Pith. sign in

REVIEW 4 major objections 6 minor 40 references

Task-Specific Audio Coding for Machines: Machine-Learned Latent Features Are Codes for That Machine

T0 review · 4 major / 6 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 Quantizing a model's own hidden audio features into residual vector-quantization codes, then retraining that model on the codes, keeps ASR and audio-classification performance near baseline at bitrates of a few hundred bits per second.

desk verdict Sound ACoM idea with a fixable evaluation; the ultra-low-bitrate headline rests on test-set entropy and internally inconsistent numbers. read the letter →

arxiv 2507.12701 v1 pith:LYRY7COK submitted 2025-07-17 cs.SD cs.AIeess.AS

Anastasia Kuznetsova , Inseon Jang , Wootaek Lim , Minje Kim This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords audiocodingformachinestask-specificcompressionresidualvectorquantizationautomaticspeechrecognitionclassificationmodelsplittingdiscretetokensneuralcodec
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 a way to compress audio not for human listening but for the machine that will consume it: take a pretrained speech or audio model, quantize one of its internal feature layers with residual vector quantization, and retrain the rest of the model on the quantized codes. The claim is that this keeps downstream accuracy close to the unquantized baseline while transmitting only a few hundred bits per second, far below what a waveform codec would need, and that the resulting discrete tokens are codes for that specific task. On LibriSpeech ASR, a two-codebook system at 500 bps raw reaches 2.21% word error rate versus 2.01% for the continuous baseline; on UrbanSound8k audio classification, a one-codebook system at 200 bps raw reaches about 80% accuracy, matching the continuous baseline. The practical interest is that a separate neural audio codec becomes unnecessary: the early layers of the model act as the encoder on the device, the codebook indices travel to the cloud, and the remaining layers finish the task.

What carries the argument

The carrying mechanism is residual vector quantization inserted at a chosen hidden layer. The encoder is the first M layers of the downstream model, whose output h_M is replaced by the sum of nearest codewords from K codebooks, C_{:,v*_1,1} + ... + C_{:,v*_K,K}; only the K indices v* are transmitted. Training minimizes a task loss (CTC plus label-smoothing KL for ASR, additive angular margin softmax for audio classification) together with codebook and commitment losses, so the codes are shaped by what the task needs rather than by waveform reconstruction. Bitrate is reported two ways: raw bitrate R*K*ceil(log2 V), and an entropy lower bound R*H(C) computed from the empirical codeword frequency on the test set, which the paper motivates by citing Huffman-style entropy coding.

What would settle it

Train the same quantized ASR and audio-classification models, build a Huffman or arithmetic coder from codeword frequencies measured only on the training split, and count actual transmitted bits plus WER or accuracy on a held-out or out-of-domain set; if the real bitrate exceeds the reported entropy bitrate or the accuracy gap widens, the central compression claim is weakened.

Watch

Extended reading notes

Core claim

The central discovery is that a quantized intermediate representation of an already trained downstream model is itself an efficient code for that model, provided the quantization is trained jointly with the task objective and codebook regularizers. The authors show that residual vector quantization (RVQ) applied to a chosen layer's output can compress the features so aggressively that the transmitted token stream is on the order of 200 to 500 bits per second, with word error rate or classification accuracy staying within a small margin of the continuous-feature model. At the same time, the conventional pipeline of encoding the waveform with a high-fidelity neural codec, transmitting, reconstructing, and then running the downstream model is shown to be both more compute-heavy and worse for noisy speech (ASR test-other 4.99% versus 7.12%). The paper argues this makes the approach architecture-independent: because any internal layer can be quantized, the split between on-device encoder and cloud model can be adjusted to match device compute constraints.

Load-bearing premise

The ultra-low bitrate numbers rely on entropy computed from test-set codeword usage with no actual entropy coder implemented, and the paper assumes that 'appropriate regularization' will make the method work for any chosen layer or architecture.

Editorial extensions

If this is right

  • Edge-to-cloud audio pipelines can drop the standalone codec entirely: the early layers of the task model encode and quantize on-device, and only codebook indices are transmitted.
  • Machine-consumed audio can operate at bitrates of a few hundred bits per second, and near 130 bps with entropy coding in the one-codebook ASR case, roughly two to three orders of magnitude below typical waveform codec rates.
  • The quantization layer choice becomes a tunable complexity knob: deeper quantization improves accuracy (for example, ASR test-other improves from 5.33% at layer 2 to 4.99% at layer 8) at the cost of more on-device computation.
  • The paper presents the approach as transferable to any task and architecture that can be split into two parts, as long as the quantization is regularized appropriately.

