Pith. sign in

REVIEW 4 major objections 7 minor 34 references

GhostRNN: Reducing State Redundancy in RNN with Cheap Operations

T0 review · 4 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper claims that an RNN's hidden state is partly redundant, and that the redundant part can be regenerated from a smaller intrinsic set by cheap operations, giving ~40% compression with no accuracy loss.

desk verdict A clean, plausible adaptation of GhostNet to RNN hidden states, but the empirical case rests on point estimates with no variance; needs error bars before I'd trust the 'similar performance' claim. read the letter →

arxiv 2411.14489 v1 pith:LMR3LQ4H submitted 2024-11-20 cs.CL cs.AIcs.SDeess.AS

Hang Zhou , Xiaoxu Zheng , Yunhe Wang , Michael Bi Mi , Deyi Xiong , Kai Han This is my paper · ORCID
classification cs.CLcs.AIcs.SDeess.AS
keywords GhostRNNcompressionhiddenstateredundancycheapoperationsstateskeywordspottingspeechenhancementGRU
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 tries to show that the hidden states of recurrent neural networks contain redundant dimensions that can be exploited for model compression on speech tasks. It proposes GhostRNN, in which a small number of intrinsic states are computed by the usual RNN equations, and the remaining ghost states are produced by cheap learned linear transformations plus an activation. On keyword spotting and speech enhancement, the resulting models keep accuracy and signal quality essentially unchanged while using about 40% fewer parameters and fewer FLOPs. If the redundancy is as generic as the paper suggests, this gives a simple way to fit competitive RNNs into low-resource devices.

What carries the argument

The load-bearing object is the ghost-state generation equation $g_t = \phi(h_t)$, where $\phi$ is a cheap operation composed of a learned linear layer and an activation function. This is what converts a smaller intrinsic hidden vector into the concatenated full state $[h_t; g_t]$ carried to the next time step. The machinery also includes the parameter-count identity $\text{Param}_\phi = (D/r)(D - D/r)$, which keeps the extra cost quadratic in the intrinsic dimension but divided by $r^2$ relative to a full GRU matrix. This identity is what guarantees the roughly $r$-fold reduction in parameters and MACs.

What would settle it

Train the same keyword spotting and speech enhancement models on tasks whose hidden states are known to be irreducibly high-dimensional, or measure the PCA contribution rate across a variety of RNN architectures: if a well-trained RNN shows near-zero cosine similarity among most dimensions and needs close to 100% of singular values for 99% energy, then generating ghost states from a half-sized intrinsic set should noticeably degrade accuracy or SDR, and the roughly 40% compression would fail to match the baseline.

Watch

Extended reading notes

Core claim

The central discovery is empirical: in a trained GRU hidden-state trajectory, roughly half of the singular values account for 99% of the PCA energy, and several hidden-state components have cosine similarity near 1 while others are near 0. The paper reads this as partial redundancy in the state space. GhostRNN formalizes the exploitation by computing an intrinsic state $h_t$ of dimension $D/r$ with the standard GRU equations, then generating a ghost state $g_t = \phi(h_t)$ through a learned linear map followed by an activation, and feeding the concatenation $[h_{t-1}, g_{t-1}]$ into the next time step. This replaces a fraction of the recurrent matrix multiplications with much cheaper operations. Across GRU-style keyword spotting and speech enhancement models, a model with roughly 40% fewer parameters matches or slightly beats its larger baseline, including about 0.1% higher accuracy on Google Speech Commands and about 0.1 dB higher SDR and Si-SDR than matched-parameter baselines.

Load-bearing premise

The weakest premise is that the redundant dimensions of a trained RNN hidden state can always be regenerated from a smaller intrinsic set by a simple learned linear-plus-activation map without losing task-relevant information, a claim the paper supports with PCA and cosine-similarity evidence from a single GRU model rather than across architectures.

Editorial extensions

If this is right

  • At roughly 40% parameter compression, GhostRNN matches or slightly exceeds the 500K GRU on Google Speech Commands (94.79% vs 94.68% accuracy).
  • Because ghost states are produced once per time step and reused in all subsequent recurrences, per-step FLOPs drop approximately in proportion to the parameter reduction.
  • The construction is not tied to GRU: the same intrinsic-plus-cheap-ghost pattern is claimed to apply to LSTM and other RNN variants.
  • When the cheap operation is parameter-free, the compression ratio can reach the full factor $r$; with a learned linear layer the saving is slightly less but still large.

