Efficient and Distortion-less Spectrum Multiplexer via Neural Network-based Filter Banks

Jiazhao Wang; Wenchao Jiang

REVIEW 4 major objections 6 minor 33 references

Efficient and Distortion-less Spectrum Multiplexer via Neural Network-based Filter Banks

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a neural network constrained to be an oversampled polyphase filter bank can multiplex multiple narrowband IoT streams into a single wideband signal with near-ideal fidelity and up to a 35x speedup over direct methods.

desk verdict A useful, honest engineering paper: learned polyphase filter banks for spectrum multiplexing work as advertised, but the evaluation misses the most relevant baseline and papers over a couple of informal steps. read the letter →

arxiv 2507.17106 v1 pith:OECUA5YH submitted 2025-07-23 eess.SP

Jiazhao Wang , Wenchao Jiang This is my paper · ORCID

classification eess.SP

keywords spectrummultiplexingpolyphasefilterbankneuralnetworksoftware-definedradioInternetofThingsoversamplingmodel-drivenlearningsignalinterpolation

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 spectrum multiplexer—the baseband module that combines several narrowband IoT streams into one wideband signal—can be built from a neural network whose architecture is exactly an oversampled polyphase filter bank. The claim is that gradient descent can replace manual filter design, and that the same network, run on a GPU, removes the efficiency penalty of filtering. In evaluation, the learned filter bank reaches -39 dB normalized mean squared error, BER close to the ideal reference with up to 10 dB SNR gain over a DFT-based multiplexer, up to 35x faster execution than direct interpolation, and packet reception ratios near 98% in field tests. If true, filter-bank design for this task stops being hand-tuned and becomes a trainable, hardware-accelerated software module.

What carries the argument

The central object is the NNPFB, a neural network-based polyphase filter bank in which the analysis arm is a multi-channel transposed convolution followed by a DFT and the synthesis arm is an IDFT followed by a strided convolution; the only trainable parameters are the polyphase filter kernels, initialized with windowed sinc filters. The load-bearing identity is the oversampling condition $K = M I$: the number of sub-bands $K$ is the product of the decimation factor $M$ and an oversampling ratio $I > 1$, so sub-band spectra do not alias after downsampling. The stateless polyphase decomposition removes the need for stateful index alternation, so standard convolution layers with fixed stride implement the interpolation and decimation. This mapping is what lets a generic optimizer tune the filters and lets GPU tensor libraries execute them.

What would settle it

Compute the residual between the high-rate training target downsampled by the interpolation factor and the low-rate training input; a nonzero residual means the MSE objective is biased. Then feed a unit impulse through the trained NNPFB and through an exact polyphase decimator with the same coefficients; any sample offset or spurious response identifies whether the strided convolution reproduces true decimation.

Watch

Extended reading notes

Core claim

The central claim is that an interpretable neural network with the topology of an oversampled polyphase filter bank can learn the analysis and synthesis filter taps that make multiplexing nearly distortionless, while the same network serves as a practical implementation because deep-learning libraries accelerate it. The paper builds the NNPFB by mapping interpolation to transposed convolution, decimation to strided convolution, and DFT/IDFT to fixed linear layers; only the polyphase filter kernels are trainable, initialized from windowed sinc filters. Trained on pairs of low-rate and high-rate modulated waveforms, the network achieves -39.49 dB NMSE for 4x QPSK interpolation and BER curves matching the standard reference, beating the DFT-based method by up to 10 dB SNR and running up to 35x faster than the direct approach on a GPU. The paper also shows that the trained kernels recover sensible synthesis filters and that field packet reception ratios match single-stream baselines for both homogeneous and heterogeneous IoT scenarios.

Load-bearing premise

The method assumes that the low-rate training waveform is exactly the decimated version of the high-rate target waveform, and that a neural strided convolution reproduces true decimation; if either fails, the filters are trained against a biased reference and the reported error numbers would not measure true distortion.