Reading between the lines

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

  • If codeword distributions are stable across domains, entropy-coded bitrates may routinely land below 200 bps, but the paper's entropy estimates are measured on test data; an actual coder trained on training data could reveal a real-world bitrate premium.
  • The same recipe of quantizing a model's own features may transfer to non-audio modalities such as video or sensor streams, where a pretrained task model's internal layer is the natural code space.
  • A testable extension is to measure robustness under distribution shift: retrain the tokenizer on one corpus and evaluate WER or accuracy on out-of-domain audio, checking both bitrate and task performance.
  • Because only task-specific codes travel over the wire, the scheme implicitly hides raw audio from the cloud; quantifying the privacy or information-leakage properties of these codes is a natural follow-up the paper does not address.
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 / 6 minor

Summary. The manuscript proposes a task-specific audio coding for machines (ACoM) method in which an intermediate feature representation of a pretrained downstream model is quantized with residual vector quantization (RVQ), and the remaining layers are finetuned with a task loss plus codebook/commitment losses. The method is evaluated on LibriSpeech ASR with a Conformer-Transformer and on UrbanSound8k audio classification with an ECAPA-TDNN model, reporting WER 2.21 vs. 2.01 (clean) and 4.99 vs. 4.52 (other) at a raw bitrate of 500 bps, and classification accuracy 80.2% vs. 79.7% at a raw 200 bps, with entropy-based bitrates as low as 131.62 bps. The paper also compares on-device GMACs against a conventional pipeline that uses DAC as a neural codec.

Significance. If the bitrate claims were validated by an implemented entropy coder, the paper would be a useful contribution to ACoM and split inference: it reuses the downstream model's own early layers as the encoder, avoids a standalone codec, and shows that task-guided RVQ can preserve downstream accuracy at very low bitrate. Strengths include the clear formulation in Eqs. (1)-(2), the use of task loss with RVQ regularizers, hyperparameter optimization, and the comparison against a DAC-based pipeline. However, the central quantitative claim rests on test-set entropy estimates and on tables with internal inconsistencies; these issues must be resolved before the contribution is fully established.

major comments (4)
  1. [Section 2.2, Eq. (5); Tables 1-2] The reported entropy bitrates violate the stated lower bound in several rows. In Table 1, the K=2, layer-8 row lists BR_raw=500 bps and BR_ent=547.59 bps, even though Eq. (5) states that BR_ent is a lower bound; the same violation occurs for both K=2 rows in Table 2 (947.02 and 924.00 bps versus 800 bps raw) and for the DAC 750 reference row (1528.95 bps versus 750 bps). Moreover, the implied frame rates are inconsistent across rows of the same task: ASR K=1 rows imply R=25, K=2 rows imply R=41.67, and K=12 rows imply R=27.78, while the AC rows imply yet another rate. The tables need to be recomputed with a single explicit frame rate and a consistent definition of entropy before any bitrate comparison can be trusted.
  2. [Section 2.2, Eqs. (3)-(5)] The 'ultra-low bitrate' headline is computed as the empirical entropy of codeword usage over the test set, and no entropy coder is implemented. The paper itself states that the authors 'assume that entropy coding can follow up' and that they report entropy-based bitrates 'as the model's compression performance.' A real Huffman or arithmetic coder would be designed from training-set frequencies, but Eq. (3) uses 'the entire test samples,' making the reported values optimistic lower bounds that do not demonstrate an achievable transmission rate under distribution shift. I request either implementing an actual entropy coder on a held-out set and reporting the realized bitrate, or clearly labeling the numbers as lower-bound estimates and additionally reporting training-set entropy.
  3. [Section 3.3; Table 1] The GMAC comparison with DAC is unfair as presented. The text says that GMACs are computed for 'DAC's encoding, RVQ, and decoding processes,' yet the column is labeled 'on-device' and the 12.30 GMAC figure is compared against the proposed method's on-device encoder plus RVQ. In the conventional pipeline of Fig. 1a, the decoder runs in the cloud, not on the device. Including DAC's decoder in the on-device column inflates the baseline; please report DAC's encoder+RVQ GMACs separately from the decoder.
  4. [Section 3.2; Tables 1-2] No error bars or repeated-run variability are reported. UrbanSound8k uses 10-fold cross-validation, yet Table 2 gives a single accuracy per configuration; the claimed parity with the continuous baseline (80.2 vs. 79.7) may be within run-to-run or fold variance. Similarly, the ASR WER differences of 0.2-0.3% are reported without variance. Please report mean±std over folds and seeds to support the 'minimal loss of downstream performance' claim.
minor comments (6)
  1. [Eq. (4)] Please specify that the logarithm in the entropy calculation is base 2, so that the resulting entropy is expressed in bits.
  2. [Section 3.1, Eq. (6)] The AC objective is said to use the codebook regularizer 'as in Eq. 6,' but Eq. 6 is ASR-specific because it contains CTC and KL terms; please provide the exact loss equation used for audio classification.
  3. [Table 2] The units in Table 2 are labeled MMACs while Table 1 and the text use GMACs; please unify the units and convert consistently.
  4. [Tables 1-2 and Introduction] There are several typos: 'Resuts' in Table 1, 'cudio classification' and 'accyracy' in Table 2, 'continous' in Section 4, and 'recogntion' in the Introduction.
  5. [Section 4] The claim that the scheme is 'universal and transferrable to many speech and audio tasks' is stronger than the evidence, since only two tasks and two architectures are tested and the 'appropriate regularization' is not specified; please soften or provide supporting experiments.
  6. [References] References [8] and [29] are the same DAC paper (NeurIPS and arXiv versions); consider citing them once.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the ultra-low-bitrate claim rests on unverified test-set entropy lower bounds, but the method's evaluation is independent and self-contained.

