Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Inter-Speaker Relative Cues for Text-Guided Target Speech Extraction

T0 review · 4 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 Inter-speaker relative cues — how the target differs from the interfering speaker — are enough to guide text-based target speech extraction, avoiding fixed attribute categories.

desk verdict The relative-cue formulation and the 100k-mixture dataset are genuinely new and useful; the robustness claims, however, are not backed by stratified results and there is no fixed-attribute baseline. read the letter →

arxiv 2506.01483 v3 pith:5767CA7M submitted 2025-06-02 eess.AS cs.SD

Wang Dai , Archontis Politis , Tuomas Virtanen This is my paper
classification eess.AScs.SD
keywords targetspeechextractioninter-speakerrelativecuestext-guidedseparationmultilingualmixturesattributespre-trainedencodersnaturallanguagepromptsreverberant
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes that a text prompt for target speech extraction should describe how the target speaker differs from the interfering speaker, rather than assigning each speaker an absolute label like "high pitch" or "middle-aged." It builds a five-language, two-speaker mixture dataset in which continuous attributes such as pitch, speaking rate, duration, loudness, distance, age, and temporal order are converted into relative cues (higher/lower, faster/slower, first/second, and so on) when their measured difference passes an auditory threshold, while discrete attributes such as gender, language, and emotion stay categorical. Across twelve cues, the paper reports that combining all cues gives better SI-SDR improvement than random subsets, and that gender and temporal order are the most reliable cues across languages and reverberant conditions. If correct, the result means dataset builders can expand text-guided extraction by adding new corpora and recomputing pairwise comparisons instead of re-annotating every utterance with fixed categories.

What carries the argument

The load-bearing object is the relative-cue label, a text token generated by comparing one attribute of target and interference speech. Continuous attributes are converted through thresholds from hearing studies into three-way labels (higher/lower/similar, faster/slower/similar, etc.), discrete attributes into same-versus-category labels, and only the 'different' cues are written into a templated natural-language prompt. The same comparison is used at training and inference, and a fine-tuned instruction-tuned language model embeds the prompt while a dual-path transformer mask network, optionally fed with pre-trained CNN features from overlapped-speech representation learning, separates the mixture under an SI-SDR loss.

What would settle it

Train the same model on mixtures where attribute differences lie just on either side of a threshold (e.g., 4.9 vs 5.1 Hz pitch difference) in a held-out language or reverberation condition, and measure per-cue SI-SDRi; if 'similar'-labeled pairs are clearly extractable or 'different'-labeled pairs are not, the audibility thresholds are false. Equally decisive would be changing a threshold (say, 10 Hz instead of 5 Hz) and seeing whether the ranking of cue effectiveness changes substantially.

Watch

Extended reading notes

Core claim

The central claim is that inter-speaker relative cues are sufficient to specify and extract a target voice from a mixture, and that they are more flexible than fixed speech-attribute categories. The authors construct twelve cues by comparing target and interference speech, grouping continuous differences into discrete labels only when the difference exceeds an auditory threshold (e.g., 5 Hz for pitch level, 15% for speaking rate, 0.1 s for temporal order); discrete attributes keep their category labels when the two speakers differ. Trained on 100,000 two-speaker mixtures spanning five languages with reverberation, the model achieves 11.4 dB SI-SDRi when all cues are given at inference, versus 10.1 dB with random cue subsets, and gender and temporal order reach 11.0 and 10.4 dB respectively, appearing reliable across language and room acoustics. The authors take this as evidence that the relative formulation, not any single absolute attribute, carries most of the discriminative information, and that richer cue combinations help rather than confuse the model.

Load-bearing premise

The weakest assumption is that the chosen thresholds (5 Hz of pitch, 15% speaking rate, 0.1 seconds of timing, etc.) mark audible differences in the same way for every language and every reverberant room; if a threshold is wrong, cues labeled 'similar' may actually discriminate speakers, or cues labeled 'different' may be imperceptible, and the measured cue effectiveness collapses.

Editorial extensions

If this is right

  • Expanding the dataset to a new language needs only a new single-speaker corpus and pairwise attribute comparisons, not a re-annotation of every utterance's absolute attributes.
  • Supplying all relevant cues at inference is better than a random subset: 11.4 vs 10.1 dB SI-SDRi in the proposed model, confirming that more contrastive information helps.
  • Gender and temporal order survive cross-lingual and reverberant conditions, so system designers can rely on them when only minimal text is available.
  • Cues such as pitch level, loudness, distance, speaking duration, and language each give near or above 3 dB gains, while age category is the only cue that hurts separation.
  • Fine-tuning the pre-trained CNN encoder raises SI-SDRi and PESQ across nearly all cue conditions, so representation pre-training is a practical component of relative-cue target speech extraction.