Editorial extensions

If this is right

A spectrum multiplexer can be implemented as a GPU-accelerated neural module with only about 127 trainable parameters, replacing hand-designed filters.
The oversampled polyphase structure reaches an NMSE around -39 dB, so multiplexed waveforms can approach the ideal reference rather than the spectrally leaked DFT-based output.
Because BER curves track the ideal reference with up to 10 dB SNR gain over DFT-based multiplexing, standard receivers can decode multiplexed streams without modification.
Field PRRs for three simultaneous ZigBee streams and a Wi-Fi/ZigBee mix stay within a few points of single-stream baselines, so heterogeneous IoT traffic can share one front end.
The same architecture can be reversed to split a wideband signal into narrowband streams, giving a trainable receiver-side channelizer.

Reading between the lines

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

The paper leaves unverified that the low-rate training input is exactly the decimated version of the high-rate target, and if that consistency fails, the reported NMSE values are measured against a biased reference; a reader should test this by comparing $x'[L n]$ with $x[n]$ on training pairs.
The admitted indexing difference between Equation 11 and true decimation could create a fractional-sample delay that the learned kernels absorb; this would be invisible in BER and NMSE but might matter for time-sensitive protocols, and it could be exposed by feeding a unit impulse through the network.
The same model-driven mapping should extend to non-integer sample-rate conversions by adjusting the oversampling ratio, so a natural extension is training one NNPFB for fractional rates and comparing against optimal polyphase resamplers.

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.

Desk Editor's Note

read the letter

First: this is a legitimate new result and the central claim holds up. The authors take the textbook oversampled polyphase filter bank, re-express it as a constrained neural network (transposed conv as interpolation, conv as decimation, fixed DFT/IDFT layers), and train only the polyphase filter taps against modulated waveforms. The held-out NMSE (–33 to –39 dB), the Wi-Fi BER curves tracking the ideal reference, and the field PRR results (81–98%) are internally consistent. The 802.11n test is genuinely out-of-distribution, which answers the circularity worry. And with only 127 trainable parameters, the model-driven initialization is doing real work.

What's new: applying the NN-as-DSP program to spectrum multiplexing, with a training scheme that removes manual filter design. That's a modest but real step beyond the authors' prior work. The write-up is clear, and the interpretation of the trained filters in Figure 15 is a nice touch.

The soft spots are in the evaluation, not the math. The biggest gap is the missing baseline: they never compare against a conventionally designed oversampled PFB, which is the direct competitor for the 'we eliminate manual design' claim. The DFT-based method is a weak strawman given its known spectral leakage. Without a PFB baseline, the claim that NNPFB matches an ideal PFB is supported only by low NMSE, not by comparison. Second, the 35x efficiency headline compares NNPFB-on-GPU against baseline-on-CPU; on CPU, NNPFB is slower than DFT-based, which the authors admit. Third, there are no error bars or run counts anywhere, so the small NMSE differences between methods are hard to interpret.

On the stress-test concern about §4.3 training-target consistency: I don't think it's fatal. The paper is sloppy in not stating that x[n] = x'[L n], but achieving –39 dB NMSE on held-out signals is hard to explain unless the targets are almost exactly the L-fold oversampled inputs. If the two modulators produced inconsistent waveforms, there would be an irreducible MSE floor well above –30 dB. So this is a documentation gap, not a demonstrated flaw. Still, it should be fixed and verified explicitly.

This paper deserves a serious referee. The core engineering result is real, the field experiments are a cut above typical simulation-only work, and the missing baseline and stats are fixable. I'd send it out, though I wouldn't hold it up as a model of rigor.

Referee Report

4 major / 6 minor

Summary. The paper proposes a neural-network-based realization of oversampled polyphase filter banks, called NNPFB, for baseband spectrum multiplexing. The analysis side is an NNPAFB that decomposes low-rate narrowband streams into oversampled sub-bands, and the synthesis side is an NNPSFB that recombines them into a wideband signal; the only trainable kernels are the transposed-convolution and convolution weights corresponding to the polyphase filters, with DFT/IDFT and complex-to-real mapping blocks fixed. Training is performed with an MSE loss against high-rate waveforms generated by modulators run at different sampling rates. The authors report NMSE as low as -39 dB, BER close to the ideal reference for a simulated 802.11n stream, up to 35x execution-time improvement on GPU, and field PRRs close to single-stream baselines for ZigBee and combined Wi-Fi/ZigBee transmissions. The central claim is that gradient-based training can replace manual filter-bank design while retaining the fidelity of an ideal oversampled PFB and benefiting from GPU acceleration.

