Pith. sign in

REVIEW 3 major objections 5 minor 43 references

CAPS: A Cascaded Reconstruction Model to Power Saving in Hearables Using Sub-Nyquist Sampling with Bandwidth Extension

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read CAPS deliberately undersamples hearable audio to 4 kHz at 8 bits, then reconstructs clean wideband speech on the phone, cutting ADC power 3.31x while keeping quality.

desk verdict A real engineering contribution with an oversold battery-life claim: the ADC-only 3.31x saving is real, but the paper's own power law contradicts it by ~29x, and the system-level extrapolation is unsupported. read the letter →

arxiv 2607.19434 v1 pith:VD4N2VXH submitted 2026-07-21 cs.SD

Tarikul Islam Tamiti , Sajid Fardin Dipto , Luke Baja-Ricketts , David Vergano , Anomadarshi Barua This is my paper · ORCID
classification cs.SD
keywords hearablessub-Nyquistsamplingbandwidthextensionspeechenhancementboneconductionmicrophonelow-powerADCcascadedneuralnetworkstreaminginference
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 establish that a hearable can save a large fraction of its ADC power by deliberately sampling at 4 kHz with 8-bit resolution, far below the usual 24 kHz/12-bit, and then handing the low-quality signal to a small neural network running on a paired phone, which rebuilds clean wideband speech. The proposed system, CAPS, combines bandwidth extension and multimodal speech enhancement, using both air- and bone-conduction microphones, in one cascaded model. Measured ADC current drops from 2.325 mW to 0.702 mW, a 3.31x reduction, while objective quality scores stay level with or above larger baselines, and on-phone inference finishes in 55.11 ms per frame, under the 150 ms streaming threshold. A sympathetic reader would take this as evidence that the power-quality trade-off in hearables can be moved off the ear and onto the pocket.

What carries the argument

The load-bearing object is the cascaded reconstruction chain: SEN (a 2D convolutional U-Net with a Mamba bottleneck) performs spectral bandwidth extension; UN (four transposed-convolution stages with dilated residual stacks) performs 256x waveform upsampling; APEN (two coupled depthwise/pointwise 1D-convolution streams) separates amplitude and wrapped phase, uses the bone-conduction waveform as a noise-free reference, and applies anti-wrapping phase loss. Together the chain converts the deliberately degraded 4 kHz/8-bit input into a clean 16 kHz or 22 kHz signal, so the hearable can stay power-lean while the reconstruction cost is paid on the phone's larger battery.

What would settle it

Measure total current draw of a complete hearable (ADC, Bluetooth radio, codec, DSP, and amplifier) at {24 kHz, 12-bit} versus {4 kHz, 8-bit} with CAPS in the loop; if the total battery-life extension is far below 3.31x, the headline claim is not an end-to-end property of the device, only of the ADC block.

Watch

Extended reading notes

Core claim

CAPS claims that sub-Nyquist, low-bit ADC sampling in hearables is usable in practice if a cascaded neural generator is available on the paired mobile device. The capture side reduces the ADC setting from {24 kHz, 12-bit} to {4 kHz, 8-bit}, a 3.31x measured power reduction. The reconstruction side first runs a spectral U-Net (SEN) to produce a high-resolution time-frequency representation, then a HiFi-GAN-style upsampler (UN) expands the waveform 256x, and finally an amplitude-phase enhancement network (APEN) fuses the cleaner bone-conduction vibration signal with the enhanced air-conduction stream, separately restoring amplitude and wrapping-corrected phase. Across VCTK speech and MagnaTagA

Load-bearing premise

The 3.31x saving is an ADC-only measurement; Section 6 explicitly says codec and encryption are excluded, and Section 2.1's formula P = k * fs * 2^N implies a far larger saving, so the headline factor rests on the measured ADC current table rather than an end-to-end device budget.

