Aller au contenu
ModelPrésentéExterne
Télécharger

MagiCodec 16k 50hz

MagiCodec is a single-layer, streaming codec model that delivers state-of-the-art audio quality and highly model-able discrete tokens.

TypeOtherVersionRevision 36b011fBClasse

Source du modèle

Type
Other
Taille du fichier
800 MB
Version
vRevision 36b011f
Mis à jour
22 sept. 2026
Téléchargements
72
Source
Hugging Face

Extrait de la source

MagiCodec is a single-layer, streaming codec model that delivers state-of-the-art audio quality and highly model-able discrete tokens.

Sources

1 source
Hugging FacePrincipalVersion 36b011f5e36591c1f28213d882547bbb92851774
huggingface.co/Ereboas/MagiCodec_16k_50hz/tree/36b011f5e36591c1f28213d882547bbb92851774

Vérifié 5 sept.

Artefacts du modèle

1 artefact

MagiCodec-50Hz-Base.ckpt

ckpt · 800 MB · SHA-256 30e05d3f8ec1…a134 · Hugging Face

Télécharger

Extraits de sources

2 extraits

--- language: - en license: mit pipeline_tag: audio-to-audio tags: - pytorch --- # MagiCodec: Simple Masked Gaussian-Injected Audio Codec for High-Fidelity Reconstruction & Generation [![github](https://img.shields.io/badge/Code-Repo-black?logo=github)](https://github.com/Ereboas/MagiCodec) [![arXiv](https://img.shields.io/badge/%F0%9F%93%84%20ArXiv-Paper-<COLOR>.svg)](https://arxiv.org/abs/2506.00385) [![demo](https://img.shields.io/badge/%F0%9F%94%97%20MagiCodec-Demo-blue)](https://ereboas.github.io/MagiCodec/) [![model](https://img.shields.io/badge/%F0%9F%A4%97%20MagiCodec-Models-blueviolet)](https://huggingface.co/Ereboas/MagiCodec_16k_50hz) MagiCodec is a **single-layer**, **streaming** codec model that delivers state-of-the-art audio quality *and* highly model-able discrete tokens. This is the code for the MagiCodec neural audio codec presented in the paper [MagiCodec: Simple **Ma**sked **G**aussian-**I**njected Audio **Codec** for High-Fidelity Reconstruction and Generation](https://arxiv.org/pdf/2506.00385) [[abs](https://arxiv.org/abs/2506.00385)]. ## ✨ Key Features - **Single‑layer streaming design** – lightweight causal Transformer that drops straight into audio‑native LLMs with minimal latency. - **Low bit‑rate & compute‑efficient** – <= 850 bps and <= 40 ms look‑ahead keep bandwidth and FLOPs tiny for on‑device use. - **Rich acoustic *and* semantic information** – captures fine‑grained detail plus high‑level semantics, achieving top scores in both waveform reconstruction and downstream tasks. ## 🎧 Samples To get a quick sense of our codec’s performance, please check out the [Sample Page](). **Comprehensive benchmarks covering a variety of baselines are available on this page.** ## 🛠️ Installation Our released code is based on a specific version ([2.3.6](https://github.com/Dao-AILab/flash-attention/releases/tag/v2.3.6)) of [FlashAttention](https://github.com/Dao-AILab/flash-attention/tree/92dd5703ecdb99aa4a4aee9817f28557907403a2). Please note that different (including newer or older) versions of flash-attention may introduce changes to the C/Python interfaces called by our custom ops. Compatibility issues may arise if you use a different version of flash-attention. We recommend creating a fresh conda environment for installation. ### Env Setup ```bash conda create -n MagiCodec python=3.9 -y conda activate MagiCodec ``` ### Basic R...

Source context: 72 downloads · 1 likes · Pipeline audio-to-audio · Repo Ereboas/MagiCodec_16k_50hz