A battery-free earbud or hearing aid could run on radio energy. The access point sends one signal, and the device splits the incoming power in two. One part charges the device and the other part is decoded into audio. The more power goes to charging, the noisier the decoded audio becomes. Wireless engineers choose this split by trading data rate, or bit error probability, against harvested energy. Suppose the audio arrives as tokens from a neural codec, meaning a compressor that encodes sound as a stack of discrete codes, coarse at the bottom and fine at the top. Then the error rate that engineers optimize says little about what listeners actually hear. Standard audio quality metrics can also mislead in the opposite direction. On one neural codec they stop improving after three of nine layers, while listeners keep preferring more layers.
This note asks a narrow question. Does the choice of quality metric change where a simultaneous wireless information and power transfer (SWIPT) receiver should set its power split? If so, can a receiver that is tuned against real listening scores harvest more energy for the same perceived quality? The proposal builds that receiver in simulation, measures how far apart the operating points chosen by different metrics sit, and tests whether the difference survives two simpler explanations: a lookup table built from the codec's clean bitrate-quality curve, and a better objective metric.
Core Claim
A power-splitting receiver sends a fraction $\rho$ of the incoming radio power to its information decoder and $1-\rho$ to its rectifier, the diode circuit that turns radio power into stored charge. The receiver also chooses $L$, the number of codec layers it decodes. For a quality measure $m$, the operating point is
$$ (\rho^*_m, L^*_m) = \arg\max_{\rho, L} \; Q_m(\rho, L) \quad \text{s.t.} \quad E(\rho) \ge E_{\min}. $$
Here $Q_m$ is audio quality under measure $m$, and $E(\rho)$ is harvested energy under a non-linear rectifier model. The object of study is the metric-induced regret. This is the listener score lost when the receiver uses $(\rho^*_m, L^*_m)$ instead of the listener-optimal point, at equal harvested energy.
We claim this regret is large under signal-matching metrics for a specific reason. The receiver's information decoder needs about $-60$ dBm, while the harvester needs about $-10$ dBm (Optimal Design of SWIPT-Aware Fog Computing Networks). That gap pushes good operating points toward very small $\rho$. Small $\rho$ is exactly the regime where token errors are nonzero but not catastrophic. In that regime a metric like log-spectral distance (LSD) is flat above three layers (Geometric Iterative Retrieval for Neural Audio Codec Resynthesis). It will therefore endorse starving the decoder of the upper layers that listeners still value.
The second, sharper claim is about corrupted tokens. We expect a wrongly decoded upper-layer token to sound worse than a dropped one, because it adds harshness where a drop only adds smoothness. If that holds, the receiver should decide layer by layer whether to decode or discard, based on its estimated error rate. Neither the bit-error view nor the clean bitrate-quality curve captures this decision.
Research Program
Channel side. We reuse the short-packet, finite-constellation analysis of Simultaneous Information and Energy Transmission with Short Packets and Finite Constellations and Achievable Information-Energy Region in the Finite Block-Length Regime with Finite Constellations. That work bounds decoding error and energy outage for BPSK, QPSK and 16-QAM under a non-linear rectifier. It also locates a "knee" in the rate-energy curve, beyond which small energy gains cost a lot of rate. Its error model stops at the physical symbol. We carry each symbol error forward into a specific codec token at a specific layer.
Codec side. Tokens come from the Descript Audio Codec with nine layers of residual vector quantization, where each layer quantizes what the previous layers left over (High-Fidelity Audio Compression with Improved RVQGAN). A second codec is EnCodec, whose bitrate is set by dropping layers (High Fidelity Neural Audio Compression). Packets carry one layer group each, so the base layer can be given stronger coding than the upper layers.
Receiver side. The receiver chooses $\rho$ and a per-layer decode-or-discard rule from the pilot-estimated signal-to-noise ratio. Discarded layers are either zeroed or refilled by a model that predicts them from the coarse layers. For refilling we use the contrastive retrieval model of Schmidt-Traub et al., which picks plausible codebook vectors rather than averaging them. That paper assumes error-free tokens; we test it under realistic partial corruption. The existing ResilientAudio repository is a starting point for simulating channel-corrupted codec streams.
Measurement side. Every configuration is scored four ways: LSD, scale-invariant signal-to-distortion ratio (SI-SDR), ViSQOL v3 (ViSQOL v3), and crowdsourced ITU-T P.808 mean opinion scores. P.808 was validated at 0.90 Spearman correlation against lab tests (The INTERSPEECH 2020 Deep Noise Suppression Challenge). For music we add MUSHRA-style comparisons.
The listening scores are expensive, so we also fit a small listener-calibrated proxy on held-out ratings over the $(\rho, L, \text{error rate})$ grid. The receiver policy is tuned against this proxy. The proxy is not claimed as a new metric; it is only a cheap stand-in for a known-to-be-expensive ground truth.
We deliberately do not pursue two nearby ideas. The first is constellation shaping that maps codebook geometry onto the radio constellation, which would need its own theory. The second is edge offloading as in the fog-computing setting; its bit-volume model is orthogonal to this question.
Baselines
Do nothing. This baseline shows what current practice would choose. The receiver always decodes all nine layers and picks $\rho$ to meet a fixed decoding-error target, as a standard SWIPT design would. It is naive but it is the default. Beating it means higher P.808 scores at equal harvested energy, or more harvested energy at equal P.808 scores.