Editorial extensions

If this is right

  • If the measurements hold in a full device, hearable battery life could rise by up to 3.31x without requiring users to accept noticeably worse speech quality.
  • Streaming speech enhancement becomes feasible on ordinary phones: 55.11 ms per inference is below the 150 ms one-way latency recommendation.
  • The same 11 MB model can run on different phone hardware (Pixel 7 and Galaxy S21) with only modest slowdown, suggesting the approach is portable.
  • Joint bandwidth extension and multimodal enhancement at 8-bit ADC resolution means low-cost vibration sensors and simple ADCs become acceptable for noisy real-world use.
  • Because the model also works on music, the reconstruction approach is not speech-specific.

Reading between the lines

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

  • Editorial inference: the 3.31x figure is an ADC-only saving; the paper's own conclusion section says codec and encryption are excluded, and it does not account for BLE, DSP, or amplifier draw, so the true end-to-end battery gain for a real earbud is likely smaller.
  • Editorial inference: the paper's power model P = k * fs * 2^N would predict a far larger ratio than 3.31x, so the headline saving rests on the measured ADC current table, not the formula; a device-level power model is the next test.
  • Editorial inference: the 'sample poor, reconstruct rich' strategy could generalize beyond speech to any wearable sensor that can offload reconstruction to a phone, such as vibration sensing or bio-signals.
  • Editorial inference: a testable extension is to push the sampling point lower, e.g., 2 kHz at 6 bits, and measure the quality cliff, trading more power for the increasing difficulty of bandwidth extension.
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

3 major / 5 minor

Summary. CAPS proposes a cascaded reconstruction model (SEN, UN, APEN) that reconstructs wideband, clean audio from sub-Nyquist, low-bit ADC captures of acoustic and bone-conduction signals. The authors report that reducing sampling from {24 kHz, 12-bit} to {4 kHz, 8-bit} yields a 3.31x power reduction and an ideal ~3.31x battery-life increase for hearables; on quality metrics CAPS is competitive with or better than six baselines on VCTK and MagnaTagATune, with 1.36 ms desktop / 55.11 ms Pixel7 inference, below the ITU streaming threshold. They evaluate across 8/10/12-bit resolutions, on two mobile platforms, and provide ablations of the main components.

Significance. If the quality and latency results hold, CAPS is a meaningful contribution: it is the first system in the cited comparison set to jointly address sub-Nyquist sampling, low-bit ADC resolution, bandwidth extension, and multimodal speech enhancement. The evaluation is comparatively thorough—multiple objective metrics, two datasets, two mobile platforms, and ablations—and the mobile inference figure (55.11 ms on Pixel7) is well below the ITU G.114 bound. The paper also includes an honest limitation statement in §6. However, the headline power-saving and battery-life claims are not supported at system level; the power contribution is demonstrated only for the ADC, and the paper's own power model is inconsistent with the measurement table. These issues are central to the paper's main claim and need to be addressed before acceptance.

