---
license: cc-by-nc-4.0
language:
  - en
tags:
  - mlx
  - audio
  - music
  - audio-super-resolution
  - audiosr
  - latent-diffusion
  - quantized
  - int4
base_model: haoheliu/audiosr_basic
library_name: mlx
pipeline_tag: audio-to-audio
---

# AudioSR-MLX-4bit

- [speech-swift](https://github.com/soniqo/speech-swift) — Apple SDK
- [soniqo.audio](https://soniqo.audio) — website
- [blog](https://soniqo.audio/blog) — blog

MLX port of [AudioSR](https://github.com/haoheliu/versatile_audio_super_resolution) —
a **multi-step latent-diffusion audio super-resolution model** — quantized to
**INT4** weight-only for on-device inference on Apple Silicon. Upsamples
any-rate input (mono) to **48 kHz** with a 200-step DDIM solver (v-prediction
+ classifier-free guidance) over an AudioLDM-style VAE latent, then decodes
via a HiFi-GAN vocoder. AudioSR is the **teacher** model that
[FlashSR](https://huggingface.co/aufklarer/FlashSR-MLX-4bit) was distilled
from — FlashSR runs the same architecture in 1 step instead of 200, at the
cost of some quality.

## Model

| | |
|---|---|
| Total parameters | 672 M (VAE 223 M + UNet 258 M + Vocoder 190 M) |
| Diffusion | 200-step DDIM, v-prediction, cosine schedule, η=0 |
| Classifier-free guidance | scale 3.5 (cond + uncond passes per step) |
| Quantization | INT4 weight-only, group size 64, mode `mlx_affine_flat` |
| Format | MLX safetensors (single combined bundle) |
| Sample rate | 48 kHz mono out (any-rate mono in) |
| Frame length | 10.24 s (491 520 samples) per forward |
| Bundle size | 365 MB on disk |
| Source | [haoheliu/audiosr_basic](https://huggingface.co/haoheliu/audiosr_basic) |

## Files

| File | Size | Description |
|---|---|---|
| `model.safetensors` | 365 MB | INT4-quantized VAE + UNet + HiFi-GAN weights |
| `config.json` | ~80 KB | Sub-model configs + quantization metadata + per-tensor shape table for dequant-on-load |

The three sub-models share one safetensors file with `vae.*`, `ldm.*`, `voc.*`
key prefixes. `config.quantized_shapes` records each tensor's pre-flatten
shape so `mx.dequantize` can rebuild conv weight tensors at load time.

## Performance (Apple Silicon, M-series, 10.24 s @ 48 kHz)

| Metric | Value |
|---|---|
| Real-time factor (wall / audio) | **1.87** |
| Load time | 0.06 s |
| SNR vs FP16 reference | **+29.2 dB** |
| Peak amplitude | 0.500 (renormalised per upstream) |

AudioSR is markedly slower than FlashSR (RTF 1.87 vs FlashSR's 1.10 on the
same hardware) because the 200-step DDIM is the inherent bottleneck. Choose
AudioSR when the teacher's quality is preferred; choose FlashSR for
real-time on-device use.

## Usage

```python
from huggingface_hub import snapshot_download
import mlx.core as mx
import numpy as np
import scipy.io.wavfile as wf
from scipy.signal import resample_poly

bundle = snapshot_download("aufklarer/AudioSR-MLX-4bit")
# See https://github.com/soniqo/speech-swift for production usage.

# Toy Python demo (requires the matching MLX AudioSR runtime):
sr, audio = wf.read("lr.wav")
audio = audio.astype(np.float32) / 32767.0
audio_48 = resample_poly(audio, 48000, sr).astype(np.float32)

from audiosr import AudioSR
model = AudioSR(bundle)
hr = model(mx.array(audio_48), steps=200, cfg_scale=3.5, seed=42)
mx.eval(hr)
wf.write("hr.wav", 48000, (np.clip(np.array(hr), -1, 1) * 32767).astype(np.int16))
```

## Source

- Upstream: [haoheliu/audiosr_basic](https://huggingface.co/haoheliu/audiosr_basic)
- Paper: [AudioSR: Versatile Audio Super-resolution at Scale](https://arxiv.org/abs/2309.07314)
- Distilled student (1-step): [FlashSR](https://github.com/jakeoneijk/FlashSR_Inference)

## License

**CC-BY-NC 4.0** — inherited from upstream AudioSR weights. Non-commercial use only.