Reading between the lines

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

  • A natural next experiment the paper does not run is to keep the same model but replace thresholded labels with raw continuous differences (e.g., 'pitch 32 Hz higher') and compare; equal performance would show the thresholds are not load-bearing.
  • The relative-cue principle should transfer to non-speech separation, such as isolating a musical instrument or sound event by contrast with co-occurring sounds, since the target description is relational rather than absolute.
  • The negative age-cue result may be an artifact of scarce, emotionally colored age-annotated data rather than a property of age perception; more age-balanced corpora would settle it.
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

4 major / 5 minor

Summary. This paper proposes a text-guided target speech extraction (TSE) approach built on inter-speaker relative cues. It constructs a two-speaker, five-language dataset with attributes such as gender, temporal order, pitch level, loudness, and speaking rate, converted into relative descriptions (e.g., "the female speaker with a higher pitch level") rather than absolute categorical labels. The authors train an encoder-mask-decoder model with a fine-tuned LLaMA-3.2-1B text encoder and optionally fine-tune WavLM Base+ CNN encoders, and report SI-SDRi and PESQ for individual cues, random cue subsets, and all cues combined. The main claims are that relative cues outperform fixed speech attribute classification, that gender and temporal order are the most robust cues across languages and reverberation, and that additional cues provide notable benefits.

Significance. If the central claim were fully validated, the relative-cue formulation would be a practical contribution to text-guided TSE, because it reduces the need for fine-grained absolute attribute estimation and simplifies dataset expansion. The paper has clear strengths: a purpose-built multilingual dataset with released creation code, a broad attribute set, and a reproducible training pipeline. However, the experiments only rank relative cues against each other; the stated advantage over fixed attribute classification is never tested, and the robustness claim rests on pooled results. The core idea is promising, but the evidence provided is incomplete for the conclusions drawn.

major comments (4)
  1. [Section 3.3, Table 3] The paper's central claim—that inter-speaker relative cues "overcome the fixed categorization of speech attributes and make the dataset much easier to expand" (Abstract, Section 1)—is not tested. The only comparisons are among relative cue configurations (individual, random, all). There is no baseline using fixed, absolute attribute labels (e.g., "female speaker with high pitch level" based on single-speaker statistics), which is the approach used in LLM-TSE [9] and TextrolMix [10] that the paper argues against. Without such a baseline, the results support an internal ranking of relative cues but not the claimed advantage over fixed categorization. I would expect at least one fixed-attribute version of the same dataset and architecture, or a direct comparison with an existing text-guided TSE method, to support the headline claim.
  2. [Section 3.3, Table 3] The claim that gender and temporal order are "most robust across languages and reverberant conditions" is supported only by pooled averages over the entire mixed test set. No per-language or per-reverberation-time breakdown is reported, so a cue could be excellent in English and Chinese but ineffective in French or under higher RT60 without changing the pooled numbers. The paper's own Section 4 states that future work will "assess generalization across different languages and reverberation conditions," which acknowledges that this assessment has not been done. Please report stratified results by language and RT60, or revise the robustness claim to refer only to the pooled test set.
  3. [Section 2.1, Table 1] The correctness of all relative-cue labels depends on the hand-set auditory thresholds (5 Hz, 15%, 0.5 m, 3 dB, 0.1 s, etc.). The temporal order threshold is justified solely by "our subjective listening test," but no details of that test (number of listeners, stimuli, procedure) are given. No sensitivity analysis is reported for any threshold. If a threshold is wrong for a particular language or acoustic condition, labels such as "similar" may actually be discriminative, or "higher/lower" may be imperceptible, which would change the measured cue effectiveness. At minimum, the listening test should be described and a sensitivity analysis over threshold values should be provided for the main cues.
  4. [Section 3.3, Table 3] The results are from a single training run per configuration; no error bars, repeated seeds, or statistical tests are reported. Differences between cues are used to rank them (e.g., Language at 6.4 dB vs. Pitch Range at 3.0 dB in the proposed model), and even smaller gaps appear among secondary cues. Without repeated-seed variation, the ranking of "notable benefits" may not be stable. Please provide multiple seeds with standard deviations or a significance test for the key pairwise comparisons.
minor comments (5)
  1. [Section 2.1] The sentence describing syllable counting contains a punctuation and notation issue: "syllable , while" has a spacing error, and "vowel letters such as 'aeiou'" should be written as "vowel letters (e.g., a, e, i, o, u)".
  2. [Section 2.2(iii)] The mixing rule is described informally; the phrase "start offset from(0, duration of longer sample−duration of shorter sample)" has a missing space, and the relationship between the overlap duration and the padding applied to the two samples is not fully specified for the case where a sample is shorter than 3 seconds.
  3. [Table 1] The speaking-rate threshold is set to 15% while the cited just-noticeable-difference references are 5–10%; a brief explanation of why the threshold is set above the JND would avoid a perceived inconsistency.
  4. [Section 3.3] The abbreviation PESQ is used without being defined at first use; please define it in Section 2 or at the start of Section 3.3.
  5. [Table 3] The negative SI-SDRi values for Age Category (e.g., −3.2 dB in Baseline) mean this cue actively degrades performance relative to the unprocessed mixture; the text should state this explicitly rather than only saying the cue "performs poorly."

