Pith. sign in

REVIEW 4 major objections 3 minor 41 references

Running Conventional Automatic Speech Recognition on Memristor Hardware: A Simulated Approach

T0 review · 4 major / 3 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 A 42-million-parameter speech recognizer can run its linear layers through simulated memristor hardware with only about a 28% relative word-error increase.

desk verdict A genuine first simulation of a 42M-parameter ASR model on realistic memristor hardware, but the headline 25% and the '3-bit' label both need correction. read the letter →

arxiv 2505.24721 v1 pith:NLAYSVBP submitted 2025-05-30 cs.LG cs.ARcs.ET

Nick Rossenbach , Benedikt Hilmes , Leon Brackmann , Moritz Gunz , Ralf Schlüter This is my paper · ORCID
classification cs.LGcs.ARcs.ET
keywords memristorsimulationautomaticspeechrecognitionConformerquantization-awaretraininganalogin-memorycomputingcrossbararrayTED-LIUMv2worderrorrate
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

Using a variability-accurate simulation of analog memristor crossbars, this paper claims that a conventional, 42-million-parameter Conformer speech recognizer can move its fixed-weight linear computations onto memristor hardware and still recognize TED talks almost as well as on a GPU. The trick is quantization-aware training that forces weights into a symmetric 3-bit format before deployment; with that preparation, the simulated device noise raises word error rate from 7.2% to 9.2% on average over ten independently programmed crossbar instances. This matters because earlier memristor studies only handled small tasks such as MNIST or single-word speech, and current prototype chips cannot physically fit large models. If the simulation is faithful, non-deterministic memristor behavior is not the obstacle to large-scale neural inference; chip-scale integration and crossbar count are.

What carries the argument

The central mechanism is a paired-memristor crossbar with binary conductance states: each weight is represented as the difference between two cells, so zero and negative weights are possible, and multiple stacked crossbars supply higher bit precision (three crossbars give 15 levels from -7 to 7). The device variability comes from the Synaptogen generative model, which reproduces programming-cycle and thermal noise learned from over three million measured electrical cycles on foundry-quality oxide memristors. The simulation pipeline quantizes inputs according to tracked DAC scales, applies voltages to simulated cells, reads output currents, applies a fitted correction factor, quantizes through the ADC, and bit-shifts results back to the model's numeric range. Quantization-aware training is the other load-bearing piece: it makes the network learn a symmetric weight distribution and tolerate the conductance noise that the hardware will inject.

What would settle it

Build or use a real 128x128 paired-memristor crossbar of the modeled 130 nm STMicroelectronics devices with 8-bit DAC/ADC, program the QAT-trained 3-bit weights, and measure TED-LIUMv2 dev WER; if the observed WER falls outside the simulation's 9.0-9.4% range (averaged over at least ten programming cycles), the device model does not capture the hardware's true error statistics.

Watch

Extended reading notes

Core claim

The paper's central claim is that a standard large ASR model can be mapped onto realistically modeled memristor hardware with limited accuracy loss. Concretely, a 12-layer Conformer encoder with 42M parameters, trained with quantization-aware learning at 3-bit symmetric weights, runs all of its parameterized linear layers through a simulated crossbar array of paired binary-conductance memristors (three stacked crossbars per weight, 8-bit DAC/ADC, 128x128 tiling), with only the depthwise convolution left out. The result on TED-LIUMv2 dev is 9.2% average WER, up from a 7.2% non-quantized baseline; the spread across 10 device programming runs is 9.0-9.4%. The authors interpret this as showing that the intrinsic stochasticity and nonlinearity of current memristor devices is tolerable for a large NLP-scale task, provided the model is trained with quantization noise from the start.

Load-bearing premise

The reliability of the result rests on the Synaptogen model being an accurate proxy for a real memristor chip used in this configuration, including programming yield, drift, tiling, and DAC/ADC behavior, and on the exact FP32 execution of attention and normalization not masking errors that would appear in an integrated system.