Reading between the lines

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

  • Editorial inference: the PCA and cosine-similarity redundancy signature is measured on one DCCRN GRU with 128 hidden units, so the 2:1 intrinsic-to-ghost split may not be optimal for other sizes; the ratio $r$ could be treated as a tunable hyperparameter.
  • Editorial inference: if the redundancy is a property of gradient-trained recurrent dynamics rather than of speech specifically, the same compression should transfer to RNNs for automatic speech recognition, acoustic echo cancellation, and other sequence tasks, which would be a direct test.
  • Editorial inference: a sharper comparison would pit GhostRNN against simply halving the hidden size with a wider input projection, isolating whether the ghost-state structure itself, rather than the parameter reduction, drives the preserved performance.
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 / 7 minor

Summary. GhostRNN proposes to compress RNNs by exploiting redundancy in hidden states. The method keeps a small set of 'intrinsic' states computed by a vanilla GRU/LSTM recurrence, then generates 'ghost' states via a cheap operation (a linear layer plus activation, Eq. 9), concatenating intrinsic and ghost states as the input to the next time step. The parameter-count analysis (Eqs. 10-12) shows a reduction by roughly the ratio r of the total state size to the intrinsic state size. Experiments on keyword spotting (Google Speech Commands) and speech enhancement (LibriMix, with DCCRN and GRU-TasNet baselines) report that GhostRNN matches or slightly exceeds baseline accuracy/SDR at roughly 40% fewer parameters and MACs. The paper concludes that GhostRNN reduces memory and computation while keeping performance similar.

Significance. If the empirical claims hold, GhostRNN is a simple and broadly applicable plug-in compression module for RNN-based speech models, and the extension of the cheap-operation principle from CNNs (GhostNet [16]) to recurrent hidden states is a reasonable and non-obvious step. The parameter-count derivation in Sec. 2.3 is correct and checkable by hand, and the compression is by construction; the matched-parameter comparisons in Tables 1-2 are an appropriate experimental design; and the redundancy analysis in Sec. 2.2 provides a concrete, falsifiable motivation. The contribution is currently limited, however, by the absence of any variance estimate for the headline margins (~0.1% accuracy, ~0.1 dB SDR), by the lack of measured memory/latency to back the abstract's 'memory usage' phrasing, and by the single-model basis of the redundancy observation. With the requested statistical and ablation support, the paper would be a useful contribution to efficient speech modeling.

