EnCodec

EnCodec

EnCodec is Meta AI's high-fidelity neural audio codec that compresses audio at extremely low bitrates using a streaming encoder-decoder with residual vector quantization (RVQ). It serves as the audio tokenizer for MusicGen, AudioGen, and the broader AudioCraft ecosystem.

Key Features

• Extreme compression — 24 kHz stereo audio at 24 kbps; 48 kHz stereo at 24 kbps; monaural at 1.5 kbps
• Residual Vector Quantization — multi-codebook RVQ enables hierarchical audio token representations
• Streaming architecture — processes audio in real-time with minimal latency (10 ms chunks)
• Perceptual losses — spectral and feature-matching discriminators preserve perceptual quality
• Scalable bandwidth — a single model gracefully degrades across 1.5 / 3 / 6 / 12 / 24 kbps
• Foundation for LLMs — convert audio to discrete tokens for language model conditioning

Quick Start

pip install encodec

import torchaudio
from encodec import EncodecModel
from encodec.utils import convert_audio

model = EncodecModel.encodec_model_24khz()
model.set_target_bandwidth(6.0)

wav, sr = torchaudio.load("audio.wav")
wav = convert_audio(wav, sr, model.sample_rate, model.channels)

encoded_frames = model.encode(wav.unsqueeze(0))

Install via ai-supply

npx ai-supply add encodec-neural-audio-codec

Curated mirror of the open-source EnCodec (MIT). Get it from the source.