Upscale Audio (UniverSR)

@woongzip11 VersionEnhance

Lifts a band-limited AudioFrame to 48 kHz mono with the vocoder-free UniverSR model, which super-resolves directly in the complex STFT domain via flow matching and reconstructs through an inverse STFT; output rate and layout are fixed.

How it fits

Typical backends

Lift a narrowband recording to 48 kHz before transcription: the file source paces frames, this component super-resolves without a vocoder, and the recognizer reads the result out as JSON.

Denoise first, then super-resolve: a speech denoiser strips background noise so the model receives a clean band-limited signal before reconstructing the high band and transcribing.

Caveats

I/O contractThe output AudioFrame is always 48 kHz mono regardless of input rate or channel layout; the input is downmixed and resampled to the nearest supported bandwidth, so downstream rate expectations must match.
FallbackAn empty input frame is passed through unchanged without invoking the model.
LimitationThe model targets inputs in the 8-24 kHz range and treats anything above that as the top supported bandwidth; it handles speech, music, and sound effects with a single checkpoint.
LatencyThe flow-matching solver runs on a GPU; more solver steps and higher-order methods add fidelity at the cost of latency, while the guidance strength controls how closely the output follows the low-band conditioning.
State lifetimeThe weights source, solver settings, and device are all read once at startup and the model is loaded a single time, so changing any of them requires a restart.

Versions

b58e0689defaultlatest
Automated release
6/24/2026