major comments (4)
  1. [Tables 1-2; Sec. 3.2, 3.4-3.5] The paper's central claim is that GhostRNN 'keeps performance similar' while compressing parameters by ~40%, but the supporting evidence is a set of point estimates with no variance: the KWS section states that each model was trained three times and averages are reported (Sec. 3.2), yet the individual runs and their spread are never shown, and the SE results in Table 2 appear to be single runs. All reported margins (accuracy +0.11 percentage points over the 498K GRU; SDR differences of +0.12, +0.13, and -0.06 dB across Table 2) are of a size that typical seed-to-seed variation in speech tasks can reverse. Because the entire contribution is an empirical efficiency claim, please report per-seed results, standard deviations or confidence intervals, and ideally a paired significance test for the KWS and SE comparisons; without this, the headline conclusion is statistically unsupported.
  2. [Sec. 3; Eqs. (9)-(12)] The proposed architecture is under-specified: the value of the ratio r (or, equivalently, the intrinsic-state dimension) and the form of the cheap operation phi used in the experiments are never stated. The reader cannot determine how GhostRNN-292K, DCCRN Ghost128, or GhostRNN512-TasNet were configured, which prevents reproduction and independent verification. Please report r, the intrinsic and ghost dimensions, and the exact phi used for each experiment, and add a sensitivity analysis over r (e.g., r = 2, 4, 8) so that the trade-off between compression and performance is quantified.
  3. [Abstract; Sec. 3.4-3.5] The abstract states that GhostRNN 'significantly reduces the memory usage (~40%) and computation cost,' but Tables 1-2 report only parameter counts and MACs; no actual memory footprint, peak activation memory, latency, or energy is measured. Since the stated motivation is deployment on low-resource devices, parameter-count-based inference is a legitimate but incomplete proxy: the RNN hidden state itself is tiny relative to model weights, and MACs do not capture operational memory traffic. Please either measure actual memory and runtime on the target hardware, or rephrase the claims to refer explicitly to parameter count and MACs.
  4. [Sec. 2.2; Eq. (9)] The key architectural assumption — that the redundant dimensions of a hidden state can be regenerated from the intrinsic states by a learned linear transform plus activation without losing task-relevant information — is motivated by a PCA and cosine-similarity analysis of a single trained model (one DCCRN GRU with 128 units, Sec. 2.2). The paper provides no evidence that this regeneration assumption holds across architectures, layer depths, or scales, which creates a circularity risk because the observation and the validation use the same task family. Please add either (i) reconstruction-error or downstream-task measurements of the ghost operation on different RNN variants (e.g., LSTM, different GRU widths), or (ii) ablations comparing the proposed phi against alternative cheap operations and against a matched-parameter plain GRU, to establish generality.
minor comments (7)
  1. [Eq. (7)] The candidate equation contains a parenthesis imbalance ("rt * (Whch(t-1) + bhc) + Wgcgt-1 + bgc)"); please rewrite it so the scope of the tanh is unambiguous.
  2. [Sec. 2.2] The initialization of the ghost state g0 at the first time step is never defined; please specify g0 (e.g., zeros).
  3. [Sec. 2.3] The text states that the total parameter count 'will be compressed by the factor r'; this is only an upper bound, since Eqs. (11)-(12) give a compression ratio strictly less than r for finite feature dimension. Please rephrase to 'approximately r' or 'up to r' for consistency with the formulas.
  4. [Various] Typographical errors: 'vanila' (Sec. 2.3), 'vanllia' (Conclusions), 'roubustness' (Sec. 3.1), and 'Simliar' (Introduction); also, the notation 'statei' in Sec. 2.2 is inconsistently formatted.
  5. [Table 1] The claim of a '0.1% accuracy improvement' is better stated as +0.11 percentage points (about 0.12% relative); also clarify whether the reported parameter counts include the classification head and feature extractor, since the baseline GRU sizes (498K/295K) do not obviously match a pure 10-dimensional-input GRU.
  6. [Sec. 3.5] The text refers to the 'librimix1 dataset' while Sec. 3.1 describes LibriMix; please use the dataset name and the train-360 subset designation consistently.
  7. [Sec. 2, Conclusions] The paper says GhostRNN 'can be applicable to other RNNs, e.g., LSTM,' but no LSTM experiment appears; either add an LSTM result or soften the claim, since the conclusions also list LSTM extension as future work.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GhostRNN's architecture is motivated by an empirical redundancy observation and validated on independent KWS/SE experiments; self-citations are not load-bearing.

full rationale

The paper's central claim is empirical: GhostRNN compresses RNN hidden states by generating ghost states from intrinsic states via cheap linear transforms, and this is evaluated against GRU and other baselines on keyword spotting and speech enhancement. The motivating observation—that trained RNN hidden states show redundancy by PCA and cosine similarity—is measured on a real trained DCCRN GRU model (Sec. 2.2) and is not defined in terms of the GhostRNN outcome. The proposed equations (5)–(9) define the architecture, and the parameter-count analysis (Eqs. 10–12) is a direct arithmetic derivation, not a prediction that reduces to a fitted input. The citation to the authors' prior GhostNet paper [16] supplies the general idea of cheap operations, but the paper does not rely on [16] as a uniqueness theorem or as proof of GhostRNN's effectiveness; the effectiveness is established by new experiments on public datasets. The lack of variance reporting and the small observed margins are legitimate statistical-correctness concerns, but they are not circularity: the performance numbers are not forced by construction, by self-definition, or by self-citation. Hence no specific circular step can be quoted, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central claim rests on the assumption that ghost states can cheaply replace redundant hidden dimensions. The compression ratio r and activation phi are free choices, and the redundancy observation comes from a single model. These are design choices rather than derived laws, so the ledger is modest but not empty.

