File size: 1,670 Bytes
f026b5b
 
 
b67b9e8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
---
license: mit
---

# [WIP] SNAC 🍿

Multi-**S**cale **N**eural **A**udio **C**odec (SNAC) compressess 44.1 kHz audio into discrete codes at a low bitrate.

See GitHub repository: https://github.com/hubertsiuzdak/snac/

## Overview

SNAC encodes audio into hierarchical tokens similarly to SoundStream, EnCodec, and DAC. However, SNAC introduces a simple change where coarse tokens are sampled less frequently,
covering a broader time span.

This can not only save on bitrate, but more importantly this might be very useful for language modeling approaches to
audio generation. E.g. with coarse tokens of ~10 Hz and a context window of 2048 you can effectively model a
consistent structure of an audio track for ~3 minutes.

## Usage

Install it using:

```bash
pip install snac
```

A pretrained model that compresses audio into discrete codes at a 2.2 kbps bitrate is available
at [Hugging Face](https://huggingface.co/hubertsiuzdak/snac). It uses 4 RVQ levels with token rates of 12.5, 25, 50, and
100 Hz.

To encode (and reconstruct) audio with SNAC in Python, use the following code:

```python
import torch
from snac import SNAC

model = SNAC.from_pretrained("hubertsiuzdak/snac").eval().cuda()
audio = torch.randn(1, 1, 44100).cuda()  # B, 1, T

with torch.inference_mode():
    audio_hat, _, codes, _, _ = model(audio)
```

⚠️ Note that `codes` is a list of token sequences of variable lengths, each corresponding to a different temporal
resolution.

```
>>> [code.shape[1] for code in codes]
[13, 26, 52, 104]
```

## Acknowledgements

Module definitions are adapted from the [Descript Audio Codec](https://github.com/descriptinc/descript-audio-codec).