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: every central claim is a direct empirical measurement on a purpose-built dataset, with no fitted parameter renamed as a prediction and no load-bearing self-citation.

full rationale

The paper makes no derivation-style claim that reduces by construction to its own inputs. Relative cues are defined by comparing target and interference attributes against thresholds (Table 1), most of which are taken from cited hearing studies; the temporal-order threshold comes from the authors' own listening test, but it is an input label-definition choice, not a parameter fitted to the reported SI-SDRi/PESQ outcomes. The headline results ('combining all relative cues yields better performance than random subsets', 'gender and temporal order being the most robust') are direct comparisons in Table 3 of models trained on the same cue-labeling scheme, so the evaluation is self-contained rather than circular. The authors cite prior work for WavLM, DPT, and LLM components, but none of those citations carries the paper's conclusion, and none is a self-citation. The manuscript's own acknowledged weaknesses—age-category underperformance due to limited annotated data, the unvalidated 0.1 s temporal-order threshold, and the absence of per-language/per-reverberation breakdowns—are correctness and generalizability concerns, not circularity: they do not show that a stated result is equivalent to its inputs. The 'robust across languages' claim rests on pooled averages, which is an evidentiary weakness that could undermine the strength of the conclusion, but it is not a logical or definitional circularity. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on eight hand-set cue thresholds, several annotation and feature-extraction assumptions, and the learnability of LLM conditioning. The paper introduces no new physical or architectural entity; 'inter-speaker relative cues' is a labeling scheme, not a postulated mechanism.

free parameters (8)
  • Speaking rate threshold = 15%
    Differences above 15% are labeled faster/slower; below are similar. Cited hearing studies give 5-10% noticeable, but the paper does not test sensitivity.
  • Speaking duration threshold = 15%
    Differences above 15% are labeled longer/shorter; below are similar. No sensitivity analysis is reported.
  • Pitch level threshold = 5 Hz
    Differences above 5 Hz are labeled higher/lower; below are similar. Cited as 3-5 Hz perceptible, chosen value is at the upper edge.
  • Pitch range threshold = 25%
    Differences above 25% are labeled wider/narrower; below are similar. Cited as 10-20% perceptible, chosen value is larger.
  • Distance threshold = 0.5 m
    Differences above 0.5 m are labeled farther/nearer; below are similar. Cited as 20-30 cm noticeable, chosen value is larger.
  • Age category threshold = 10 years
    Differences above 10 years are labeled older/younger; below are similar. Cited as 5-10 years discernible, chosen value is at the upper edge.
  • Loudness threshold = 3 dB
    Differences above 3 dB are labeled louder/quieter; below are similar. Cited as 3 dB noticeable.
  • Temporal order threshold = 0.1 s
    Onset differences above 0.1 s are labeled first/second; below are similar. Justified only by the authors' subjective listening test.
assumptions (5)
  • domain assumption Auditory thresholds from hearing studies transfer to simulated reverberant mixtures across five languages.
    Used in Section 2.1 to discretize continuous differences; if thresholds do not transfer, cue labels are noisy.
  • domain assumption The twelve speech attributes are estimated accurately enough from corpus annotations, pYIN, Montreal Forced Aligner, and Silero VAD.
    Errors in mean F0, duration, or pause detection propagate into relative-cue labels and then into the measured cue effectiveness.
  • domain assumption The LLM (LLaMA 3.2 1B) can convert relative-cue text prompts into embeddings that identify the target speaker.
    The whole method relies on this mapping being learnable with LoRA fine-tuning; the paper does not ablate the LLM.
  • domain assumption Splitting each corpus into two non-overlapping parts prevents speaker and content leakage between training and test mixtures.
    Stated in Section 2.2(i); leakage would inflate measured performance.
  • domain assumption Two-speaker mixtures with simulated RT60 0.3-0.6 s and SIR -6 to 6 dB represent the 'reverberant conditions' claimed in the conclusions.
    The paper generalizes to reverberant conditions from this narrow simulated range.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inter-Speaker Relative Cues for Text-Guided Target Speech Extraction." pith.science (2026). https://pith.science/paper/5767CA7M

