Physical reservoir computing (PRC) exploits the intrinsic, nonlinear dynamics of physical materials to overcome the fundamental bandwidth and energy limitations of conventional von Neumann architectures (Fig. 1A). However, a critical gap remains between its conceptual validation and real-world implementation, stemming from the challenge of achieving high performance while maintaining a hardware-feasible, real-time architecture [1]. This challenge arises because extracting sufficiently high-dimensional information from a material's response typically demands implementation-heavy extraction schemes with massive readout channels. Conventional time multiplexing uses sequential time-domain sampling that requires substantial hardware overhead in the form of complex integrated circuits (ICs) for precise, synchronized clocking and high-speed sampling (Fig. 1B). Frequency-domain features of material dynamics have emerged as an alternative, powerful computational resource, but existing demonstrations generally rely on long-observation-window spectroscopic measurements that are themselves hardware-intensive and unsuited to real-time operation. Moreover, real-time operation inherently limits observation time, so only coarse, discretized spectra remain accessible at high processing speed due to a fundamental time-frequency trade-off (Fig. 1C) — raising the open question of whether coarse spectra can still support complex neuromorphic computation.

Here, we present spectral dynamics reservoir computing (SDRC), a framework that answers this question affirmatively by extracting reservoir states directly from coarse spectra using a bank of selective analogue filters and envelope detectors (Fig. 1D), enabling high-performance computation with substantially fewer readout channels than conventional extraction schemes [2]. We demonstrate SDRC experimentally in a spin-wave physical reservoir — an established PRC platform owing to its rich nonlinear dynamics and GHz-range operation [3–5] — implementing the entire readout with passive filters and diodes, achieving strong benchmark performance and 98.0% accuracy on real-world spoken-digit recognition without cochleagram preprocessing. These results show that coarse spectra, rather than a limitation of real-time operation, can be harnessed as an efficient computational resource, establishing a hardware-efficient route toward real-time neuromorphic processors.

1. H. Kurebayashi et al., Nat. Rev. Phys. 8, 208(2026) DOI: 10.1038/s42254-025-00918-1
2. J. Chen et al., arXiv:2603.04901 (2026) DOI: 10.48550/arXiv.2603.04901
3. J. Chen et al., Phys. Rev. Appl. 23, 034045(2025) DOI: 10.1103/PhysRevApplied.23.034045
4. W. Namiki et al., Adv. Intell. Syst. 5, 2300228(2023) DOI: 10.1002/aisy.202300228
5. J. Chen et al., Phys. Rev. B 113, 064421(2026) DOI: 10.1103/29mw-s88k