TinySR Readme!
TinySR is the world's smallest super-resolution model (for audio) that I trained!
It's only 57.6 KB in size and it can easily run on a CPU or GPU.
Details
- Input Sample Rate: 16 kHz
- Output Sample Rate: 48 kHz
- Size: Only 57.6 KB
The model works in mono.
Training
I used just 46 minutes of instrumental music generated with Mubert, without using any other type of audio!
Benchmark Comparison
| Model | By Myself? | Size |
|---|---|---|
| TinySR (MihaiPopa-1/TinySR) | ✅ | 57.6KB |
| FlashSR (YatharthS/FlashSR) | ❌ | 2MB (PyTorch) / 500KB (ONNX) |
| AudioSR (haoheliu/audiosr_basic) | ❌ | 6.18GB |
Other things
I used a 16-second clip generated with Suno (with vocals) and the model worked!
It's not exactly like the ground truth (and not perfect), but it's better than I thought!