, B. Lay, S. Watanabe, A. Richard, and T. Gerkmann, “EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,” in Interspeech, 2024

    2024

  17. [25]

    Darpa timit acoustic-phonetic continuous speech corpus cd-rom. nist speech disc 1-1.1,

    J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren, “Darpa timit acoustic-phonetic continuous speech corpus cd-rom. nist speech disc 1-1.1,” NASA STI/Recon Technical Report N, vol. 93, p. 27403, 1993

    1993

  18. [26]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar et al., “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971 , 2023

    2023 arXiv

  19. [27]

    Audio entailment: Assessing deductive reasoning for audio understanding,

    S. Deshmukh, S. Han, H. Bukhari, B. Elizalde, H. Gam- per, R. Singh, and B. Raj, “Audio entailment: Assessing deductive reasoning for audio understanding,” arXiv preprint arXiv:2407.18062, 2024

    2024 arXiv

  20. [28]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” ICLR, 2015

    2015

  21. [29]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al., “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.