Significance. If the implementation-consistency issues are resolved, the paper makes a useful methodological contribution: it gives a clean, interpretable mapping from oversampled polyphase filter banks to standard neural-network layers, with only 127 trainable parameters, and it provides both simulation and field evidence. The held-out 802.11n experiment in Section 6.3 is genuinely external to the QPSK/ZigBee/BLE training set, and the BER curves and field PRR results are mutually consistent with the low-distortion claim. The paper also ships a code link and reports hardware-acceleration results on two platforms. The main significance is practical and architectural rather than theoretical; the load-bearing points that need attention are the consistency of the training target with true interpolation, the equivalence of the convolutional decimation layer to polyphase decimation, the absence of a conventional oversampled PFB baseline, and the fairness of the GPU-vs-CPU efficiency comparison.

major comments (4)

[Section 4.3, 'Training waveform generation'; Table 2] The paper does not establish that the high-rate target x'[l] is the exact L-fold oversampled version of the low-rate input x[n]. The text says only that the two waveform generators use modulators 'with the same characteristics' at different sampling rates and that the resulting signals 'exhibit similar waveforms.' If the two implementations discretize the pulse shaper differently, the target is not in the range of the filter-bank cascade and the MSE loss contains an irreducible bias; the NMSE values in Table 2 would then not measure interpolation fidelity against the true high-rate signal. Please state and verify the consistency condition x[n] = x'[L n] for a bandlimited waveform, or generate the target by resampling a single continuous-time waveform at the two rates.
[Section 4.1, Eq. (11); Section 3.3, Eq. (9)] The paper acknowledges that the convolutional-layer model 'differs in the indexing' from true decimation. Since the Conv #1 layer in the NNPSFB is the trainable decimation stage, the indexing and padding conventions determine which input samples contribute to each output sample. Please provide a derivation or an explicit implementation-level verification that Eq. (11), under the PyTorch padding/stride/grouping conventions actually used, is equivalent to the polyphase decimation in Eq. (9). If it is not equivalent, the claim that the architecture faithfully realizes the stateless polyphase decomposition of Section 3.3 does not carry over to the neural implementation.
[Section 6, Tables 2 and 4, Figure 16] The evaluation never compares the NNPFB against a conventionally designed oversampled PFB, although the abstract and contributions claim the performance of a theoretically ideal oversampled PFB. Table 2 compares only with U-Net- and ResNet-based models, and Figure 16 with direct and DFT-based methods. Add a baseline implemented with a standard oversampled polyphase filter bank (for example, using the same Kaiser-windowed sinc used for initialization) to the NMSE and BER comparisons, or soften the claim to 'low distortion' rather than 'theoretically ideal PFB.'
[Section 6.4, Figure 17] The reported 35x speedup compares NNPFB on GPU with direct and DFT methods running on CPU, while the text states that unaccelerated NNPFB is only slightly longer than the DFT-based method. The efficiency claim should be framed as a system-level GPU-acceleration result rather than an intrinsic algorithmic gain, and ideally the direct and DFT baselines should also be measured with GPU acceleration, or the platform asymmetry should be stated prominently in the abstract and conclusion where the 35x figure appears.

minor comments (6)

[Section 2, Eq. (1)] The anti-imaging filter is denoted g in Eq. (1) but f in the surrounding text; align the notation.
[Section 3.3, Eq. (7)] The exponential W^{kρ}_k appears to have a typographical error: the subscript should be K, not the summation index k; otherwise the expression is undefined outside the sum.
[Section 4.3, 'Initialization'; Table 1] The paper reports 127 trainable parameters but does not list the kernel lengths, polyphase branch counts, or the exact filter lengths used in the experiments; include these configuration details for reproducibility.
[Section 6.4, text following Figure 17] The phrase 'increases nearly exponentially' is inaccurate for message lengths 16, 32, 64, and 127 bytes; the plotted scaling appears closer to linear or sub-linear, and the wording should be revised.
[Section 6.5, Tables 4 and 5] The field PRR results are reported without confidence intervals or the number of packets transmitted; adding these would strengthen the claim that the NNPFB multiplexer performs comparably to single-stream baselines.
[Footnote 1 (code availability)] The GitHub repository name 'Repo4Sub/Sensys2026' appears to be a submission placeholder; the final version should point to a permanent, accessible repository or DOI.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; central claims are supported by held-out tests and external benchmarks.