Error-probability-optimal unequal protection. This is the strongest wireless-side competitor. It uses the same layered packets, but protection and $\rho$ are allocated to minimize a weighted decoding error probability using the finite-blocklength bounds of ul Zuhra et al. It gets the same channel, the same codec and the same search budget over $(\rho, \text{code rates})$ as our policy. It is strong because unequal error protection already captures "base layers matter more." The question is whether its weights, chosen from bit counts, match what listeners value.
Composition lookup. This is the strongest competitor overall. Take the rate-energy frontier, convert rate into a number of layers, and read quality off the codec's clean layer-versus-quality curve. The curve is measured once with P.808 on uncorrupted DAC tokens. This baseline ignores corrupted-but-decoded tokens and nothing else, so it isolates our second claim. It gets the same listener rating budget, spent on clean decodes instead of corrupted ones.
Objective-metric-tuned receiver. This uses our full decode-or-discard policy, but tuned against ViSQOL v3 and, separately, against LSD instead of the listener proxy. It isolates whether listener calibration is the key ingredient.
Ablation without discarding. Our policy with discarding disabled, so that all received layers are decoded however noisy they are. This isolates the decode-or-discard rule.
Reference numbers. On the clean nine-layer DAC benchmark, contrastive retrieval reaches an LSD of 10.61 against 12.42 for cross-entropy prediction. It also holds a 45-point lead on a 100-point listening scale (Geometric Iterative Retrieval). DAC reaches an SI-SDR of 10.75 at 8 kbps, against EnCodec's 8.36 at 12 kbps (Improved RVQGAN). No reported result was found on Lacuna for any neural codec transmitted over a SWIPT link. The baselines above therefore have to be run by us, not quoted.
Why not just…?
Why not just use Opus over a standard SWIPT link? EnCodec at 3 kbps was rated above Opus at 12 kbps (High Fidelity Neural Audio Compression). At equal quality, Opus needs about four times the airtime, and that airtime comes out of harvesting. Opus also has no layer structure to truncate on the fly. We still include Opus at 6 to 12 kbps as a reference curve, scored with ViSQOL v3, which was tuned to behave monotonically on Opus. If Opus matched the neural stack at equal harvested energy, the premise would fail, and we would report that.
Why not just compose the rate-energy curve with the codec's clean quality-versus-bitrate curve? This is the real threat. If quality depends only on how many layers arrive intact, the whole problem reduces to a lookup, and our regret measurement adds nothing beyond "use listener scores, not LSD." The composition-lookup baseline tests this directly. If its predicted frontier sits within the P.808 confidence interval of the measured frontier across the low-$\rho$ regime, we drop the corruption claim and publish the lookup as the recommended design. The proposal only has substance beyond that if corrupted tokens behave differently from dropped ones.
Why not just optimize ViSQOL v3 or another objective metric and skip listeners? ViSQOL v3 was built around aligned, full-reference comparisons of classical codecs. Its authors and later work note that full-reference metrics penalize generated detail that does not line up sample by sample (ViSQOL v3; Diffiner). That is precisely what refilled upper layers produce. Still, it may be good enough in this regime. The objective-metric-tuned baseline measures its regret directly. If that regret is small, the practical recommendation becomes "tune on ViSQOL," which is much cheaper than listening tests.
Why not just stay with decoding error probability? Error probability is rigorous and tractable (ul Zuhra et al.). It weights every bit equally, however, while listeners do not weight layers equally (Geometric Iterative Retrieval). The error-probability-optimal unequal-protection baseline gives this view its best form, so any gain we report is over a fair version of it.
Proposed Experiments
The main experiment sweeps $\rho$ from near zero to 0.5, all three constellations, packet lengths in the short-packet range, and received power levels straddling the knee of the rate-energy curve. Test material is LibriTTS speech at 24 kHz (LibriTTS) and MUSDB and MTG-Jamendo music encoded with DAC and EnCodec. For each energy target $E_{\min}$ we compute every baseline's operating point and collect P.808 ratings. Music gets MUSHRA ratings.
Decisive check one. Is the LSD-chosen and ViSQOL-chosen $\rho$ different from the listener-chosen $\rho$ by more than the listener confidence interval, and does that gap cost P.808 score at equal energy? If not, metric divergence does not matter for power splitting, and the note becomes a negative result.
Decisive check two. At matched token error rates in layers four to nine, do listeners prefer discarding (zeroing or retrieval-refilled) over decoding the corrupted tokens? This is the one outcome the composition lookup cannot produce.
Decisive check three. At a fixed P.808 score, how much more energy does the full listener-tuned policy harvest than the error-probability-optimal baseline? We report this in microjoules per second of audio across channel conditions. We also include a sensitivity analysis with imperfect channel estimates, since multi-antenna and reconfigurable-surface systems are known to degrade under poor channel feedback.
A secondary measurement records decoder compute when fewer layers are decoded. On a harvested budget, decoding cost competes with radio energy, even though EnCodec already runs about ten times faster than real time on one laptop CPU thread.
References
- Optimal Design of SWIPT-Aware Fog Computing Networks
- Geometric Iterative Retrieval for Neural Audio Codec Resynthesis
- Simultaneous Information and Energy Transmission With Short Packets and Finite Constellations
- Achievable Information-Energy Region in the Finite Block-Length Regime with Finite Constellations
- High-Fidelity Audio Compression with Improved RVQGAN
- High Fidelity Neural Audio Compression
- ViSQOL v3: An Open Source Production Ready Objective Speech and Audio Metric
- The INTERSPEECH 2020 Deep Noise Suppression Challenge: Datasets,\n Subjective Testing Framework, and Challenge Results
- Diffusion-based Signal Refiner for Speech Enhancement and Separation
- LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech