My Intel Core Ultra 7 laptop has an NPU, but in my everyday work it rarely had anything to do. Leaving it idle felt wasteful, so I started looking for a useful local experiment.
That became Fan Composer: a small, open-source Windows app that listens to laptop fan noise and adds a quiet soundscape around it.
The question is simple: could a steady fan sound feel less annoying if it became part of a musical background?
This is not active noise cancellation. It adds sound; it does not remove the fan noise or lower its sound pressure.
Three ways to listen
- Harmonic: a continuous drone and overtones guided by a stable fan pitch estimate.
- Nature: synthesized wind and wave-like textures, without recorded nature samples.
- Generative: a seeded musical motif with an ambient background and slow changes.
All three are synthesized locally. There are no cloud music APIs or downloaded music samples.
A short sound preview
0:00 Harmonic ยท 0:15 Nature ยท 0:30 Generative. These are offline renders from the actual Rust synthesis engine, paired with a static screenshot of the stopped app. Preview gain is raised by about 12 dB for audibility. This is not a recording of live microphone analysis or NPU inference.
Automatic volume is the core feature
You set a maximum volume, and the app gradually adjusts playback as the estimated fan/background level changes. Automatic volume is always on; there is no extra switch to manage.
There is a catch: the microphone can also hear the app's own sound. Simply following microphone loudness could make playback chase itself. Fan Composer uses its output signal as a spectral reference and holds or reduces amplification when the estimate is uncertain.
This remains an experimental background estimate. Other steady sounds and room changes can affect it; it is not an acoustic echo canceller or a reliable fan-versus-speech classifier.
What the NPU actually does
The NPU runs a small music-policy model through OpenVINO, once per second. It adjusts parameters such as brightness, note density, motion, and reverb. Rust on the CPU handles FFT analysis, volume control, and audio synthesis.
The app tries NPU first, then OpenVINO CPU, then CPU rules. It checks successful inference and the actual execution device, rather than assuming that an enumerated NPU is doing the work. The current backend is visible in the interface.
The bundled model imitates a rule-based policy. I am not claiming that it makes better music or saves power compared with the CPU; those comparisons still need proper measurements.
Try it on Windows
- Download the Windows x64 release and extract the entire ZIP.
- Run
fan-composer.exeand choose your speaker. - Choose a mode, set a comfortable maximum volume, and press Start.
- Allow the initial 3.5-second quiet measurement, then listen. Stop is always available.
Nothing plays automatically at launch. Microphone audio stays in memory and is not recorded or uploaded. Settings and explicit feedback stay local.
For Rust developers, the stack is iced + CPAL/WASAPI + rustfft, with the OpenVINO C API for the music policy. Build on Windows with the Rust MSVC toolchain and C++ Build Tools:
git clone https://github.com/zendoclab/fan-composer.git
cd fan-composer
cargo build --release --locked
cargo test --release --locked
The build guide also covers packaging the model and NPU runtime. A source build without that runtime still works through CPU rules.
What is established so far?
Version 0.3.0 passed 26 automated tests, including synthetic analysis, synthesis safety, and automatic-volume checks. Real NPU inference and a short, silent Windows audio I/O run were also verified. These checks do not establish long-session stability or improved listening comfort.
My next question is whether adaptive sound is actually more comfortable than fan noise alone or a fixed ambient track, at comparable overall loudness. Listener feedback and controlled comparisons are welcome.
- Source, documentation, and downloads: GitHub
- Methods and engineering results: Zenodo preprint
Built by Kyu Lee, cording.ai. Source code and the locally generated policy model are MIT-licensed.
Top comments (0)