Editorial extensions

If this is right

  • Post-training quantization is not viable for this route: at 3-bit weights it degrades WER from 7.2% to 30.7%, while quantization-aware training at the same precision stays at 8.3% before hardware simulation.
  • If the simulation transfers to real hardware, binary-paired memristor cells are precise enough for large ASR with 3-bit effective weights; multi-level conductance states are not required.
  • The run-to-run variation across programmed devices is small (standard deviation about 0.1 WER, max-min 0.4), so device mismatch alone need not force per-device retraining.
  • For a fixed-weight linear layer to move to crossbars, the paper's recipe is complete: retrain with symmetric fake quantization, map to stacked paired cells, and read out with the correction factor; no further hardware-specific adaptation is needed.
  • With sufficient crossbars on a chip, the same mapping should apply to other transformer-style encoders whose parameterized operations are linear, not only to Conformer.

Reading between the lines

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

  • The paper only simulates the linear layers; attention, layer norm, feature extraction, and decoding stay in exact floating point. A full hardware deployment would move some of those operations onto the same or neighboring chips, and their non-idealities are not captured in the reported WER.
  • Because the memory footprint scales as one paired crossbar stack per weight bit, the real constraint at 3-bit precision is chip crossbar capacity; the simulation suggests algorithmic readiness may arrive before hardware can fit a whole 42M-parameter encoder.
  • A direct testable extension would be to rerun the same QAT-plus-simulation pipeline on an RNN-Transducer or streaming attention model; if its WER degradation stays near the same 10-30% relative band, the conclusion would generalize beyond offline CTC Conformers.
  • The abstract's '25%' degradation corresponds to the best of the ten simulated device draws (9.0% WER), while the reported average is 9.2%, a 27.8% relative increase over baseline.
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 / 3 minor

Summary. The manuscript presents a PyTorch-based simulation framework that connects the Synaptogen memristor device model to a conventional ASR pipeline. The authors train a 42M-parameter Conformer CTC model on TED-LIUMv2 with quantization-aware training (QAT), then simulate execution of its linear layers on paired memristor crossbars with 8-bit DAC/ADC quantization, 128x128 tiling, and 10 independently programmed device draws. They report word error rates for 8-bit down to 2-bit weight resolutions and conclude that, with QAT, a 3-bit weight setting limits the relative WER degradation to 25% compared with the non-quantized baseline. The main contributions claimed are a large-scale memristor simulation for speech recognition, a public implementation, and a demonstration that QAT can make a realistic device model usable without hardware-specific tuning.

Significance. If the result holds, this is a useful step beyond the MNIST-scale or isolated-word demonstrations common in the memristor-ML literature. The evaluation is externally grounded on a standard ASR benchmark rather than a synthetic proxy, the device noise is modeled from measured programming cycles, and the authors report multiple device draws with means and spreads. The comparison of PTQ and QAT in Table 2 is informative and the release of code supports reproducibility. However, the headline quantitative claim is not precisely supported: the 25% figure is a best-case draw rather than the average behavior, and the label '3-bit' is ambiguous because the three-crossbar stack implements 15 weight levels. The simulation also covers only the linear layers, with attention, normalization, feature extraction, and decoding running in exact FP32, so the abstract's phrasing about how a full ML system 'would behave' overstates the scope.

