SNAC

1.2.1 verified Fri May 01 auth: no python

Multi-Scale Neural Audio Codec for audio compression, supporting 24 kHz, 32 kHz, and 44 kHz sampling rates. This is a PyTorch-based library for encoding audio into discrete codes (suitable for language modeling) and decoding back to waveform. Current version 1.2.1 has a stable API with `encode` and `decode` methods.

pip install snac

Common errors

error ImportError: cannot import name 'SNAC' from 'snac' ↓

cause SNAC was not installed correctly or an incompatible version is installed.

fix

Ensure you installed the correct package: pip install snac. Check that you are not shadowing the package with a local file named snac.py.

error RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! ↓

cause The model is on one device, but the input tensor is on another.

fix

Move the model and input to the same device: model = model.to('cuda'); audio = audio.to('cuda').

error OSError: Can't load tokenizer for 'hubertsiuzdak/snac_24khz' ↓

cause The model name is incorrect or the model is not publicly accessible.

fix

Use a valid Hugging Face model ID (e.g., 'hubertsiuzdak/snac_24khz', 'hubertsiuzdak/snac_32khz', 'hubertsiuzdak/snac_44khz') or ensure your network can access huggingface.co.

Warnings

gotcha The `encode` method returns a list of tensors (one per layer) in version <1.2.0, but returns a single stacked tensor in 1.2.0+. Check your version and adjust code accordingly. ↓

fix Upgrade to >=1.2.0 or use `codes = model.encode(audio)` and handle list.

deprecated Loading models from a local filepath was broken in 1.2.0 and fixed in 1.2.1. If you use `SNAC.from_pretrained('./local_model')`, ensure version >=1.2.1. ↓

fix Upgrade to 1.2.1 or use a Hugging Face model ID.

gotcha The model expects audio resampled to the model's sample rate (24kHz, 32kHz, or 44kHz). Failure to resample will produce garbled output. ↓

fix Resample input audio to match the model's sample rate before encoding.

breaking Version 1.0.0 introduced a completely new architecture and model zoo. Models from v0.x (if they existed) are incompatible. ↓

fix Use only v1.x models and upgrade to latest version.

Imports

SNAC
```
from snac import SNAC
```

Quickstart

Load a pretrained model, encode audio to discrete codes, and decode back to audio.

import torch
from snac import SNAC

model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
audio = torch.randn(1, 1, 24000)  # 1 second of 24 kHz audio
codes = model.encode(audio)
print("Codes shape:", codes.shape)
reconstructed = model.decode(codes)
print("Audio shape:", reconstructed.shape)