NovaSR: A 52KB audio model running 3,500x faster than real-time

NovaSR is an ultra-lightweight open-source audio upsampling model weighing in at just 52KB. Its single mission is simple: upscale muffled, low-res 16kHz audio into crisp, clear 48kHz sound.

On a single A100 GPU, NovaSR hits an astonishing 3,600x real-time speed. That means it can process a full hour of audio in roughly one second, requiring practically zero compute or memory overhead.

How it works under the hood

While most modern audio enhancement models rely on heavy diffusion architectures that take up gigabytes of storage, NovaSR strips the neural network down to the absolute bone.

It runs on fewer than 10 tiny Conv1D convolutional layers paired with Snake activation functions borrowed from BigVGAN, preserving frequency quality while keeping parameters negligible.

The training process was equally modest: the author trained it on just 100 hours of public audio data from the mls_sidon and VCTK datasets.

Official benchmarks highlight the contrast. AudioSR sits at around 2,000MB and crawls at 0.6x real-time speed. FlashSR is 1,000MB at 14x real-time, and FlowHigh is 450MB at 20x. NovaSR delivers 3,600x real-time speed while being thousands of times smaller.

Running it and trade-offs

Thanks to its microscopic size, NovaSR does not need an H100 or high-end rig. It runs comfortably on consumer GPUs and laptops. On CPUs, loading it without half-precision provides an additional 3x to 4x speed bump.

That compact size does come with trade-offs. NovaSR strictly accepts 16kHz input, and the output fidelity falls into the low-to-medium quality range. For users who need arbitrary sampling rates and superior fidelity, the creator also released LavaSR, a 50MB follow-up model.

If you want to test it locally, installation is a one-line pip command from the GitHub repository. There is also an interactive Hugging Face Space for instant testing, and a Kaggle notebook covering custom training scripts.

Why it matters

For voice developers, NovaSR solves a frustrating bottleneck in speech synthesis. Many fast TTS engines output low-bandwidth 16kHz audio to keep latency low. Dropping NovaSR onto the output pipeline upgrades that speech to 48kHz almost instantaneously.

Because it only takes up 52KB, it can also be embedded directly into edge hardware like smart earbuds, conference microphones, and mobile calling apps to restore lost high-frequency voice details in real time.

We think NovaSR represents a welcome departure from brute-force scaling. It proves that clever, specialized architectures can deliver massive real-time utility on constrained hardware without dragging along the baggage of multi-gigabyte models.