major comments (3)
  1. [§5.2, Table 5; Abstract; §6] The headline '3.31x reduction in power consumption in hearables' and 'increase battery life by ~3.31x' is computed solely from Table 5, which reports NRF52840 ADC current at two configurations. This is an ADC-only measurement. The paper's own §6 explicitly excludes codec and encryption, and BLE transmission, sensor conditioning, and hearable-side DSP are not included. For a real hearable, the battery-life multiplication factor depends on the ADC's fraction f of the total power budget: the factor is at most 1/((1-f)+f/3.31), approaching 3.31 only as f→1. Since the ADC is typically not the dominant consumer in a hearable (BLE radios alone draw tens of mA during transmission), the claimed battery-life improvement is unsupported. Please supply a system-level power budget or explicitly restrict the claims to ADC power and remove the battery-life generalization.
  2. [§2.1 vs Table 5] The stated power model P = k·fs·2^N predicts a 96x reduction for the headline change from {24 kHz, 12-bit} to {4 kHz, 8-bit}: (24/4)×(2^12/2^8) = 6×16 = 96. Table 5 reports only a 3.31x ratio. Similarly, §2.1's claimed 2.45x saving between {16 kHz, 12-bit} and {4 kHz, 8-bit} would be 64x under the same model. The discrepancy (roughly a factor of 29 for the headline claim) is far too large for measurement noise; it implies fixed overheads or nonlinear power scaling that the text does not state. As written, the theoretical model contradicts the measurements, so it cannot be used to motivate extrapolation from the measured ADC table to system-level battery life.
  3. [§5.2, Table 7] The text states that CAPS will restore low-resolution audio to 'provide the same audio quality.' Table 7 shows a small but consistent degradation at the headline 8-bit operating point relative to 12-bit: PESQ 2.93 vs 2.99, VISQOL 4.09 vs 4.15, SI-SDR 16.84 vs 16.99. CAPS still outperforms the best baselines, so the quality contribution is not invalidated, but the 'same audio quality' claim is stronger than the evidence. Please qualify this statement with the measured 8-bit results or provide a statistical test showing the differences are not meaningful.
minor comments (5)
  1. [Abstract] The abstract reports 'inference time of 1.36 ms' in the context of mobile platforms. Table 4 shows 1.36 ms is the desktop GPU number; the Pixel7 inference time is 55.11 ms. Please state both explicitly to avoid misrepresenting the mobile result.
  2. [§3.4] The multi-period loss equation is not written as a mean absolute error: as printed, it appears to be the absolute value of the sum of differences rather than the sum of absolute per-element differences. Please add the missing absolute-value bars/summation notation so the loss matches the MAE description.
  3. [Table 7] The HiFi++ row for 8-bit lists '0.91/93'; this appears to be a typo for '0.91/0.93'.
  4. [§5.1 heading] The heading reads 'Multimodel SE'; this should be 'Multimodal SE.'
  5. [§5.2, Table 6] Please clarify in the table caption or text whether 'Power for CAPS only' is the measured power of the model on Pixel7 or a simulation, and how the 'Total power' is derived. Also specify the measurement conditions (e.g., screen state, radio state) since the 1.15 W average is later compared with battery capacity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the power-saving headline is a direct ADC measurement, and the quality results come from held-out comparative evaluation; the inconsistency with the paper's theoretical power model is a correctness/scope issue, not a circular derivation.

full rationale

The paper's central quantitative claim—3.31x power reduction in hearables—is taken directly from Table 5, which reports measured NRF52840 ADC currents at {4 kHz, 8-bit} (234 µA / 0.702 mW) and {24 kHz, 12-bit} (775 µA / 2.325 mW), with the paper stating "2.325/0.702 = 3.31x power." This is an empirical measurement ratio, not a quantity derived from a model whose inputs already contain the answer. The theoretical relation P = k·fs·2^N in §2.1 is used only as motivational context and is, if anything, inconsistent with the measured Table 5 values (it would imply a much larger ratio); the paper does not fit the constant k to the data and then call the result a prediction. The choice of {4 kHz, 8-bit} as the operating point is a design decision, not a fitted parameter, and comparing it against a conventional {24 kHz, 12-bit} baseline is standard evaluation, not circular reasoning. The speech-quality results in Tables 3, 4, 7, and 8 are obtained by training CAPS and comparing against six external baselines on the same noisy speech/music conditions; no baseline is derived from CAPS's own outputs. There is no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in through the authors' prior work; cited methods such as HiFi-GAN, MelGAN, Mamba, and ITU G.114 are independent external sources. The paper's own §6 limitation—excluding codec and encryption, and not accounting for the full hearable power budget—undermines the system-level battery-life extrapolation, but that is a validity/scope concern about an unsupported generalization, not a circularity in the derivation. Therefore no circular step is exhibited, and the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central assumptions are design-domain assumptions about ADC power scaling, the recoverability of wideband speech from 4 kHz 8-bit input, and the usefulness of bone-conduction for phase reconstruction. The inconsistent power law is the most fragile item: it is asserted in §2.1 but not reflected in the measured Table 5.