major comments (4)
  1. [Abstract; §6.3, Table 3] The abstract's claim of limiting relative WER degradation to 25% is not what Table 3 shows on average. For the 3-bit row, the mean WER is 9.2% versus the 7.2% baseline, which is a 27.8% relative degradation; the 25% figure corresponds to the minimum over the ten device runs (9.0% WER). Please state the mean as the headline result, or explicitly label 25% as the best-case draw, and report the average relative degradation alongside the range.
  2. [§2.3; §5] The term '3-bit weight precision' is not well-defined. Section 2.3 describes a stack of three crossbars with paired memristors yielding 15 weight levels from -7 to 7, i.e., log2(15) ≈ 3.9 bits of resolution, and the example in Section 5 repeatedly refers to a '3-crossbar stack' with levels -7 to 7. The text is also internally inconsistent: the example is introduced with the phrase 'for a 4-bit resolution' while describing three crossbars. The paper does not state whether the '3-bit' rows in Tables 2 and 3 use an 8-level QAT codebook mapped onto the 15 physical levels or the 15-level codebook itself. Please define the codebook and the mapping explicitly; otherwise the central precision claim is ambiguous.
  3. [§2.2; §2.3; §7] The realism of the simulation is asserted rather than validated for the specific deployment scenario. Section 2.2 states that Synaptogen was trained on measured electrical characteristics from over 3 million programming cycles of STMicroelectronics 130 nm devices, but no evidence is provided that the same error statistics hold for the paper's use of 128x128 paired crossbars, the particular voltage programming routine, and the assumed 8-bit DAC/ADC conversions. Since the conclusion that memristor non-determinism 'poses no restrictions' depends on this transfer, please either provide calibration or validation evidence, or explicitly condition the conclusions on the Synaptogen device model and its measured-device basis.
  4. [Abstract; §3; §5] Only the linear layers of the Conformer encoder are simulated on the memristor model; feature extraction, the down-sampling convolutional stack, the depthwise convolution, self-attention dot products, layer normalization, and decoding run in exact FP32. The abstract's phrase 'how an ML system with millions of parameters would behave on memristor hardware' therefore overstates the evidence. The fraction of trainable weights inside the crossbars is large, but that is not the same as the fraction of compute or the fraction of sensitivity to noise, and the paper does not quantify those. Please rephrase the claims to refer to the linear operations, and ideally quantify how much of the encoder's computation is simulated.
minor comments (3)
  1. [§1; §4] There are small typos: 'Artifical' in the first sentence of the Introduction and 'usualy' in Section 4 should be corrected to 'Artificial' and 'usually' respectively.
  2. [§5] The example in Section 5 gives a weight scaling factor of 70 for a 3-crossbar stack without showing the underlying formula; please add the general expression relating the observed value range, the number of stacked crossbars, and the scale factors.
  3. [Table 2] The PTQ entry at 2-bit resolution is listed as 'n.a.', but the reason is not stated; a sentence explaining whether this is due to numerical overflow, unstable training, or an arbitrary cutoff would improve the table.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the WER result is an external measurement on TED-LIUMv2 through a device model fitted to measured electrical data; the reported '3-bit' and '25%' discrepancies are labeling and reporting issues, not circular steps.

full rationale

The central claim—that a 42M-parameter Conformer ASR system can run its linear layers through simulated analog memristor crossbars and retain reasonable WER—is an empirical measurement against the external TED-LIUMv2 corpus, not a derivation from the paper's own assumptions. The WER values in Table 3 come from actual simulated runs over 10 randomly drawn device instances and are decoded with an external search and language model. The correction factor c in Section 2.3 is a disclosed calibration constant fitted to the Synaptogen device model's conductance statistics, not to WER, so it cannot force recognition performance. Synaptogen [8], although co-authored by an author of this paper, is grounded in 'measured electrical characteristics of over 3 million programming cycles' on commercial foundry devices, which is independent external evidence under the review rules. The paper's own Section 7 further limits the claims (no energy/speed assumptions, single device type, linear layers only), reinforcing that the result is a conditional measurement. Two reviewer concerns—the ambiguous '3-bit' label given the 15-level crossbar stack (log2(15) ≈ 3.9 bits) and the abstract's '25%' matching only the best of 10 runs while the mean degradation is 27.8%—are precision and reporting issues, not circular ones. No fitted quantity is renamed as a prediction, no uniqueness theorem is imported from prior work, and no known result is repackaged as a derivation. Finding: no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central result rests on a validated-looking but unvalidated device model, a calibration constant c, and an exact-computation assumption for the unsimulated parts of the pipeline. These are disclosed, but they are assumptions rather than independent evidence, so the paper's main number should be read as conditional on Synaptogen's fidelity.