free parameters (3)
  • ratio r (intrinsic-to-total state ratio) = inferred to be 2 from parameter counts
    The hidden dimension is divided by r and the remaining dimensions are generated by cheap ops. The experiments do not state r explicitly, but the reported parameter counts imply r=2. Performance depends on this choice and no sensitivity analysis is provided.
  • ghost activation function phi = unspecified
    Eq. (9) defines ghost states as phi(h_t), but the exact nonlinearity is never given. Results may depend on it.
  • layout of cheap-operation layers = unspecified, one linear layer assumed in Eq. (12)
    The parameter-count equation assumes a single linear layer of size (S/r) by (S - S/r); the actual implementation is not detailed.
assumptions (3)
  • domain assumption Hidden-state redundancy in RNNs is general and can be exploited by cheap linear transforms.
    The paper observes PCA and cosine similarity on a single trained DCCRN GRU model (Sec. 2.2, Figs. 1-2) and generalizes to all RNNs and tasks without further evidence.
  • ad hoc to paper A small intrinsic state plus cheap-operation-generated ghost states can represent the information of the full hidden state.
    This is the core architectural assumption of Eqs. (5)-(9); no formal guarantee or independent test is given.
  • standard math Standard GRU update equations (Eqs. 1-4) are correct background.
    The GRU equations are taken from prior literature [1,2] and are used as the starting point for GhostRNN.
invented entities (1)
  • Ghost state g_t
    purpose: To replace redundant dimensions of the hidden state with cheaply generated approximations, reducing parameter and FLOP counts.
    The ghost state is an architectural construct evaluated only on the paper's own benchmarks; it makes no falsifiable prediction outside those tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GhostRNN: Reducing State Redundancy in RNN with Cheap Operations." pith.science (2026). https://pith.science/paper/LMR3LQ4H

@misc{pith2026241114489,
  author       = {Pith},
  title        = {Pith review of: GhostRNN: Reducing State Redundancy in RNN with Cheap Operations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LMR3LQ4H}},
  note         = {Machine review of arXiv:2411.14489}
}
read the original abstract

Recurrent neural network (RNNs) that are capable of modeling long-distance dependencies are widely used in various speech tasks, eg., keyword spotting (KWS) and speech enhancement (SE). Due to the limitation of power and memory in low-resource devices, efficient RNN models are urgently required for real-world applications. In this paper, we propose an efficient RNN architecture, GhostRNN, which reduces hidden state redundancy with cheap operations. In particular, we observe that partial dimensions of hidden states are similar to the others in trained RNN models, suggesting that redundancy exists in specific RNNs. To reduce the redundancy and hence computational cost, we propose to first generate a few intrinsic states, and then apply cheap operations to produce ghost states based on the intrinsic states. Experiments on KWS and SE tasks demonstrate that the proposed GhostRNN significantly reduces the memory usage (~40%) and computation cost while keeping performance similar.

Figures

Figures reproduced from arXiv: 2411.14489 by the authors.

Figure 1
Figure 1. The cosine similarity matrix of RNN hidden states (first 12 dimensions) in a well-trained DCCRN GRU with 128 hidden units. the number of gate matrices may undermine the exploration of contextual information. In this work, we have empirically observed redundancy in hidden states of RNN models in addition to that in gate matrices studied in aforementioned previous works. We hence propose to fully explore the redundanc… view at source ↗
Figure 2
Figure 2. The value of hidden states at different indexes. be constructed. Furthermore, as shown in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Schematic diagram of the RNN layer and proposed Ghost RNN. ghost states. The concatenated states are used as the input for the next time step. Specifically, at time step t the GhostRNN receives both the previous hidden state ht−1 and correspond￾ing ghost state gt−1 synchronously with the current input fea￾ture xt. This process is repeated to complete the calculation of the GhostRNN. As mentioned above, the GhostRNN … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 19 canonical work pages

  1. [16]

    Acoustic echo cancellation by combining adaptive digital filter and recurrent neural network,

    L. Ma, H. Huang, P. Zhao, and T. Su, “Acoustic echo cancellation by combining adaptive digital filter and recurrent neural network,” arXiv preprint arXiv:2005.09237, 2020

    arXiv 2005

  2. [1]

    Introduction Recent years have witnessed that substantial improvements have been made in a wide range of speech tasks with the rapid development of neural networks. Among these neural networks, RNNs, e.g., LSTMs [1] or GRUs [2], are widely employed in various speech-related tasks in low-resource devices (e.g., mo- bile phones), such as KWS [3, 4], SE [5, ...

    work page

  3. [2]

    Without loss of generality, we use GRU to illustrate the definition of GhostRNN

    Proposed Method In this section, we elaborate the proposed GhostRNN with de- tails in model compression. Without loss of generality, we use GRU to illustrate the definition of GhostRNN. Our method can be applicable to other RNNs, e.g., LSTM. 2.1. RNN RNNs are a class of model structures that utilize hidden states to store and leverage contextual informati...

    work page

  4. [3]

    Experiments Experiments on two tasks were conducted to evaluate the effec- tiveness of our method: KWS and SE. Table 1: Overall performance of KWS System # Params # MACs Accuracy (%) GRU 498K 24.1M 94.68 GRU 295K 14.2M 94.49 Li-GRU 334K 16.1M 93.49 SITGRU 334K 16.1M 94.26 GhostRNN 292K 14.0M 94.79 3.1. Datasets The Google Speech Commands dataset v0.02 [18...

    work page

  5. [4]

    Conclusions In this paper, we have presented GhostRNN for RNN model compression based on the observation of the redundancy in hid- den states. In our GhostRNN, given the intrinsic hidden states, the extreme low-cost transformation layers are applied to gen- erate the ghost states which significantly reduces the number of parameters and the computation cos...

    work page

  6. [5]

    • GRU-TasNet

    model and replaced the LSTM with GRU, in which two sizes of the hidden unit 128 and 80 were chosen to construct the different baseline models with different size. • GRU-TasNet. This model is optimized by the Time-domain Audio Separation Network, which consists of three parts: a 1-D convolutional encoder, a 1-D deconvolutional decoder, and a Deep LSTM sepa...

    work page

  7. [6]

    Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech sepa- ration,

    Y . Luo, Z. Chen, and T. Yoshioka, “Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech sepa- ration,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 46–50

    2020

  8. [7]

    2022D01D43)

    Acknowledgements Deyi Xiong was partially supported by the Natural Sci- ence Foundation of Xinjiang Uygur Autonomous Region (No. 2022D01D43). We gratefully acknowledge the support of MindSpore [27], CANN(Compute Architecture for Neural Net- works) and Ascend AI Processor used for this research. We would like to thank the anonymous reviewers for their insi...

    work page

Show all 34 references
  1. [8]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  2. [9]

    Empirical evalu- ation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evalu- ation of gated recurrent neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, 2014

    2014 arXiv

  3. [10]

    Hello edge: Keyword spotting on microcontrollers,

    Y . Zhang, N. Suda, L. Lai, and V . Chandra, “Hello edge: Keyword spotting on microcontrollers,” arXiv preprint arXiv:1711.07128 , 2017

    2017 arXiv

  4. [11]

    Streaming keyword spotting on mobile devices,

    O. Rybakov, N. Kononenko, N. Subrahmanya, M. Visontai, and S. Laurenzo, “Streaming keyword spotting on mobile devices,” arXiv preprint arXiv:2005.06720, 2020

    2005 arXiv

  5. [12]

    Dccrn: Deep complex convolution re- current network for phase-aware speech enhancement,

    Y . Hu, Y . Liu, S. Lv, M. Xing, S. Zhang, Y . Fu, J. Wu, B. Zhang, and L. Xie, “Dccrn: Deep complex convolution re- current network for phase-aware speech enhancement,” arXiv preprint arXiv:2008.00264, 2020

    2008 arXiv

  6. [13]

    Simliar to Li- GRU, Fanta et al

    propose the Embedded Gated Recurrent Unit, which has only one gate with the Single Gate Mechanism. Simliar to Li- GRU, Fanta et al. [14] discard the reset gate of GRU and re- place the activation function Tanh with Sigmoid in their pro- posed SITGRU. Zhang et al. [15] compress...

    2024 arXiv

  7. [14]

    Exploring architectures, data and units for streaming end-to-end speech recognition with rnn-transducer,

    K. Rao, H. Sak, and R. Prabhavalkar, “Exploring architectures, data and units for streaming end-to-end speech recognition with rnn-transducer,” in2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2017, pp. 193–199

    2017

  8. [15]

    Nonlinear residual echo sup- pression using a recurrent neural network

    L. Pfeifenberger and F. Pernkopf, “Nonlinear residual echo sup- pression using a recurrent neural network.” in Interspeech, 2020, pp. 3950–3954

    2020

  9. [17]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  10. [18]

    Gate-variants of gated recurrent unit (gru) neural networks,

    R. Dey and F. M. Salem, “Gate-variants of gated recurrent unit (gru) neural networks,” in 2017 IEEE 60th international midwest symposium on circuits and systems (MWSCAS). IEEE, 2017, pp. 1597–1600

    2017

  11. [19]

    Light gated recurrent units for speech recognition,

    M. Ravanelli, P. Brakel, M. Omologo, and Y . Bengio, “Light gated recurrent units for speech recognition,” IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 2, no. 2, pp. 92–102, 2018

    2018

  12. [20]

    An optimized recurrent unit for ultra- low-power keyword spotting,

    J. Amoh and K. M. Odame, “An optimized recurrent unit for ultra- low-power keyword spotting,” Proceedings of the ACM on Inter- active, Mobile, Wearable and Ubiquitous Technologies , vol. 3, no. 2, pp. 1–17, 2019

    2019

  13. [21]

    Sitgru: single-tunnelled gated re- current unit for abnormality detection,

    H. Fanta, Z. Shao, and L. Ma, “Sitgru: single-tunnelled gated re- current unit for abnormality detection,”Information Sciences, vol. 524, pp. 15–32, 2020

    2020

  14. [22]

    Simplifying neural machine translation with addition-subtraction twin-gated recurrent networks,

    B. Zhang, D. Xiong, J. Su, Q. Lin, and H. Zhang, “Simplifying neural machine translation with addition-subtraction twin-gated recurrent networks,” arXiv preprint arXiv:1810.12546, 2018

    2018 arXiv

  15. [23]

    Ghost- net: More features from cheap operations,

    K. Han, Y . Wang, Q. Tian, J. Guo, C. Xu, and C. Xu, “Ghost- net: More features from cheap operations,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, 2020, pp. 1580–1589

    2020

  16. [24]

    Learning long-term de- pendencies with gradient descent is difficult,

    Y . Bengio, P. Simard, and P. Frasconi, “Learning long-term de- pendencies with gradient descent is difficult,” IEEE transactions on neural networks, vol. 5, no. 2, pp. 157–166, 1994

    1994

  17. [25]

    Speech commands: A dataset for limited-vocabulary speech recognition,

    P. Warden, “Speech commands: A dataset for limited-vocabulary speech recognition,” arXiv preprint arXiv:1804.03209, 2018

    2018 arXiv

  18. [26]

    End-to-end low resource keyword spotting through character recognition and beam-search re-scoring,

    E. T. Mekonnen, A. Brutti, and D. Falavigna, “End-to-end low resource keyword spotting through character recognition and beam-search re-scoring,” in ICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8182–8186

    2022

  19. [27]

    Librimix: An open-source dataset for generalizable speech separation,

    J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vin- cent, “Librimix: An open-source dataset for generalizable speech separation,” arXiv preprint arXiv:2005.11262, 2020

    2005 arXiv

  20. [28]

    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 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  21. [29]

    Wham!: Extend- ing speech separation to noisy environments,

    G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, D. Crow, E. Manilow, and J. L. Roux, “Wham!: Extend- ing speech separation to noisy environments,” arXiv preprint arXiv:1907.01160, 2019

    1907 arXiv

  22. [30]

    As- teroid: the PyTorch-based audio source separation toolkit for re- searchers,

    M. Pariente, S. Cornell, J. Cosentino, S. Sivasankaran, E. Tzinis, J. Heitkaemper, M. Olvera, F.-R. St ¨oter, M. Hu, J. M. Mart ´ın- Do˜nas, D. Ditter, A. Frank, A. Deleforge, and E. Vincent, “As- teroid: the PyTorch-based audio source separation toolkit for re- searchers,” in...

    2020

  23. [31]

    Real-time single-channel dereverbera- tion and separation with time-domain audio separation network

    Y . Luo and N. Mesgarani, “Real-time single-channel dereverbera- tion and separation with time-domain audio separation network.” in Interspeech, 2018, pp. 342–346

    2018

  24. [32]

    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 Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 626–630

    2019

  25. [33]

    An al- gorithm for intelligibility prediction of time–frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An al- gorithm for intelligibility prediction of time–frequency weighted noisy speech,” IEEE Transactions on Audio, Speech, and Lan- guage Processing, vol. 19, no. 7, pp. 2125–2136, 2011

    2011

  26. [34]

    Mindspore,

    Huawei, “Mindspore,” https://www.mindspore.cn/, 2020

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.