@misc{pith2026250601483,
  author       = {Pith},
  title        = {Pith review of: Inter-Speaker Relative Cues for Text-Guided Target Speech Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5767CA7M}},
  note         = {Machine review of arXiv:2506.01483}
}
read the original abstract

We propose a novel approach that utilizes inter-speaker relative cues to distinguish target speakers and extract their voices from mixtures. Continuous cues (e.g., temporal order, age, pitch level) are grouped by relative differences, while discrete cues (e.g., language, gender, emotion) retain their categorical distinctions. Compared to fixed speech attribute classification, inter-speaker relative cues offer greater flexibility, facilitating much easier expansion of text-guided target speech extraction datasets. Our experiments show that combining all relative cues yields better performance than random subsets, with gender and temporal order being the most robust across languages and reverberant conditions. Additional cues, such as pitch level, loudness, distance, speaking duration, language, and pitch range, also demonstrate notable benefits in complex scenarios. Fine-tuning pre-trained WavLM Base+ CNN encoders improves overall performance over the Conv1d baseline.

Figures

Figures reproduced from arXiv: 2506.01483 by the authors.

Figure 1
Figure 1. A block diagram of the used target speech extraction system 3.2. Training and Network Configurations We adopted a relative large kernel size of 80 and stride of 40 for both the Conv1d encoder and ConvTranspose1d decoder, simi￾lar to [8], which proved efficient in preliminary experiments. To match the time resolution of the Conv1d encoder, only the first four temporal convolution blocks of WavLM Base+ CNN en￾coders w… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

  1. [9]

    Selective listening by syn- chronizing speech with lips,

    Z. Pan, R. Tao, C. Xu, and H. Li, “Selective listening by syn- chronizing speech with lips,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 1650–1664, 2022

    work page 2022

  2. [10]

    Separate anything you describe,

    X. Liu, Q. Kong, Y . Zhao, H. Liu, Y . Yuan, Y . Liu, R. Xia, Y . Wang, M. D. Plumbley, and W. Wang, “Separate anything you describe,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, 2024

    2024

  3. [1]

    Introduction Target speech extraction (TSE) methods aim at isolating a spe- cific speaker’s voice from a multi-talker mixture. This process uses cues associated with the desired speaker, such as a pre- recorded enrollment speech that highlights the speaker’s vocal characteristics [1, 2], a spatial cue indicating the direction from which the speaker is spe...

    work page

  4. [2]

    Inter-Speaker Relative Cues for Text-Guided Target Speech Extraction

    Dataset Construction In this study, mixture signals consist of speech from two speak- ers: one as the target and the other as interference. We construct relative cues from various speech attributes to specify and ex- tract the target speaker’s voice. These cues are formatted as nat- ural language descriptions (text prompts) for target speech ex- traction....

    work page Pith review arXiv 2025

  5. [3]

    Please extract the female speaker with a quieter loudness in the audio

    Experimental Setup 3.1. Proposed Method and Baseline The architecture of our proposed method is depicted in Figure 1. We adopt the conventional encoder-mask-decoder framework for speech separation, where the text prompt, describing cues of the target speaker, guides the extraction of the target speaker from the mixture speech. To enhance learning of mixed...

    work page

  6. [4]

    Experimental results suggest the effectiveness of incorporating multiple cues and leveraging pre-trained representations from WavLM Base+ CNN encoders

    Conclusions and Future Work This study explored rich relative cues between speakers for text- guided target speech extraction. Experimental results suggest the effectiveness of incorporating multiple cues and leveraging pre-trained representations from WavLM Base+ CNN encoders. The findings also reveal that different cues have varying impacts and contribu...

    work page

  7. [5]

