Preserving a Prompt Through Air and Hardware

SWhisper’s acoustic delivery layer tackles a different problem from prompt construction. Once a voice-compatible instruction has been produced, how can its meaning survive a path that includes air, a commodity speaker, an unfamiliar microphone, nonlinear hardware, and environmental variation? The focus here is channel-inversion pre-compensation: modeling how the receiver changes a signal and modifying the transmitted waveform so that the recovered baseband more closely resembles the intended speech. Understanding this distinction lets the reader judge SWhisper as an engineering system rather than as a claim that a high-frequency carrier is automatically reliable.

The receiver path is difficult because it is not a single, stable filter. It combines propagation with microphone behavior, and the behavior can change with frequency, device, position, and surroundings. A method that works for one recorded waveform may fail when a different microphone attenuates a component or when reflections alter the relative timing of components. SWhisper’s central design decision is to treat those transformations as a channel to be measured and approximated, not as random imperfections that can be ignored.

A channel is a transformation, not just noise

In communications engineering, a channel describes how an input becomes an output. For SWhisper, the input is an intended baseband speech signal that will ultimately be used by a speech-driven model. The output is the signal recovered after near-ultrasonic transmission and microphone processing. If the channel were known and stable, the sender could choose a transmitted waveform whose output matched the target. The difficulty is that the mapping includes nonlinear frequency coupling: energy at one input frequency can contribute to several output frequencies after the microphone responds to the carrier and embedded voice content.

That is why SWhisper estimates an empirical transfer matrix rather than assuming one gain value for the whole signal. Conceptually, the matrix records how components placed at different input frequencies appear in the recovered output. The study obtains this behavior from probe signals sent through the relevant acoustic path under multiple devices and environments. The resulting average is intended to capture recurring behavior across receivers, including the cross-frequency effects introduced by demodulation and hardware nonlinearity.

The word empirical is important. The model is not derived solely from an ideal microphone circuit, because the useful behavior depends on the complete path that the signal actually travels. It includes the speaker’s high-frequency response, propagation through the room, the microphone diaphragm and preamplifier, front-end filtering, and the processing that follows capture. Measuring the compound behavior gives SWhisper a practical approximation of what a target waveform is likely to become.

The target of the compensation is not the carrier itself. It is the recovered baseband speech. The system begins with a desired speech spectrum and asks what transmitted spectrum would make the channel output resemble that target. In a simple model, this looks like applying the inverse of the channel mapping.