free parameters (3)
  • ADC power-law constant k
    Introduced in §2.1 as P = k·fs·2^N. No value is fitted or reported, and the formula is inconsistent with Table 5; the 3.31x saving comes from measured rows, not from k.
  • ADC operating point {4 kHz, 8-bit} = 4 kHz, 8-bit
    Selected as the low-power endpoint; the headline 3.31x saving is relative to {24 kHz, 12-bit}. No joint optimization over fs and bit resolution is shown.
  • Multi-period loss periods p={5,7} = 5,7
    Chosen to reduce model size versus HiFi-GAN's five periods; §5.5 ablates removal of the loss but not the period choice.
assumptions (5)
  • domain assumption ADC power is proportional to fs·2^N and dominates hearable energy use
    Used in §2.1 to motivate sub-Nyquist low-bit sampling; contradicted by Table 5 and not inclusive of radio, codec, DSP, or encryption power.
  • domain assumption The {4 kHz, 8-bit} sub-Nyquist signal contains enough information for a learned model to reconstruct 16-24 kHz clean speech
    Central premise of the BWE approach; the model learns this mapping from paired data, but no information-theoretic justification is given, and §5.3 shows degradation at 8-bit.
  • domain assumption Bone-conduction vibration is less noisy than air-conduction audio and can guide phase reconstruction
    Used in APEN §3.3; standard in multimodal SE, but depends on sensor placement and coupling in the authors' prototype.
  • standard math ITU G.114 150 ms one-way delay is the correct streaming bound
    Used in §5.1 to judge real-time capability; a reasonable and externally established threshold.
  • ad hoc to paper Flattening the 2D spectrogram into a sequence for Mamba preserves enough spatial structure for the bottleneck
    §3.1; an ad hoc adaptation of Mamba to 2D features. The §5.5 ablation against Transformers shows similar metrics, providing empirical but not theoretical support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CAPS: A Cascaded Reconstruction Model to Power Saving in Hearables Using Sub-Nyquist Sampling with Bandwidth Extension." pith.science (2026). https://pith.science/paper/VD4N2VXH

@misc{pith2026260719434,
  author       = {Pith},
  title        = {Pith review of: CAPS: A Cascaded Reconstruction Model to Power Saving in Hearables Using Sub-Nyquist Sampling with Bandwidth Extension},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VD4N2VXH}},
  note         = {Machine review of arXiv:2607.19434}
}
read the original abstract

Hearables are wearable computers worn on the ear. Bone conduction microphones are used with air conduction microphones in hearables for multimodal speech enhancement in noisy conditions. Despite this potential, current models largely fail to explore how jointly reducing sampling bit resolution and sampling frequency in analog-to-digital converters (ADCs) of hearables impacts both power usage and audio quality. Furthermore, current frameworks cannot do sub-Nyquist sampling in hearables because they lack a method to reconstruct wideband signals from narrowband components. We therefore propose CAPS, which (i) intentionally employs sub-Nyquist sampling and low bit resolution in ADCs, achieving a 3.3x reduction in power consumption in hearables, and (ii) supports streaming operation on mobile platforms with an inference time of 1.36 ms and a memory footprint of 11.04 MB. CAPS ensures robust speech intelligibility in real-world settings, bridging the gap between efficiency and power savings.

Figures

Figures reproduced from arXiv: 2607.19434 by the authors.

Figure 1
Figure 1. (Left) Overview of our appraoch. (Middle & Right) Our proposed design and different modules – UN & APEN – of CAPS. • CAPS will handle dynamic interferences from extremely noisy conditions in mobile scenarios using joint BWE and multimodal SE algorithms and reconstruct audio from low sampling frequency and bit resolutions. CAPS adopts a cascaded model ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. A prototype of hearable. As our low-resolution audio sampled from BCMs will go through joint BWE and multimodal SE in a mobile platform, we choose Google Pixel7 having the Google Tensor G2 chipset and Mali-G710 MP7 GPUs as the mobile platform. 4.2. Dataset Collection Although there are a few datasets [29, 3] available for multi￾modal SE, they don’t have simultaneous data from both vibra￾tion and accelerometer sensor… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 1 canonical work pages

  1. [1]

    Traditionally, air conduction microphones (ACMs) are used in hearables that are prone to background noise

    Introduction A hearable is a wearable computer that is worn on the ear. Traditionally, air conduction microphones (ACMs) are used in hearables that are prone to background noise. To solve this problem,bone conduction microphones (BCMs) are commonly used with ACMs as a conditional signal enhancer for multi- modal speech enhancement (SE) in noisy conditions...

  2. [2]

    Af- ter sampling, audio codecs compress data to reduce the bitrate, saving transmission energy and bandwidth

    Lowering sampling frequency and bit resolution:Au- dio or vibration signals from ACMs and BCMs, respectively, are first sampled at Nyquist rates (greater than 16 kHz) and over 12-bit resolutions by the analog-to-digital converter (ADC). Af- ter sampling, audio codecs compress data to reduce the bitrate, saving transmission energy and bandwidth. Later, mul...

  3. [3]

    Lacking in multimodal SE methods:State-of-the-art (SOTA) multimodal SE methods [4, 5, 6, 7, 1, 8, 2, 9, 10, 11, 12, 13, 14, 15, 16, 4, 7, 17] have either one or multiple of the following limitations for which they are not suitable for our low- power hearables:(i)The multimodal SE algorithms do not con- sider lower bit resolution and low power applications...

  4. [4]

    Preliminary 2.1. Power at Sub-Nyquist Frequencies and Bit Resolutions The power consumptionPof ADCs in hearables increases with sampling rates and resolutions, followingP=k·f s ·2 N , where kis a proportionality constant,Nis the bit resolution, andf s is the sampling frequency of ADCs. Traditional methods require ADCs to operate at high sampling frequenci...

  5. [5]

    Architecture Design CAPS is engineered to achieve the following objectives: • CAPS will enable streaming enhancement and low-power and low-memory solutions that will make CAPS deployable on mobile platforms. arXiv:2607.19434v1 [cs.SD] 21 Jul 2026 Spectral Enhanceme nt Network (U-Net) Upsampling Network Amplitude-Phase Enhancement Network (U-Net) Sub-Nyqui...

    arXiv 2026

  6. [6]

    Implementation 4.1. Wearable Platform and Mobile Platform Design We use a piezo-resistive vibration sensor (part# CEB-27032- L100) [27] and an accelerometer (part# 352C33) [28] as BCMs, placed on an off-the-shelf frame for collecting vibration near the earbone (see Fig. 2). We use the built-in ADC from the NRF52840 chip to sample the signals from the BCMs...

  7. [7]

    Performance Evaluation 5.1. Real-Time BWE and Multimodel SE with Inference Please note that VibV oice and SEANet are multimodal SE mod- els by default, whereas TFiLM, AERO, EBEN, and HiFi++ are single-modal SE models. To compare CAPS with the single- modal SE models, we convert our multimodal CAPS into a single-modal model by using a single-input network ...

  8. [8]

    Conclusion and Limitation The low latency of CAPS while performing joint BWE and mul- timodal SE enables streaming enhancement, low power, and low memory solutions, making CAPS deployable on mobile platforms. Therefore, CAPS is designed to meet the high de- mands of smart hearables in low-power and real-world usage by bridging the gap between power and th...

Show all 43 references
  1. [9]

    Generative AI Use Disclosure We acknowledge that we have used Elicit for finding relevant papers, and used ChatGPT for debugging codes, and finding grammatical errors

  2. [10]

    Tramba: A hybrid transformer and mamba architecture for practical audio and bone conduction speech super resolution and enhancement on mobile and wearable platforms,

    Y . Sui, M. Zhao, J. Xia, X. Jiang, and S. Xia, “Tramba: A hybrid transformer and mamba architecture for practical audio and bone conduction speech super resolution and enhancement on mobile and wearable platforms,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ub...

    2024

  3. [11]

    Seanet: A multi-modal speech enhancement network,

    M. Tagliasacchi, Y . Li, K. Misiunas, and D. Roblek, “Seanet: A multi-modal speech enhancement network,”arXiv preprint arXiv:2009.02095, 2020

    2009 arXiv

  4. [12]

    End-to-end multi-modal speech recognition on an air and bone conducted speech corpus,

    M. Wang, J. Chen, X.-L. Zhang, and S. Rahardja, “End-to-end multi-modal speech recognition on an air and bone conducted speech corpus,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 513–524, 2022

    2022

  5. [13]

    Clearspeech: Im- proving voice quality of earbuds using both in-ear and out-ear microphones,

    D. Ma, T. Dang, M. Ding, and R. K. Balan, “Clearspeech: Im- proving voice quality of earbuds using both in-ear and out-ear microphones,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 7, no. 4, pp. 170:1– 170:25, 2023

    2023

  6. [14]

    Earspeech: Exploring in-ear occlusion effect on earphones for data-efficient airborne speech enhancement,

    F. Han, P. Yang, Y . Zuo, F. Shang, F. Xu, and X.-Y . Li, “Earspeech: Exploring in-ear occlusion effect on earphones for data-efficient airborne speech enhancement,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 8, no. 3, pp. 104:1–10...

    2024 doi

  7. [15]

    Wearse: Enabling streaming speech enhancement on eyewear using acoustic sens- ing,

    Q. Zhang, K. Guo, Y . Yang, and D. Wang, “Wearse: Enabling streaming speech enhancement on eyewear using acoustic sens- ing,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 9, no. 1, pp. 1–30, 2025

    2025

  8. [16]

    Vibvoice: Towards bone-conducted vibration speech enhancement on head-mounted wearables,

    L. He, H. Hou, S. Shi, X. Shuai, and Z. Yan, “Vibvoice: Towards bone-conducted vibration speech enhancement on head-mounted wearables,” inProceedings of the 21st ACM International Con- ference on Mobile Systems, Applications and Services. ACM, 2023, pp. 356–369

    2023

  9. [17]

    Enabling real-time on-chip audio super resolution for bone-conduction mi- crophones,

    Y . Li, Y . Wang, X. Liu, Y . Shi, S. Patel, and S.-F. Shih, “Enabling real-time on-chip audio super resolution for bone-conduction mi- crophones,”Sensors, vol. 23, no. 1, p. 35, 2022

    2022

  10. [18]

    Aero: Audio super resolution in the spectral domain,

    M. Mandel, O. Tal, and Y . Adi, “Aero: Audio super resolution in the spectral domain,” inICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  11. [19]

    Eben: Extreme bandwidth extension network applied to speech signals captured with noise-resilient body-conduction microphones,

    J. Hauret, T. Joubaud, V . Zimpfer, and ´E. Bavu, “Eben: Extreme bandwidth extension network applied to speech signals captured with noise-resilient body-conduction microphones,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICA...

    2023

  12. [20]

    Neu- ral vocoder is all you need for speech super-resolution,

    H. Liu, W. Choi, X. Liu, Q. Kong, Q. Tian, and D. Wang, “Neu- ral vocoder is all you need for speech super-resolution,” inInter- speech, 2022

    2022

  13. [21]

    Denoising diffusion probabilis- tic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilis- tic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  14. [22]

    Nu-wave: A diffusion probabilistic model for neural audio upsampling,

    J. Lee and S. Han, “Nu-wave: A diffusion probabilistic model for neural audio upsampling,” inInterspeech 2021, 2021, pp. 1634– 1638

    2021

  15. [23]

    Nu-wave 2: A general neural audio up- sampling model for various sampling rates,

    S. Han and J. Lee, “Nu-wave 2: A general neural audio up- sampling model for various sampling rates,” inInterspeech 2022, 2022, pp. 4401–4405

    2022

  16. [24]

    Temporal film: Capturing long-range sequence dependen- cies with feature-wise modulations

    S. Birnbaum, V . Kuleshov, Z. Enam, P. W. W. Koh, and S. Er- mon, “Temporal film: Capturing long-range sequence dependen- cies with feature-wise modulations.”Advances in Neural Infor- mation Processing Systems, vol. 32, 2019

    2019

  17. [25]

    Self-attention for audio super-resolution,

    N. C. Rakotonirina, “Self-attention for audio super-resolution,” in 2021 IEEE 31st International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 2021, pp. 1–6

    2021

  18. [26]

    HiFi++: Towards perceptually enhanced and computationally ef- ficient neural vocoders,

    W. Kim, H. Kang, Y . Kim, K. Jung, J. S. Lee, and S.-H. Lee, “HiFi++: Towards perceptually enhanced and computationally ef- ficient neural vocoders,” inProc. Interspeech, 2023, pp. 4374– 4378

    2023

  19. [27]

    Recommen- dation G.114: One-way transmission time,

    International Telecommunication Union, “Recommen- dation G.114: One-way transmission time,” https: //www.itu.int/rec/T-REC-G.114, 2003, accessed: 2025-04- 10

    2003

  20. [28]

    Type 4192: 1/2

    Br ¨uel & Kjær Sound & Vibration Measurement A/S, “Type 4192: 1/2” Pressure-field Microphone – High Sensitiv- ity,” https://www.bksv.com/en/transducers/microphones/ microphone-cartridges/4192, 2002, accessed: 2025-04-10

    2002

  21. [29]

    nrf52840 product specification v1.1,

    N. Semiconductor, “nrf52840 product specification v1.1,” https:// infocenter.nordicsemi.com/pdf/nRF52840 PS v1.1.pdf, 2018, ac- cessed: 2025-04-26

    2018

  22. [30]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    2023 arXiv

  23. [31]

    Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,”Advances in neural information processing systems, vol. 33, pp. 17 022– 17 033, 2020

    2020

  24. [32]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016

    2016 arXiv

  25. [33]

    Gaussian error linear units (gelus),

    D. Hendrycks and K. Gimpel, “Gaussian error linear units (gelus),”arXiv preprint arXiv:1606.08415, 2016

    2016 arXiv

  26. [34]

    Neural speech phase prediction based on parallel estimation architecture and anti-wrapping losses,

    Y . Ai and Z.-H. Ling, “Neural speech phase prediction based on parallel estimation architecture and anti-wrapping losses,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  27. [35]

    Mel- gan: Generative adversarial networks for conditional waveform synthesis,

    K. Kumar, R. Kumar, T. De Boissiere, L. Gestin, W. Z. Teoh, J. Sotelo, A. De Brebisson, Y . Bengio, and A. C. Courville, “Mel- gan: Generative adversarial networks for conditional waveform synthesis,”Advances in neural information processing systems, vol. 32, 2019

    2019

  28. [36]

    CEB-27032-L100: Contact Micro- phone,

    CUI Devices, “CEB-27032-L100: Contact Micro- phone,” https://www.cuidevices.com/product/audio/speakers/ contact-microphones/ceb-27032-l100, 2023, accessed: 2025-04- 10

    2023

  29. [37]

    352C33: ICP® Quartz Shear Ac- celerometer,

    PCB Piezotronics, Inc., “352C33: ICP® Quartz Shear Ac- celerometer,” https://www.pcb.com/products?m=352C33, 2024, accessed: 2025-04-10

    2024

  30. [38]

    Vibravox: A dataset of french speech captured with body-conduction audio sensors,

    J. Hauret, M. Olivier, T. Joubaud, C. Langrenne, S. Poir ´ee, V . Zimpfer, and ´E. Bavu, “Vibravox: A dataset of french speech captured with body-conduction audio sensors,”arXiv preprint arXiv:2407.11828, 2024

    2024 arXiv

  31. [39]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (ver- sion 0.92),

    J. Yamagishi, C. Veaux, K. MacDonaldet al., “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (ver- sion 0.92),”University of Edinburgh. The Centre for Speech Tech- nology Research (CSTR), pp. 271–350, 2019

    2019

  32. [40]

    Whisper: Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Whisper: Robust speech recognition via large-scale weak supervision,” https://github.com/openai/whisper, 2022, ope- nAI Technical Report

    2022

  33. [41]

    Freesound technical demo,

    F. Font, G. Roma, and X. Serra, “Freesound technical demo,” in Proceedings of the 21st ACM international conference on Multi- media, 2013, pp. 411–412

    2013

  34. [42]

    Onnx-tensorflow: Open neural network ex- change (onnx) backend for tensorflow,

    O.-T. Community, “Onnx-tensorflow: Open neural network ex- change (onnx) backend for tensorflow,” https://github.com/onnx/ onnx-tensorflow, 2024, gitHub repository

    2024

  35. [43]

    Evalu- ation of algorithms using games: The case of music tagging,

    E. Law, M. West, M. Mandel, M. Bay, and J. S. Downie, “Evalu- ation of algorithms using games: The case of music tagging,” in Proceedings of the 10th International Society for Music Informa- tion Retrieval Conference (ISMIR), 2009

    2009

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.