    V oice- Filter: Targeted V oice Separation by Speaker-Conditioned Spec- trogram Masking,

    Q. Wang, H. Muckenhirn, K. Wilson, P. Sridhar, Z. Wu, J. R. Her- shey, R. A. Saurous, R. J. Weiss, Y . Jia, and I. L. Moreno, “V oice- Filter: Targeted V oice Separation by Speaker-Conditioned Spec- trogram Masking,” in Proc. INTERSPEECH , 2019, pp. 2728– 2732

    work page 2019

  8. [6]

    Speakerbeam: Speaker aware neural network for target speaker extraction in speech mixtures,

    K. ˇZmol´ıkov´a, M. Delcroix, K. Kinoshita, T. Ochiai, T. Nakatani, L. Burget, and J. ˇCernock`y, “Speakerbeam: Speaker aware neural network for target speaker extraction in speech mixtures,” IEEE Journal of Selected Topics in Signal Processing , vol. 13, no. 4, pp. 800–814, 2019

    2019

Show all 36 references
  1. [7]

    L- spex: Localized target speaker extraction,

    M. Ge, C. Xu, L. Wang, E. S. Chng, J. Dang, and H. Li, “L- spex: Localized target speaker extraction,” in IEEE International Conference on Acoustics, Speech and Signal Processing , 2022, pp. 7287–7291

    2022

  2. [8]

    Multimodal speakerbeam: Single channel target speech extrac- tion with audio-visual speaker clues

    T. Ochiai, M. Delcroix, K. Kinoshita, A. Ogawa, and T. Nakatani, “Multimodal speakerbeam: Single channel target speech extrac- tion with audio-visual speaker clues.” in INTERSPEECH, 2019, pp. 2718–2722

    2019

  3. [11]

    Target sound extraction with variable cross- modality clues,

    C. Li, Y . Qian, Z. Chen, D. Wang, T. Yoshioka, S. Liu, Y . Qian, and M. Zeng, “Target sound extraction with variable cross- modality clues,” in IEEE International Conference on Acoustics, Speech and Signal Processing, 2023, pp. 1–5

    2023

  4. [12]

    Listen, chat, and edit: Text-guided soundscape modification for enhanced auditory experience,

    X. Jiang, C. Han, Y . A. Li, and N. Mesgarani, “Listen, chat, and edit: Text-guided soundscape modification for enhanced auditory experience,” arXiv preprint arXiv:2402.03710, 2024

    2024 arXiv

  5. [13]

    Typing to listen at the cocktail party: Text-guided target speaker extraction,

    X. Hao, J. Wu, J. Yu, C. Xu, and K. C. Tan, “Typing to listen at the cocktail party: Text-guided target speaker extraction,” arXiv preprint arXiv:2310.07284, 2023

    2023 arXiv

  6. [14]

    Beyond speaker identity: Text guided target speech ex- traction,

    M. Huo, A. Jain, C. P. Huynh, F. Kong, P. Wang, Z. Liu, and V . Bhat, “Beyond speaker identity: Text guided target speech ex- traction,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing, 2025, pp. 1–5

    2025

  7. [15]

    Textrolspeech: A text style control speech cor- pus with codec language text-to-speech models,

    S. Ji, J. Zuo, M. Fang, Z. Jiang, F. Chen, X. Duan, B. Huai, and Z. Zhao, “Textrolspeech: A text style control speech cor- pus with codec language text-to-speech models,” in IEEE Inter- national Conference on Acoustics, Speech and Signal Processing, 2024, pp. 10 301–10 305

    2024

  8. [16]

    Some effects of speaking rate on phonetic percep- tion,

    J. L. Miller, “Some effects of speaking rate on phonetic percep- tion,” Phonetica, vol. 38, no. 1-3, pp. 159–180, 1981

    1981

  9. [17]

    On the just noticeable difference for tempo in speech,

    H. Quen ´e, “On the just noticeable difference for tempo in speech,” Journal of Phonetics, vol. 35, no. 3, pp. 353–362, 2007

    2007

  10. [18]

    B. C. Moore, An Introduction to the Psychology of Hearing . Brill, 2012

    2012

  11. [19]

    The psychophysics of human sound localization,

    J. Blauert and S. Hearing, “The psychophysics of human sound localization,” in Spatial Hearing. MIT Press Cambridge, MA, USA, 1997

    1997

  12. [20]

    Linville, Vocal Aging

    S. Linville, Vocal Aging. Singular Thomson Learning, 2001

    2001

  13. [21]