full rationale

The derivation chain is empirical rather than circular. The quantizer is inserted into a pretrained ASR/AC model and trained jointly with a task loss plus codebook and commitment losses (Eq. 6); the reported WER and accuracy on held-out test sets are independent measurements, not quantities re-derived from the training objective. The reported BR_ent values are empirical entropies of test-set codeword usage (Eqs. 3-5), so the 'less than 200 bps' figures are lower-bound estimates conditioned on the test distribution and on a hypothetical entropy coder, not achieved transmission rates; no entropy coder is implemented. This is a limitation in the strength of the quantitative claim, and the tables also show suspicious BR_ent/BR_raw inconsistencies (e.g., ASR 2-codebook 8th-layer row: BR_ent 547.59 vs BR_raw 500) that should be corrected before the numbers are used, but no fitted parameter is renamed as a prediction. The only self-citations (refs [15] and [18]) appear in related-work examples and are not load-bearing: the method's validity rests on the implemented experiments, not on those citations. Accordingly, no circular step is present; the score is 1 only to reflect minor, non-load-bearing self-citations.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method uses only existing components (RVQ, codebooks, task losses). No new physical entities or theoretical constructs are introduced. The free parameters are standard hyperparameters; their exact values are not fully reported.

free parameters (6)
  • Number of RVQ codebooks K = 2-12 (ASR), 1-2 (AC)
    Chooses the number of quantization stages; directly sets bitrate and model complexity. Selected via TPE.
  • Codebook size V = 1024 or 8192 (ASR); 32-1024 (AC)
    Number of codewords per codebook; larger V increases bitrate and capacity.
  • Code dimension D = 8, 64, or 512
    Dimension of each codeword; tuned for each task.
  • RVQ regularization weight beta = tuned (AC), unspecified (ASR)
    Balances codebook/commitment losses against task loss; only reported for AC tuning.
  • Task loss weight lambda = not reported
    Eq. 6 defines lambda and beta, but only beta is said to be tuned.
  • Quantized layer M = 1, 2, 4, 6, 8, 12
    Depth of the split point; affects bitrate, accuracy, and on-device compute.
assumptions (4)
  • domain assumption The downstream model can be split into device-side and cloud-side parts without loss of generality.
    Section 2.1 states this as the main assumption.
  • domain assumption RVQ training with codebook and commitment losses leads to a usable discrete representation for the task model.
    Used in all experiments; no theoretical guarantee.
  • domain assumption The empirical distribution of codewords on the test set is representative for entropy coding.
    Eqs. 3-5 compute entropy from test samples.
  • domain assumption Standard deep learning training with TPE hyperparameter search produces the reported results.
    Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Specific Audio Coding for Machines: Machine-Learned Latent Features Are Codes for That Machine." pith.science (2026). https://pith.science/paper/LYRY7COK

