Upscale Audio (AudioSR)

@haoheliu1 VersionEnhance

Reconstructs the missing high band of a 2-16 kHz AudioFrame with the AudioSR latent-diffusion model and emits a 48 kHz mono AudioFrame; output rate and layout are fixed regardless of the input.

How it fits

Typical backends

Lift a narrowband recording to 48 kHz before transcription: the file source paces frames, this component restores the high band, and the recognizer reads the cleaned audio out as JSON.

Denoise first, then super-resolve: a speech denoiser removes background noise so the diffusion model is fed the clean low-pass signal it expects before upsampling 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 internally, so downstream rate expectations must match.
FallbackAn empty input emits an empty 48 kHz frame, and a frame too short to upsample is passed through resampled to 48 kHz without invoking the model.
AccuracyThe model expects a clean low-pass-limited signal; strong noise, reverb, or lossy-codec artifacts in the input degrade the reconstructed high band, and the band outside the 2-16 kHz training range is not the design target.
LatencyThis is a latent diffusion model whose sampler runs many iterative steps, so a GPU is expected and CPU inference is impractical; reducing the step count trades quality for speed.
Parameter interactionThe checkpoint choice and device are hot-reloadable: changing either rebuilds the model on the next call, while the sampler step count, guidance strength, and seed take effect per call and a fixed seed makes runs reproducible.

Versions

e59b9239defaultlatest
Automated release
6/24/2026