free parameters (1)
  • current-to-value correction factor c = 8020 1/A
    Section 2.3 defines c as the optimal value with respect to a quadratic error over Synaptogen simulated paired cells. It is used in Eq. (1) through (3) to convert output currents into trained weight values for every simulated linear layer.
assumptions (4)
  • domain assumption Synaptogen's learned device model faithfully represents the statistical behavior of real memristor hardware at scale.
    Section 2.2 asserts the model was trained on measured characteristics from over 3 million programming cycles, but the paper does not validate extrapolation to a large stacked-crossbar system.
  • ad hoc to paper A fixed correction factor c can linearly map paired-cell output currents to trained weight values across all layers and bit levels.
    Section 2.3 introduces c=8020 1/A fitted to minimize quadratic error on simulated paired cells; Table 1 shows nonlinearity and noise remain for basic operations such as 1.0 x 1 with standard deviation 0.063.
  • domain assumption Operations not mapped to memristors, including self-attention dot products, layer-norm, gating, feature extraction, depthwise convolution, and decoding, can be treated as exact and error-free without biasing the comparison.
    Section 3 excludes these operations by design and Section 5 says parameter-free operations are computed as usual. No error-budget analysis is provided to show their exactness does not hide memristor-induced failures.
  • domain assumption 128x128 crossbar tiling and 8-bit DAC/ADC resolutions are realistic hardware constraints that do not distort the word error rate results.
    Section 2.3 fixes tile size at 128x128 based on a cited scaling limits reference and assumes 8-bit converters. These choices bound the simulation but their impact on the final WER is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Running Conventional Automatic Speech Recognition on Memristor Hardware: A Simulated Approach." pith.science (2026). https://pith.science/paper/NLAYSVBP

@misc{pith2026250524721,
  author       = {Pith},
  title        = {Pith review of: Running Conventional Automatic Speech Recognition on Memristor Hardware: A Simulated Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NLAYSVBP}},
  note         = {Machine review of arXiv:2505.24721}
}
read the original abstract

Memristor-based hardware offers new possibilities for energy-efficient machine learning (ML) by providing analog in-memory matrix multiplication. Current hardware prototypes cannot fit large neural networks, and related literature covers only small ML models for tasks like MNIST or single word recognition. Simulation can be used to explore how hardware properties affect larger models, but existing software assumes simplified hardware. We propose a PyTorch-based library based on "Synaptogen" to simulate neural network execution with accurately captured memristor hardware properties. For the first time, we show how an ML system with millions of parameters would behave on memristor hardware, using a Conformer trained on the speech recognition task TED-LIUMv2 as example. With adjusted quantization-aware training, we limit the relative degradation in word error rate to 25% when using a 3-bit weight precision to execute linear operations via simulated analog computation.

Figures

Figures reproduced from arXiv: 2505.24721 by the authors.

Figure 1
Figure 1. Schematic representation of a Vector-matrix￾multiplication in a 2 × 2 memristor crossbar. A weight ma￾trix is programmed as conductance level of the memristors, Gi,j (i, j ∈ [1, 2]). An input vector (V1, V2) is applied at the horizontal lines. The resulting current per column, I1,2 yields the sum of the products between the inputs and the conductance. these variabilities are comparable to the phenomena of mismatch a… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages

  1. [1]

    Introduction Artifical neural networks (ANN) play an important role in natu- ral language processing (NLP) tasks such as automatic speech recognition (ASR). The majority of current ANN architectures used in NLP such as LSTM [1] or Transformer [2] and related derivatives make use of tensor operations such as vector-matrix- multiplication (VMM). VMM-based n...

    work page

  2. [2]

    Running Conventional Automatic Speech Recognition on Memristor Hardware: A Simulated Approach

    Simulated Memristor Hardware A memristor is a passive device whose resistance changes in response to the application of a sufficiently large electric voltage [5]. After the voltage is removed, the resistance of the device is conserved until a sufficiently large voltage is applied again. By varying the voltage amplitude, the resistance of the device can be...

    work page Pith review arXiv 2025

  3. [3]

    We chose the Conformer as encoder architecture as it is widely used in current research literature [27]

    Automatic Speech Recognition Our ASR system consists of a Conformer encoder [25] with a Connectionist-Temporal-Classification (CTC) [26] output loss layer. We chose the Conformer as encoder architecture as it is widely used in current research literature [27]. While most layers in a Conformer block make use of VMMs with static matrices where modeling with...

    work page

  4. [4]

    While PTQ works well for 8-bit precision formats, the performance degrades with decreas- ing bit depth [30]

    Quantization-Aware Training Previous research utilizes already trained models, usualy using post-training quantization (PTQ) to quantize the models for fur- ther usage, as also done in MemTorch. While PTQ works well for 8-bit precision formats, the performance degrades with decreas- ing bit depth [30]. This is because lower bit precision introduces more n...

    work page

  5. [5]

    Simulated Execution of Conformer Blocks During QAT, we have observers that track the statistics of each weight and input. Assuming that for a specific linear layer we would have tracked the largest absolute input value to be 4 and the largest absolute weight value to be 0.1, we would have the input scaling factor 0.25 and a weight factor of 70 for a 3- cr...

    work page

  6. [6]

    Experimental Setup We conduct our experiments on TED-LIUMv2 (TEDv2) [17], which consists of 207 hours of TED talks

    Experiments 6.1. Experimental Setup We conduct our experiments on TED-LIUMv2 (TEDv2) [17], which consists of 207 hours of TED talks. We use the dev-set of TED-LIUMv2 for evaluation. For recognition, we use a 4-gram LM trained on the text data provided with the corpus. The model consists of 12 Conformer layers with a model di- mension of 384 and a feed-for...

    work page

  7. [7]

    As chip co-integration has been disregarded in this work, we can not make any assumptions yet about the speed and energy benefit achieved by using memristor hardware

    Limitations and Future Work This work only investigates the effect of a particular kind of memristor hardware on linear transformation layers for a sin- gle speech recognition system. As chip co-integration has been disregarded in this work, we can not make any assumptions yet about the speed and energy benefit achieved by using memristor hardware. In the...

    work page

  8. [8]

    Conclusion In this work, we presented how a Conformer-based ASR model performs on simulated memristor hardware that introduces a strong uncertainty in tensor operations. We present a PyTorch extension to the Synaptogen toolkit to easily map nn.Linear lay- ers in order to allow neural networks to be executed via a realistic memristor device simulation. We ...

    work page

Show all 41 references
  1. [9]

    Clusters4Future

    Acknowledgments This work was partially supported by NeuroSys, which as part of the initiative “Clusters4Future” is funded by the Federal Ministry of Education and Research BMBF (funding IDs 03ZU2106DA and 03ZU2106DD)

  2. [10]

    Long short-term memory,

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

    1997

  3. [11]

    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,” in Advances in Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA, Dec. 2017, pp. 6000–6010

    2017

  4. [12]

    TPU v4: An optically reconfig- urable supercomputer for machine learning with hardware support for embeddings,

    N. P. Jouppi, G. Kurian, S. Li, P. C. Ma, R. Nagarajan, L. Nai, N. Patil, S. Subramanian, A. Swing, B. Towles, C. Young, X. Zhou, Z. Zhou, and D. A. Patterson, “TPU v4: An optically reconfig- urable supercomputer for machine learning with hardware support for embeddings,” in P...

    2023

  5. [13]

    Accelerating ml workloads using gpu tensor cores: The good, the bad, and the ugly,

    B. Hanindhito and L. K. John, “Accelerating ml workloads using gpu tensor cores: The good, the bad, and the ugly,” in Proceedings of the 15th ACM/SPEC International Conference on Performance Engineering, ser. ICPE ’24. New York, NY , USA: Association for Computing Machinery, 2...

    2024

  6. [14]

    The missing memristor found,

    D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, “The missing memristor found,” Nature, vol. 453, no. 7191, pp. 80–83, May 2008

    2008

  7. [15]

    Hardware implementation of memristor-based artificial neural networks,

    F. Aguirre, A. Sebastian, M. Le Gallo, W. Song, T. Wang, J. J. Yang, W. Lu, M.-F. Chang, D. Ielmini, Y . Yang, A. Mehonic, A. Kenyon, M. A. Villena, J. B. Rold ´an, Y . Wu, H.-H. Hsu, N. Raghavan, J. Su ˜n´e, E. Miranda, A. Eltawil, G. Setti, K. Smagulova, K. N. Salama, O. Kre...