@misc{pith2026250712701,
  author       = {Pith},
  title        = {Pith review of: Task-Specific Audio Coding for Machines: Machine-Learned Latent Features Are Codes for That Machine},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LYRY7COK}},
  note         = {Machine review of arXiv:2507.12701}
}
read the original abstract

Neural audio codecs, leveraging quantization algorithms, have significantly impacted various speech/audio tasks. While high-fidelity reconstruction is paramount for human perception, audio coding for machines (ACoM) prioritizes efficient compression and downstream task performance, disregarding perceptual nuances. This work introduces an efficient ACoM method that can compress and quantize any chosen intermediate feature representation of an already trained speech/audio downstream model. Our approach employs task-specific loss guidance alongside residual vector quantization (RVQ) losses, providing ultra-low bitrates (i.e., less than 200 bps) with a minimal loss of the downstream model performance. The resulting tokenizer is adaptable to various bitrates and model sizes for flexible deployment. Evaluated on automatic speech recognition and audio classification, our method demonstrates its efficacy and potential for broader task and architectural applicability through appropriate regularization.

Figures

Figures reproduced from arXiv: 2507.12701 by the authors.

Figure 1
Figure 1. (a) The conventional ML pipeline involving a codec. The ML [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 27 canonical work pages

  1. [1]

    Coding of moving pictures and associated audio for digital storage media at up to about 1.5 Mbit/s,

    ISO/IEC 11172-3:1993, “Coding of moving pictures and associated audio for digital storage media at up to about 1.5 Mbit/s,” 1993

    work page 1993

  2. [2]

    Perceptual coding of digital audio,

    T. Painter and A. Spanias, “Perceptual coding of digital audio,” Proceed- ings of the IEEE , vol. 88, no. 4, pp. 451–515, 2000

    work page 2000

  3. [3]

    Information technology — Generic coding of moving pictures and associated audio information — Part 7: Advanced Audio Coding (AAC),

    ISO (2006) ISO/IEC 13818-7:2006, “Information technology — Generic coding of moving pictures and associated audio information — Part 7: Advanced Audio Coding (AAC),” 2006

    work page 2006

  4. [4]

    ISO/IEC - Information technology – MPEG audio technologies – Part 2: Spatial Audio Object Coding (SAOC) , ISO/IEC IS 23 003-2, 2010

    work page 2010

  5. [5]

    Transport of unified speech and audio coding (USAC),

    ISO/IEC 14496-3:2009/PDAM 3, “Transport of unified speech and audio coding (USAC),” 2011

    work page 2009

  6. [6]

    Wideband coding of speech at around 16 kbit/s using Adaptive Multi-Rate Wideband (AMR-WB),

    ITU-T Recommendation G.722.2, “Wideband coding of speech at around 16 kbit/s using Adaptive Multi-Rate Wideband (AMR-WB),” 2003

    work page 2003

  7. [7]

    Soundstream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “Soundstream: An end-to-end neural audio codec,” IEEE/ACM Trans. Audio, Speech and Lang. Proc. , vol. 30, p. 495–507, jan 2022

    work page 2022

  8. [8]

    High- fidelity audio compression with improved RVQGAN,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High- fidelity audio compression with improved RVQGAN,” in Advances in Neural Information Processing Systems (NeurIPS) , 2023

    work page 2023

Show all 40 references
  1. [9]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,” Transactions on Machine Learning Research , 2023

    2023

  2. [11]

    On generative spoken language modeling from raw audio,

    K. Lakhotia, E. Kharitonov, W.-N. Hsu, Y . Adi, A. Polyak, B. Bolte, T.-A. Nguyen, J. Copet, A. Baevski, A.and Mohamed et al. , “On generative spoken language modeling from raw audio,” Transactions of the Association for Computational Linguistics , vol. 9, pp. 1336–1354, 2021

    2021

  3. [12]

