Neuro-Wideband: Neural & Neuromorphic Signal Processing

• Neuro-Wideband (NWB) is a framework that unites neural network models, neuromorphic principles, and wideband signal processing to exceed traditional bandwidth constraints.
• It leverages nonlinear prediction, deep generative models, and self-supervised methods for applications such as enhanced audio coding, wireless sensing, and analog signal regeneration.
• NWB systems demonstrate improved performance metrics, energy efficiency, and innovative architectures in domains ranging from speech coding to event-driven semantic communications.

Neuro-Wideband (NWB) encompasses a set of paradigms and methods at the intersection of neural network models, neuromorphic principles, and wideband signal processing. NWB methodologies aim to extend, enhance, or emulate the bandwidth and time-frequency resolution of biological and engineered systems well beyond conventional digital or analog designs. These approaches span neural-network-based audio coding, neuro-inspired analog front-ends, neural self-extrapolation in sensing, and event-driven wireless communications, unified by their exploitation of neural function or learning to surpass traditional bandwidth or hardware constraints.

1. Neural Bandwidth Extension and Coding Architectures

NWB frameworks have transformed wideband speech and audio coding through nonlinear neural prediction and deep generative models. In classic sub-band speech coding, a wideband input (e.g., 16 kHz speech) is decomposed via a quadrature mirror filter (QMF) bank into sub-bands, each encoded separately. NWB replaces linear adaptive prediction with multi-layer perceptrons (MLPs) acting as nonlinear predictors for each sub-band. The canonical predictor structure is a 10–2–1 MLP (10-tap input, two hidden neurons with $\tanh$ activation, one linear output), trained per-frame by minimizing mean-squared error using the Levenberg–Marquardt algorithm. This nonlinear prediction achieves segmental SNR gains of up to 2 dB over classical linear schemes, particularly at higher bit rates (e.g., 32.72 dB SEGSNR for MLP at 80 kbps, compared to 30.91 dB for LPC-10) (Faundez-Zanuy, 2022).

To facilitate bandwidth extension from narrowband transmissions, NWB introduces synthetic wideband regeneration using a memoryless quadratic nonlinearity: the decoded low band is squared and high-pass filtered to synthesize harmonics in the missing band.

In modern audio codecs, NWB is realized by inserting two deep neural submodules around a fixed core codec. The SBG-encoder extracts side-information from the high-frequency content, which is compressed via Residual Vector Quantization (RVQ), while the SBG-decoder leverages both the decoded low band and the side-information to generate the missing spectral bands. The system is trained end-to-end with adversarial, mel-reconstruction, and feature-matching losses, with architectures drawing from ResNet and SEANet. Conditioning is provided by Temporal Feature-wise Linear Modulation (TFiLM) (Choi et al., 7 Jun 2025).

2. Neuro-Wideband Sensing and Self-Extrapolation

NWB also refers to a sensing paradigm in which a neural network extrapolates a wideband physical signal from a single narrowband measurement. In WiFi sensing, the WuKong system demonstrates that, given a 20–80 MHz Channel State Information (CSI) measurement, it is possible to directly infer wideband-equivalent CSI (eCSI) covering up to 8-fold greater bandwidth, without extra hardware or measurements, by converting implicit multipath parameters into expanded frequency response (Ji et al., 10 Jan 2026).

This is achieved with FreDiT, a Transformer-Diffusion architecture employing relative frequency embedding (RFE) for precise frequency alignment, and a conditional generative process that leverages both observed and masked (to be predicted) frequency regions. The core theoretical premise is that multipath channel parameters are frequency-invariant at a fixed location, so the mapping from observed CSI to its wideband extension is uniquely determined absent measurement noise.

Models are trained in a self-supervised fashion using masked CSI samples as input and the withheld bands as targets. Evaluation demonstrates substantial improvements in mean squared error (median MSE ≈ 0.27 for 8× bandwidth extrapolation), Pearson correlation of reconstructed channel impulse responses (AccCIR ≈ 0.72–0.78), and applicability to downstream localization and vital-sign monitoring tasks at bandwidths far exceeding the native hardware (Ji et al., 10 Jan 2026).

3. Biological and Neuromorphic NWB: Harmonic Fractal Transformation and Analog Regeneration

The neurophysiological basis for NWB lies in the capacity of neurons to process signals at frequencies orders of magnitude above their firing rate. The FitzHugh–Nagumo (FHN) model, a canonical two-variable oscillator, exhibits spike dynamics that, when driven with broadband or high-modulation-rate stimuli, implement a Harmonic Fractal Transformation (HFT) of bandwidth and sampling rate. The output spike train locks onto harmonics or subharmonics of the input rate, enabling effective up- or down-sampling beyond the spiking bandwidth (Sorokina, 2024).

This transformation follows quantized “staircase” laws:

• For $f_s < f_0$, output bandwidth $B_{\text{out}} = m f_s$, $R = 2 m f_s$ for integer $m = \left\lfloor f_0 / f_s \right\rceil$;
• For $f_s > f_0$, $B_{\text{out}} = f_s / n$, $R = 2 f_s / n$ for $n = \left\lfloor f_s / f_0 \right\rceil$.