    1974

  8. [16]

    A compute-in-memory chip based on resistive random-access memory,

    W. Wan, R. Kubendran, C. Schaefer, S. B. Eryilmaz, W. Zhang, D. Wu, S. Deiss, P. Raina, H. Qian, B. Gao, S. Joshi, H. Wu, H.- S. P. Wong, and G. Cauwenberghs, “A compute-in-memory chip based on resistive random-access memory,”Nature, vol. 608, no. 7923, pp. 504–512, 08 2022

    2022

  9. [17]

    Synaptogen: A cross- domain generative device model for large-scale neuromorphic circuit design,

    T. Hennen, L. Brackmann, T. Ziegler, S. Siegel, S. Menzel, R. Waser, D. J. Wouters, and D. Bedau, “Synaptogen: A cross- domain generative device model for large-scale neuromorphic circuit design,” IEEE Transactions on Electron Devices, vol. 71, no. 9, pp. 5345–5353, 2024

    2024

  10. [18]

    MemTorch: An Open-source Simulation Framework for Memris- tive Deep Learning Systems,

    C. Lammie, W. Xiang, B. Linares-Barranco, and M. R. Azghadi, “MemTorch: An Open-source Simulation Framework for Memris- tive Deep Learning Systems,” Neurocomputing, 2022

    2022

  11. [19]

    A flexible and fast pytorch toolkit for simulating training and infer- ence on analog crossbar arrays,

    M. J. Rasch, D. Moreda, T. Gokmen, M. Le Gallo, F. Carta, C. Goldberg, K. El Maghraoui, A. Sebastian, and V . Narayanan, “A flexible and fast pytorch toolkit for simulating training and infer- ence on analog crossbar arrays,” in 2021 IEEE 3rd International Conference on Artifi...

    2021

  12. [20]

    Fusion of memristor and digital compute-in-memory processing for energy-efficient edge computing,

    T.-H. Wen, J.-M. Hung, W.-H. Huang, C.-J. Jhang, Y .-C. Lo, H.-H. Hsu, Z.-E. Ke, Y .-C. Chen, Y .-H. Chin, C.-I. Su, W.-S. Khwa, C.-C. Lo, R.-S. Liu, C.-C. Hsieh, K.-T. Tang, M.-S. Ho, C.-C. Chou, Y .-D. Chih, T.-Y . J. Chang, and M.-F. Chang, “Fusion of memristor and digital ...

    2024

  13. [21]

    Neuromorphic circuit simulation with memristors: Design and evaluation using memtorch for mnist and cifar,

    J. Souto, G. Botella, D. Garc ´ıa, R. Murillo, and A. del Barrio, “Neuromorphic circuit simulation with memristors: Design and evaluation using memtorch for mnist and cifar,” 2024

    2024

  14. [22]

    Rxnn: A framework for evaluating deep neural networks on resistive cross- bars,

    S. Jain, A. Sengupta, K. Roy, and A. Raghunathan, “Rxnn: A framework for evaluating deep neural networks on resistive cross- bars,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 40, no. 2, pp. 326–338, 2021

    2021

  15. [23]

    MNIST handwritten digit database,

    Y . LeCun and C. Cortes, “MNIST handwritten digit database,” 2010

    2010

  16. [24]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” Tech. Rep., 2009

    2009

  17. [25]

    Speech Commands: A Dataset for Limited-V ocabulary Speech Recognition,

    P. Warden, “Speech Commands: A Dataset for Limited-V ocabulary Speech Recognition,” ArXiv e-prints, Apr. 2018

    2018

  18. [26]

    Enhancing the ted-lium corpus with selected data for language modeling and more ted talks,

    A. Rousseau, P. Del´eglise, and Y . Est`eve, “Enhancing the ted-lium corpus with selected data for language modeling and more ted talks,” in International Conference on Language Resources and Evaluation, 2014

    2014

  19. [27]

    Sisyphus, a workflow manager designed for machine translation and automatic speech recogni- tion,

    J. Peter, E. Beck, and H. Ney, “Sisyphus, a workflow manager designed for machine translation and automatic speech recogni- tion,” in EMNLP 2018: System Demonstrations, Brussels, Belgium, October 31 - November 4, 2018, pp. 84–89