    Recent advances in speech language models: A survey,

    W. Cui, D. Yu, X. Jiao, Z. Meng, G. Zhang, Q. Wang, Y . Guo, and I. King, “Recent advances in speech language models: A survey,” arXiv preprint arXiv:2410.03751, 2024

    2024 arXiv

  4. [13]

    Neural codec language models are zero-shot text to speech synthesizers,

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al., “Neural codec language models are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111 , 2023

    2023 arXiv

  5. [14]

    Beats: Audio pre-training with acoustic tokenizers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, and F. Wei, “Beats: Audio pre-training with acoustic tokenizers,” arXiv preprint arXiv:2212.09058, 2022

    2022 arXiv

  6. [15]

    Genhancer: High-fidelity speech enhancement via generative modeling on discrete codec tokens,

    H. Yang, J. Su, M. Kim, and Z. Jin, “Genhancer: High-fidelity speech enhancement via generative modeling on discrete codec tokens,” in Proc. Interspeech, 2024

    2024

  7. [16]

    DASB - discrete audio and speech benchmark,

    P. Mousavi, L. D. Libera, J. Duret, A.Ploujnikov, C. Subakan, and M. Ravanelli, “DASB - discrete audio and speech benchmark,” 2024

    2024

  8. [17]

    Codec-SUPERB: An in-depth analysis of sound codec models,

    H. Wu, H.-L. Chung, Y .-C. Lin, Y .-K. Wu, X. Chen, Y .-C. Pai, H.-H. Wang, K.-W. Chang, A. Liu, and H. Lee, “Codec-SUPERB: An in-depth analysis of sound codec models,” in Findings of the Association for Computational Linguistics ACL 2024 , Aug. 2024, pp. 10 330–10 348

    2024

  9. [18]

    Psychoacoustic calibration of loss functions for efficient end-to-end neural audio coding,

    K. Zhen, M. S. Lee, J. Sung, S. Beack, and M. Kim, “Psychoacoustic calibration of loss functions for efficient end-to-end neural audio coding,” IEEE Signal Processing Letters , vol. 27, pp. 2159–2163, 2020

    2020

  10. [19]

    Use Cases and Requirements on Audio Coding for Machines,

    MPEG, “Use Cases and Requirements on Audio Coding for Machines,” International Organisation for Standardisation, Tech. Rep. N0046. [Online]. Available: https://www.mpeg.org/standards/Explorations/46/

  11. [20]

    Wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “Wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in neural information processing systems , vol. 33, pp. 12 449– 12 460, 2020

    2020

  12. [21]

    WavLM: Large-Scale Self-Supervised Pre- Training for Full Stack Speech Processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao et al., “WavLM: Large-Scale Self-Supervised Pre- Training for Full Stack Speech Processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  13. [22]

    HuBERT: Self-Supervised Speech Representation Learn- ing by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-Supervised Speech Representation Learn- ing by Masked Prediction of Hidden Units,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 29, pp. 3451–3460, 2021

    2021

  14. [23]

    On The Landscape of Spoken Language Models: A Comprehensive Survey,

    S. Arora, K.-W. Chang, C.-M. Chien, Y . Peng, Y . Wu, H.and Adi, E. Dupoux, H.-Y . Lee, K. Livescu, and S. Watanabe, “On The Landscape of Spoken Language Models: A Comprehensive Survey,” arXiv preprint arXiv:2504.08528, 2025

    2025 arXiv

  15. [24]

    Towards audio language modeling–an overview,

    H. Wu, X. Chen, Y .-I. Lin, K. Chang, H.-L. Chung, A. H. Liu, and H. Lee, “Towards audio language modeling–an overview,” arXiv preprint arXiv:2402.13236, 2024

    2024 arXiv

  16. [25]

    Espnet-codec: Comprehensive training and evaluation of neural codecs for audio, music, and speech,

    J. Shi, J. Tian, Y . Wu, J.-W. Jung, J. Yip, Y . Masuyama, W. Chen, Y . Wu, Y . Tang, M. Baali, D. Alharthi, D. Zhang, R. Deng, T. Srivastava, H. Wu, A. Liu, B. Raj, Q. Jin, R. Song, and S. Watanabe, “Espnet-codec: Comprehensive training and evaluation of neural codecs for aud...