full rationale

The paper derives the NNPFB architecture directly from DSP equations (Sections 2-4) rather than from the authors' prior work. The polyphase filter bank structure follows Crochiere-Rabiner [10] and Harris et al. [15], which are external references. The training procedure uses a model-driven sinc initialization and an MSE loss against independently generated high-rate waveforms; the validation set is held out, and Section 6.3 tests a new signal type (802.11n Wi-Fi) absent from the QPSK/ZigBee/BLE training data. BER curves and field PRR are external benchmarks. The only self-citations ([31,32]) appear in a related-work paragraph and are not used to justify any load-bearing premise. The reader's concern about the consistency of the §4.3 training target (whether x'[l] is exactly the interpolated x[n]) is a methodological validity issue, not circularity, because the model is not defined in terms of the target and no parameter is fitted to the evaluation set. The NMSE metric coincides with the training loss, but reporting it on held-out data is standard supervised evaluation rather than a reduction of the prediction to the fit.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on classic filter bank theory, which the paper cites, plus two engineering premises it asserts rather than proves: the conv-layer and decimation equivalence under framework-specific indexing (Section 4.1, Equation 11), and the consistency of the training target (Section 4.3). The hand-chosen configuration (K = 8/32, I = 2, pi/8 prototypes) and the 127 learned filter taps are the free quantities of the method; the 'distortionless' headline is an empirical property of fitting and evaluating those taps, not a theorem. No new physical entities are introduced.

free parameters (4)

Trainable polyphase filter kernels (NNPAFB TransConv #1 and NNPSFB Conv #1) = 127 total trainable parameters, with the trained synthesis filter response shown in Fig 15.
The filter taps are the design output, learned by gradient descent to minimize MSE on the training waveforms (Sections 4.3, 6.1, Table 1). They are numbers fitted to training data; legitimate for filter design, but it means the -29.9 to -39.5 dB NMSE figures are the training objective evaluated on held-out draws, not independent measurements.
Number of sub-bands K = 8 for analysis and 32 for synthesis in the ZigBee multiplexer (Sections 6.4 and 6.5).
Chosen by hand from the 0.5 MHz sub-band interval and 16 MHz output band. The achievable NMSE and the computational cost both depend on this configuration, and no sweep over K is reported.
Oversampling ratio I = 2, the oversampling ratio used throughout Section 3.2 and the experiments.
Set to I = 2 in Section 3.2 (K = M*I). The aliasing behavior and the convergence of training depend on this choice; the paper reports no variation of I.
Prototype filter bandwidth and window = Normalized bandwidth pi/8, Kaiser window for analysis, truncated sinc for synthesis (Sections 4.3 and 6.2).
Initialization choices described in Sections 4.3 and 6.2. Figure 15 shows sinc initialization beats random initialization, so the final performance inherits from these hand-picked prototypes.

assumptions (5)

standard math Standard DFT/IDFT, interpolation, and decimation identities (Equations 1-3), and the polyphase change of variables n = rK + rho in Equations 6-9.
Used throughout Section 3; the decomposition is textbook and treated as unproved background.
domain assumption Oversampling with K = M*I removes the aliasing that practical non-ideal filters cause at critical sampling.
Section 3.2 invokes oversampled filter bank theory from [10, 15] to justify the structure; the paper does not derive reconstruction bounds for its specific filters.
domain assumption A strided convolutional layer implements decimation and a strided transposed convolutional layer implements interpolation, with framework-standard padding and strides.
Section 4.1, Observations 1-2 and Equations 10-11. The paper itself notes Equation 11 'differs in the indexing' from Equation 2; the exact equivalence when deployed in PyTorch is asserted, not proven.
domain assumption The complex-valued filter bank can be built from real-valued filter taps plus fixed DFT/IDFT and fixed complex-combination layers.
Section 4.2 'Model Simplification' drops all complex taps, relying on the prototype filters being real and symmetric; this constrains the achievable responses without an analysis of the loss.
ad hoc to paper Each training pair (x[n], x'[l]) consists of two consistent samples of the same continuous waveform, so x[n] = x'[L n].
Section 4.3 generates the target by modulating the same symbols at a different sample rate but never states the consistency condition. If the two modulator outputs are not exactly aligned and bandlimited, the MSE objective is biased and all NMSE claims are measured against a wrong reference.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient and Distortion-less Spectrum Multiplexer via Neural Network-based Filter Banks." pith.science (2026). https://pith.science/paper/OECUA5YH