    2018

  20. [28]

    Returnn: The RWTH extensible training framework for universal recurrent neural networks,

    P. Doetsch, A. Zeyer, P. V oigtlaender, I. Kulikov, R. Schl ¨uter, and H. Ney, “Returnn: The RWTH extensible training framework for universal recurrent neural networks,” in ICASSP 2017, New Orleans, LA, USA, March 5-9, 2017, pp. 5345–5349

    2017

  21. [29]

    Memristive devices for computing,

    J. J. Yang, D. B. Strukov, and D. R. Stewart, “Memristive devices for computing,” Nature Nanotechnology, vol. 8, no. 1, pp. 13–24, Jan 2013

    2013

  22. [30]

    D. J. Wouters, Y .-Y . Chen, A. Fantini, and N. Raghavan,Reliability Aspects. John Wiley & Sons, Ltd, 2016, ch. 21, pp. 597–622

    2016

  23. [31]

    Infineon Technologies AG. (2022). [Online]. Avail- able: https://www.infineon.com/cms/de/about-infineon/press/ market-news/2022/INFATV202211-031.html

    2022

  24. [32]

    Memristor-based hardware accelerators for artificial intel- ligence,

    Y . Huang, T. Ando, A. Sebastian, M.-F. Chang, J. J. Yang, and Q. Xia, “Memristor-based hardware accelerators for artificial intel- ligence,” Nature Reviews Electrical Engineering, vol. 1, no. 5, pp. 286–299, May 2024

    2024

  25. [33]

    Scaling limits of memristor-based routers for asynchronous neuromorphic systems,

    J. Chen, S. Yang, H. Wu, G. Indiveri, and M. Payvand, “Scaling limits of memristor-based routers for asynchronous neuromorphic systems,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 71, no. 3, pp. 1576–1580, 2024

    2024

  26. [34]

    Gulati, C.-C

    A. Gulati, C.-C. Chiu, J. Qin, J. Yu, N. Parmar, R. Pang, S. Wang, W. Han, Y . Wu, Y . Zhang, and Z. Zhang, Eds., Con- former: Convolution-augmented Transformer for Speech Recogni- tion, 2020

    2020

  27. [35]

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

    A. Graves, S. Fern´andez, and F. Gomez, “Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks,” in In Proceedings of the International Confer- ence on Machine Learning, ICML, 2006, pp. 369–376

    2006

  28. [36]

    End-to-end speech recognition: A survey,

    R. Prabhavalkar, T. Hori, T. N. Sainath, R. Schl¨uter, and S. Watan- abe, “End-to-end speech recognition: A survey,”IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 32, pp. 325–351, 2024

    2024

  29. [37]

    Flashlight: Enabling inno- vation in tools for machine learning,

    J. D. Kahn, V . Pratap, T. Likhomanenko, Q. Xu, A. Hannun, J. Cai, P. Tomasello, A. Lee, E. Grave, G. Avidov, B. Steiner, V . Liptchin- sky, G. Synnaeve, and R. Collobert, “Flashlight: Enabling inno- vation in tools for machine learning,” in Proceedings of the 39th Internation...

    2022

  30. [38]

    KenLM: Faster and smaller language model queries,

    K. Heafield, “KenLM: Faster and smaller language model queries,” in Proceedings of the Sixth Workshop on Statistical Machine Trans- lation. Edinburgh, Scotland: Association for Computational Linguistics, Jul. 2011, pp. 187–197

    2011

  31. [39]

    A survey of quantization methods for efficient neural network inference,

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantization methods for efficient neural network inference,” 2021

    2021

  32. [40]

    On the variance of the adaptive learning rate and beyond,

    L. Liu, H. Jiang, P. He, W. Chen, X. Liu, J. Gao, and J. Han, “On the variance of the adaptive learning rate and beyond,” in8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020, 2020

    2020

  33. [41]

    Sequence Transduction with Recurrent Neural Net- works,

    A. Graves, “Sequence Transduction with Recurrent Neural Net- works,” in Proc. ICML Workshop on Representation Learning , Edinburgh, Scotland, Jun. 2012

    2012

Pith tools

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