    2024

  17. [26]

    Recent advances in discrete speech tokens: A review,

    Y . Guo, Z. Li, H. Wang, B. Li, C. Shao, H. Zhang, C. Du, X. Chen, S. Liu, and K. Yu, “Recent advances in discrete speech tokens: A review,” arXiv preprint arXiv:2502.06490 , 2025

    2025

  18. [27]

    CosyV oice: A scalable multilingual zero-shot text-to- speech synthesizer based on supervised semantic tokens,

    Z. Du, Q. Chen, S. Zhang, K. Hu, H. Lu, Y . Yang, H. Hu, S. Zheng, Y . Gu, Z. Maet al., “CosyV oice: A scalable multilingual zero-shot text-to- speech synthesizer based on supervised semantic tokens,” arXiv preprint arXiv:2407.05407, 2024

    2024 arXiv

  19. [28]

    Soft-to-hard vector quantization for end- to-end learning compressible representations,

    E. Agustsson, F. Mentzer, M. Tschannen, L. Cavigelli, R. Timofte, L. Benini, and L. V . Gool, “Soft-to-hard vector quantization for end- to-end learning compressible representations,” in Advances in Neural Information Processing Systems (NIPS) , 2017, pp. 1141–1151

    2017

  20. [29]

    High- fidelity audio compression with improved RVQGAN,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High- fidelity audio compression with improved RVQGAN,” arXiv preprint arXiv:2306.06546, 2023

    2023 arXiv

  21. [30]

    Open-Source Conversational AI with SpeechBrain 1.0,

    M. Ravanelli, T. Parcollet, A. Moumen, S. de Langen, C.Subakan, P. Plantinga, Y . Wang, P. Mousavi, L. D. Libera, A. Ploujnikov, F. Paissan, D. Borra, S. Zaiem, Z. Zhao, S. Zhang, G. Karakasidis, S.-L. Yeh, P. Champion, A. Rouhe, R. Braun, F. Mai, J. Zuluaga-Gomez, S. M. Mousa...

    2024

  22. [31]

    SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing,

    T. Kudo and J. Richardson, “SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations . Brussels, Belgium: Ass...

    2018

  23. [32]

    Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning , 2006, pp. 369–376

    2006

  24. [33]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” CoRR, vol. abs/1412.6980, 2015

    2015 arXiv

  25. [34]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” in Advances in Neural Information Processing Systems (NIPS), 2017, pp. 6306–6315

    2017

  26. [35]

    ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” arXiv preprint arXiv:2005.07143 , 2020

    2005 arXiv

  27. [36]

    Margin matters: Towards more discriminative deep neural network embeddings for speaker recognition,

    X. Xiang, S. Wang, H. Huang, Y . Qian, and K. Yu, “Margin matters: Towards more discriminative deep neural network embeddings for speaker recognition,” in 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) . IEEE, 2019, pp...

    2019

  28. [37]

    Tree-structured parzen estimator: Understanding its algorithm components and their roles for better empirical performance,

    S. Watanabe, “Tree-structured parzen estimator: Understanding its algorithm components and their roles for better empirical performance,” arXiv preprint arXiv:2304.11127 , 2023

    2023 arXiv

  29. [38]

    Epistimio/orion: Asynchronous Distributed Hyperparameter Optimization,

    X. Bouthillier, C.Tsirigotis, F. Corneau-Tremblay, T. Schweizer, L.Dong, P. Delaunay, F.Normandin, M. Bronzi, D. Suhubdy, R. Askari, M. Noukhovitch, C.Xue, S. Ortiz-Gagn ´e, O. Breuleux, A. Bergeron, O. Bilaniuk, S. Bocco, H.Bertrand, G. Alain, D. Serdyuk, P. Henderson, P. Lam...

    2022 doi

  30. [39]

    Librispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An ASR corpus based on public domain audio books,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  31. [40]

    A Dataset and Taxonomy for Urban Sound Research,

    J. Salamon, C. Jacoby, and J. Bello, “A Dataset and Taxonomy for Urban Sound Research,” in Proceedings of the 22nd ACM International Conference on Multimedia , 2014, p. 1041–1044. [Online]. Available: https://doi.org/10.1145/2647868.2655045

    2014

  32. [41]

    On the information rate of speech communication,

    S. V . Kuyk, W. B. Kleijn, and R. C. Hendriks, “On the information rate of speech communication,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2017, pp. 5625–5629

    2017

Pith tools

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