@misc{pith2026250717106,
  author       = {Pith},
  title        = {Pith review of: Efficient and Distortion-less Spectrum Multiplexer via Neural Network-based Filter Banks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OECUA5YH}},
  note         = {Machine review of arXiv:2507.17106}
}

abstract

Spectrum multiplexer enables simultaneous transmission of multiple narrow-band IoT signals through gateway devices, thereby enhancing overall spectrum utilization. We propose a novel solution based on filter banks that offer increased efficiency and minimal distortion compared with conventional methods. We follow a model-driven approach to integrate the neural networks into the filter bank design by interpreting the neural network models as filter banks. The proposed NN-based filter banks can leverage advanced learning capabilities to achieve distortionless multiplexing and harness hardware acceleration for high efficiency. Then, we evaluate the performance of the spectrum multiplexer implemented by NN-based filter banks for various types of signals and environmental conditions. The results show that it can achieve a low distortion level down to $-39$dB normalized mean squared error. Furthermore, it achieves up to $35$ times execution efficiency gain and $10$dB SNR gain compared with the conventional methods. The field applications show that it can handle both the heterogeneous and homogeneous IoT networks, resulting in high packet reception ratio at the standard receivers up to $98\%$.

Figures

Figures reproduced from arXiv: 2507.17106 by the authors.

**Figure 2.** (a) Direct spectrum multiplexer based on time-domain interpolation and modulation. (b) DFT-based spectrum [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Diagram of interpolators and corresponding spec [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (13 more)

**Figure 4.** Figure 4: Diagram of decimators and corresponding spectrum [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 6.** Figure 6: The spectrum of sub-band signal and downsam [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

**Figure 8.** Figure 8: Diagram of the basic operation of (a) the transposed [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 7.** Figure 7: Diagrams of (a) the analysis filter bank with [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

**Figure 10.** Figure 10: Diagram of the NN-based DFT/IDFT block. FC Layer (Combination) TransConv Layer (Single-channel) NN-based Interpolator (a) FC Layer (Spread) Conv Layer (Single-channel) NN-based Decimator (b) [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]

**Figure 9.** Figure 9: (a) Diagram of the operation of a multi-channel [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]

**Figure 12.** Figure 12: Diagram of (a) the NN-based polyphase analysis filter bank and (b) the NN-based polyphase synthesis filter bank. [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]

**Figure 13.** Figure 13: The training procedure for NN-based filter banks. [PITH_FULL_IMAGE:figures/full_fig_p007_13.png]

**Figure 13.** Figure 13: A straightforward approach is to apply modulators to [PITH_FULL_IMAGE:figures/full_fig_p008_13.png]

**Figure 14.** Figure 14: (a) Training loss of the plain U-Net-based, ResNet [PITH_FULL_IMAGE:figures/full_fig_p009_14.png]

**Figure 16.** Figure 16: (a) BER curves of signals generated by different [PITH_FULL_IMAGE:figures/full_fig_p010_16.png]

**Figure 17.** Figure 17: (a) Running time on desktop PC. (b) Running time [PITH_FULL_IMAGE:figures/full_fig_p011_17.png]

**Figure 18.** Figure 18: Diagram of Wi-Fi and ZigBee signal combination [PITH_FULL_IMAGE:figures/full_fig_p011_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages

[1]

2018. Wi-Fi 6. https://www.wi-fi.org/news-events/newsroom/wi-fi-alliance- introduces-wi-fi-6

work page 2018
[2]

2024. 5G-NR. https://www.qualcomm.com/research/5g/5g-nr

work page 2024
[3]

MATLAB Signal Processing Toolbox

2024. MATLAB Signal Processing Toolbox. https://www.mathworks.com/ products/signal.html

work page 2024
[4]

MATLAB WLAN Toolbox

2024. MATLAB WLAN Toolbox. https://www.mathworks.com/products/wlan. html

work page 2024
[5]

Nvidia Jetson Orin Kit

2024. Nvidia Jetson Orin Kit. https://www.nvidia.com/en-sg/autonomous- machines/embedded-systems/jetson-orin/

work page 2024
[6]

2024. PyTorch. https://pytorch.org/

work page 2024
[7]

2024. SciPy. https://www.scipy.org/

work page 2024
[8]

Texas Instruments LAUNCHXL-CC2650

2024. Texas Instruments LAUNCHXL-CC2650. https://www.ti.com/tool/ LAUNCHXL-CC2650

work page 2024

Show all 33 references

[9]

USRP X310 SDR

2024. USRP X310 SDR. https://www.ettus.com/all-products/x310-kit/

2024
[10]

Crochiere and Lawrence R

Ronald E. Crochiere and Lawrence R. Rabiner. [n. d.]. Multirate Digital Signal Processing. Prentice-Hall
[11]

Felipe AP de Figueiredo, Ruben Mennes, Irfan Jabandžić, Xianjun Jiao, and Ingrid Moerman. 2020. A baseband wireless spectrum hypervisor for multiplexing concurrent ofdm signals. Sensors 20, 4 (2020), 1101

2020
[12]

Jesse Engel, Lamtharn Hantrakul, Chenjie Gu, and Adam Roberts. 2020. DDSP: Differentiable digital signal processing. arXiv preprint arXiv:2001.04643 (2020)

2020 arXiv
[13]

Andrea Goldsmith. 2005. Wireless communications. Cambridge university press

2005
[14]

Fred Harris, Chris Dick, Xiaofei Chen, and Elettra Venosa. 2011. Wideband 160-channel polyphase filter bank cable TV channeliser. IET Signal processing 5, 3 (2011), 325–332

2011
[15]

Harris, C

F.J. Harris, C. Dick, and M. Rice. [n. d.]. Digital Receivers and Transmitters Using Polyphase Filter Banks for Wireless Communications. 51, 4 ([n. d.]), 1395–1412. https://doi.org/10.1109/TMTT.2003.809176

2003
[16]

Fred Harris, Elettra Venosa, Xiaofei Chen, and Chris Dick. 2017. Cascade non- maximally decimated filter banks form efficient variable bandwidth filters for wideband digital transceivers. In 2017 22nd International Conference on Digital Signal Processing (DSP). IEEE, 1–5

2017
[17]

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

2016
[18]

Jakob Hoydis, Sebastian Cammerer, Fayçal Ait Aoudia, Avinash Vem, Nikolaus Binder, Guillermo Marcus, and Alexander Keller. 2022. Sionna: An open-source li- brary for next-generation physical layer research. arXiv preprint arXiv:2203.11854 (2022)

2022 arXiv
[19]

Maicon Kist, Juergen Rochol, Luiz A DaSilva, and Cristiano Bonato Both. 2018. SDR virtualization in future mobile networks: Enabling multi-programmable air-interfaces. In 2018 IEEE International Conference on Communications (ICC) . IEEE, 1–6

2018
[20]

Maicon Kist, João F Santos, Diarmuid Collins, Juergen Rochol, Luiz A DaSilva, and Cristiano Bonato Both. 2020. AIRTIME: End-to-end virtualization layer for RAN-as-a-Service in future multi-service mobile networks. IEEE Transactions on Mobile Computing 21, 8 (2020), 2701–2717

2020
[21]

Volodymyr Kuleshov, S Zayd Enam, and Stefano Ermon. 2017. Audio super resolution using neural networks. arXiv preprint arXiv:1708.00853 (2017)

2017 arXiv
[22]

Wei Liu, Joao F Santos, Jonathan van de Belt, Xianjun Jiao, Ingrid Moerman, Johann Marquez-Barja, Luiz DaSilva, and Sofie Pollin. 2020. Enabling virtual radio functions on software defined radio for future wireless networks. Wireless Personal Communications 113 (2020), 1579–1595

2020
[23]

Price, Maciej J

Nathan D. Price, Maciej J. Zawodniok, and Ivan G. Guardiola. [n. d.]. Transceivers as a Resource: Scheduling Time and Bandwidth in Software-Defined Radio. 8 ([n. d.]), 132603–132613. https://doi.org/10.1109/ACCESS.2020.3011051

2020
[24]

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceeding...

2015
[25]

Joachim Sachs and Stephan Baucke. [n. d.]. Virtual Radio: A Framework for Con- figurable Radio Networks. In Proceedings of the 4th International ICST Conference on Wireless Internet (2008). ICST. https://doi.org/10.4108/ICST.WICON2008.4925

2008
[26]

Santos, Maicon Kist, Juergen Rochol, and Luiz A

Joao F. Santos, Maicon Kist, Juergen Rochol, and Luiz A. DaSilva. [n. d.]. Virtual Radios, Real Services: Enabling RANaaS Through Radio Virtualisation. 17, 4 ([n. d.]), 2610–2619. https://doi.org/10.1109/TNSM.2020.3009863

2020
[27]

Jennifer Pearl Smith, J. I. Bailey, John Tuthill, Leandro Stefanazzi, Gustavo Can- celo, Ken Treptow, and Benjamin A. Mazin. [n. d.]. A High-Throughput Oversam- pled Polyphase Filter Bank Using Vivado HLS and PYNQ on a RFSoC. 2 ([n. d.]), 241–252. https://doi.org/10.1109/OJCAS...
[28]

Manolis Surligas, Antonis Makrogiannakis, and Stefanos Papadakis. 2015. Em- powering the IoT heterogeneous wireless networking with software defined radio. In 2015 IEEE 81st Vehicular Technology Conference (VTC Spring) . IEEE, 1–5

2015
[29]

Kun Tan, Haichen Shen, Jiansong Zhang, and Yongguang Zhang. [n. d.]. Enable Flexible Spectrum Access with Spectrum Virtualization. In2012 IEEE International Symposium on Dynamic Spectrum Access Networks (2012-10). 47–58. https: //doi.org/10.1109/DYSPAN.2012.6478115

2012
[30]

Vaidyanathan

P.P. Vaidyanathan. Jan./1990. Multirate Digital Filters, Filter Banks, Polyphase Networks, and Applications: A Tutorial. 78, 1 (Jan./1990), 56–93. https://doi.org/ 10.1109/5.52200

1990 doi
[31]

Haoyu Wang, Jiazhao Wang, Demin Gao, and Wenchao Jiang. 2024. NNCTC: Physical Layer Cross-Technology Communication via Neural Networks. arXiv preprint arXiv:2403.10014 (2024)

2024 arXiv
[32]

2024.{NN-Defined} Modulator: Reconfigurable and Portable Software Modulator on{IoT} Gateways

Jiazhao Wang, Wenchao Jiang, Ruofeng Liu, Bin Hu, Demin Gao, and Shuai Wang. 2024.{NN-Defined} Modulator: Reconfigurable and Portable Software Modulator on{IoT} Gateways. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24) . 775–789

2024
[33]

Chaorui Zhang, Peng Xie, Deyuan Li, Jiekai Zhang, and Rong Yu. 2011. Wireless home gateway: Software-defined radio architecture and applications. In IET International Conference on Communication Technology and Application (ICCTA 2011). IET. 13

2011

Pith tools

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

[1] [1]

2018. Wi-Fi 6. https://www.wi-fi.org/news-events/newsroom/wi-fi-alliance- introduces-wi-fi-6

work page 2018

[2] [2]

2024. 5G-NR. https://www.qualcomm.com/research/5g/5g-nr

work page 2024

[3] [3]

MATLAB Signal Processing Toolbox

2024. MATLAB Signal Processing Toolbox. https://www.mathworks.com/ products/signal.html

work page 2024

[4] [4]

MATLAB WLAN Toolbox

2024. MATLAB WLAN Toolbox. https://www.mathworks.com/products/wlan. html

work page 2024

[5] [5]

Nvidia Jetson Orin Kit

2024. Nvidia Jetson Orin Kit. https://www.nvidia.com/en-sg/autonomous- machines/embedded-systems/jetson-orin/

work page 2024

[6] [6]

2024. PyTorch. https://pytorch.org/

work page 2024

[7] [7]

2024. SciPy. https://www.scipy.org/

work page 2024

[8] [8]

Texas Instruments LAUNCHXL-CC2650

2024. Texas Instruments LAUNCHXL-CC2650. https://www.ti.com/tool/ LAUNCHXL-CC2650

work page 2024