    Age perceptions and eval- uative reactions toward adult speakers,

    E. B. Ryan and H. L. Capadano III, “Age perceptions and eval- uative reactions toward adult speakers,” Journal of Gerontology, vol. 33, no. 1, pp. 98–102, 1978

    1978

  14. [22]

    Zwicker and H

    E. Zwicker and H. Fastl, Psychoacoustics: Facts and Models . Springer Science & Business Media, 2013, vol. 22

    2013

  15. [23]

    Design of speech corpus for mandarin text to speech,

    J. Zhang and H. Jia, “Design of speech corpus for mandarin text to speech,” in The Blizzard Challenge 2008 Workshop, 2008

    2008

  16. [24]

    Emotional voice con- version: Theory, databases and esd,

    K. Zhou, B. Sisman, R. Liu, and H. Li, “Emotional voice con- version: Theory, databases and esd,”Speech Communication, vol. 137, pp. 1–18, 2022

    2022

  17. [25]

    Aishell-3: A multi- speaker mandarin tts corpus,

    Y . Shi, H. Bu, X. Xu, S. Zhang, and M. Li, “Aishell-3: A multi- speaker mandarin tts corpus,” inINTERSPEECH, 2021, pp. 2756– 2760

    2021

  18. [26]

    Magicdata mandarin chinese read speech corpus,

    OpenSLR, “Magicdata mandarin chinese read speech corpus,” https://openslr.org/68/, 2019, accessed: 2025-06-08

    2019

  19. [27]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in IEEE International Conference on Acoustics, Speech and Signal Processing, 2015, pp. 5206–5210

    2015

  20. [28]

    French emo- tional speech database - or´eau,

    L. Kerkeni, C. Cleder, Y . Serrestou, and K. Raouf, “French emo- tional speech database - or´eau,” Dec. 2020, Zenodo

    2020

  21. [29]

    Mls: A large-scale multilingual dataset for speech research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “Mls: A large-scale multilingual dataset for speech research,” ArXiv, vol. abs/2012.03411, 2020

    2012 arXiv

  22. [30]

    A database of german emotional speech

    F. Burkhardt, A. Paeschke, M. Rolfes, W. F. Sendlmeier, B. Weiss et al. , “A database of german emotional speech.” in INTER- SPEECH, vol. 5, 2005, pp. 1517–1520

    2005

  23. [31]

    Emo- matchspanishdb: study of speech emotion recognition machine learning models in a new spanish elicited database,

    E. Garcia-Cuesta, A. B. Salvador, and D. G. P ˜aez, “Emo- matchspanishdb: study of speech emotion recognition machine learning models in a new spanish elicited database,” Multimedia Tools and Applications, vol. 83, no. 5, pp. 13 093–13 112, 2024

    2024

  24. [32]

    gpurir: A python library for room impulse response simulation with gpu acceler- ation,

    D. Diaz-Guerra, A. Miguel, and J. R. Beltran, “gpurir: A python library for room impulse response simulation with gpu acceler- ation,” Multimedia Tools and Applications , vol. 80, no. 4, pp. 5653–5671, 2021

    2021

  25. [33]

    Target speech extraction with pre-trained self- supervised learning models,

    J. Peng, M. Delcroix, T. Ochiai, O. Plchot, S. Araki, and J. ˇCernock`y, “Target speech extraction with pre-trained self- supervised learning models,” in IEEE International Conference on Acoustics, Speech and Signal Processing , 2024, pp. 10 421– 10 425

    2024

  26. [34]

    Attention is all you need in speech separation,

    C. Subakan, M. Ravanelli, S. Cornell, M. Bronzi, and J. Zhong, “Attention is all you need in speech separation,” inICASSP 2021- 2021 IEEE International Conference on Acoustics, Speech and Signal Processing, 2021, pp. 21–25

    2021

  27. [35]

    Sdr–half- baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “Sdr–half- baked or well done?” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing , 2019, pp. 626–630

    2019

  28. [36]

    Perceptual eval- uation of speech quality (PESQ)-a new method for speech qual- ity assessment of telephone networks and codecs,

    A. Rix, J. Beerends, M. Hollier, and A. Hekstra, “Perceptual eval- uation of speech quality (PESQ)-a new method for speech qual- ity assessment of telephone networks and codecs,” in 2001 IEEE International Conference on Acoustics, Speech, and Signal Pro- cessing, vol. 2, 2001,...

    2001

Pith tools

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