NovaSR: A 52KB Audio Super-Resolution Model that Upscales 16kHz Audio to Clear 48kHz

NovaSR is an open-source audio super-resolution model weighing only 52KB—smaller than a three‑second audio clip—yet capable of enhancing fuzzy 16kHz recordings into clearer 48kHz versions.

According to the developers, the model runs extremely fast: on a single GPU it can process between 100 and 3600 seconds of audio in one second, effectively enabling near‑real‑time streaming enhancement.

Improving TTS quality: most text‑to‑speech systems output 16kHz or 24kHz audio; NovaSR can boost these outputs with virtually no additional compute cost.

Rapidly fixing low‑quality audio datasets: the model provides a quick solution for cleaning large audio corpora.

Minimal device footprint: at 52KB the model can be deployed on any hardware, including resource‑constrained edge devices.

Community testing reports that, compared with the previously released FlashSR, NovaSR handles female voices better, reducing harsh sibilance and audio artifacts. While it does not yet match the higher‑fidelity FlowHigh model, NovaSR achieves a favorable balance of speed and quality.

The model was trained on only 100 hours of audio data, and the developers acknowledge room for improvement. Nevertheless, given its tiny size, the current audio quality is already impressive.

For developers seeking a fast, lightweight audio enhancement solution, NovaSR is worth trying.

Links:

GitHub repository: https://github.com/ysharma3501/NovaSR

Model and examples: https://huggingface.co/YatharthS/NovaSR

Online demo (runs on a weak CPU at roughly 10× real‑time speed): https://huggingface.co/spaces/YatharthS/NovaSR