This staircase quantization produces fractal suppression of deviations and allows the system to process multi-octave frequency domains efficiently. The FHN oscillator, when realized in an RLC analog circuit, functions as a sigma-delta modulator, achieving first-order noise shaping without explicit delay elements and effecting the so-called 4R-regeneration: re-amplifying, re-shaping, re-timing, and re-modulating the input (Sorokina, 2024).

Simulation results confirm ultra-wideband signal acquisition at integer or sub-integer multiples of the spiking range, and 5–10 dB EVM noise-shaping gain without digital oversampling.

4. Event-Driven Semantic Communications and Neuromorphic NWB

NWB communications integrate neuromorphic sensors, spiking neural networks (SNNs), and ultra-wideband impulse-radio (IR) signaling into an event-driven semantic transmission paradigm. Each sensing device emits spikes corresponding to salient data events, and the SNN encoder converts these into sparse IR pulses transmitted over a multi-antenna fading channel. The receiver, employing a learnable SNN decoder, reconstructs or infers the original semantic content without waiting for full frames (Chen et al., 2022).

The signal model is continuous-time, with IR pulses corresponding to spike events, convolved with multipath channels. Two encoding variants are used: time hopping (random allocation within chip intervals) and learned mapping. The system leverages joint source–channel coding, event sparsity, and per-frame SNN adaptation via hypernetworks informed by pilot signals.

Empirical benchmarks in remote inference (e.g., MNIST-DVS digit classification) establish that NWB systems achieve lower time-to-accuracy and an order-of-magnitude reduction in cumulative energy consumption compared to traditional frame-based digital protocols for the same inference tasks. This efficiency derives from proportional energy use to semantic event content and the absence of extraneous transmissions during low-activity periods (Chen et al., 2022).

5. Performance Metrics and Comparative Outcomes

NWB methods are evaluated via application- and domain-specific metrics:

• In speech coding, SEGSNR gains of 0.2–2 dB are observed, especially for higher quantizer bit rates, with clear preference for allocating more bits to lower bands (Faundez-Zanuy, 2022).
• Neural audio coding demonstrates substantial improvements in normalized mean-square error (NMR), 2f-model MMS, and ViSQOL MOS compared to HE-AAC-v1, achieving equivalent perceptual quality at roughly half the side information bitrate. Subjective listening confirms NWB is preferred by listeners, especially at low bit rates (Choi et al., 7 Jun 2025).
• Wireless sensing: NWB extrapolation reduces mean ToF localization errors and enables multi-person vital sign monitoring with sub-bpm accuracy. Cross-domain generalization is maintained across hardware and environments (Ji et al., 10 Jan 2026).
• Biological NWB achieves staircase sampling-rate conversion and in-hardware noise shaping, with practical prospects for neuromorphic RF front-ends, wideband cochlear implants, and brain–machine interfaces (Sorokina, 2024).
• Semantic NWB communications yield incremental accuracy gains per event and substantial energy savings due to sparse, event-driven IR signaling (Chen et al., 2022).

A summary table compares representative outcomes:

Domain	NWB Method	Gain/Metric	Baseline
Speech Coding	MLP sub-band prediction	+2 dB SEGSNR	Linear LPC
Audio Band Ext.	DNN SBG w/side info	+50% bit-rate reduction	HE-AAC-v1
WiFi Sensing	WuKong (FreDiT) eCSI	MSE ≈ 0.27, AccCIR >0.7	VAE/tiling FIRE
Neuromorphic Comm.	SNN+IR+hypernetwork	10× energy reduction	Frame-digital
Biological UWB	FHN+RLC fractal mapping	5–10 dB EVM, octave BWs	Linear receiver

6. Limitations, Extensions, and Research Directions

Current NWB systems show method-specific limitations:

• Neural audio coding requires separate models per bit-rate and struggles with strongly tonal high-frequency content; future work aims at adaptive models, explicit harmonic priors, and run-time model simplification (Choi et al., 7 Jun 2025).
• WuKong-style extrapolation is currently validated up to 160 MHz; extension to full UWB (>500 MHz) and robust ToF estimation require improved phase calibration, larger datasets, and explicit modeling of hardware impairments (Ji et al., 10 Jan 2026).
• FHN-based analog NWB still needs hardware implementation at GHz frequencies and integration with standard CMOS or III–V platforms for communications/radar applications (Sorokina, 2024).

Research trajectories include cross-band attention for low/high band neural coupling, multi-octave neuromorphic RF front-ends using cascaded nonlinear harmonics, integration with event-based brain–machine interfaces, and lightweight models for mobile deployment.

7. Conceptual Synthesis and Prospects

Neuro-Wideband unites a spectrum of neural and neuromorphic approaches that enable wideband information processing, acquisition, transmission, and reconstruction under tight hardware, energy, or bandwidth constraints. Key distinguishing principles are nonlinear prediction, self-conditioned generative extrapolation, harmonic locking and fractal transformation, and semantically adaptive event-driven signaling. These insights, spanning audio, wireless, analog, and cognitive systems, provide a unified theory and toolkit for wideband operations in both classic and neuromorphic domains (Faundez-Zanuy, 2022, Choi et al., 7 Jun 2025, Ji et al., 10 Jan 2026, Sorokina, 2024, Chen et al., 2022).