AudioCraft

Generates high-fidelity music from text descriptions using MusicGen, a transformer-based language model that operates on discrete audio tokens produced by EnCodec. The model uses a two-stage pipeline: text conditioning through embeddings, followed by autoregressive token generation that is decoded back to waveform audio. Supports duration control, temperature sampling, and top-k/top-p filtering for output variation.

Generates diverse sound effects and ambient audio from text descriptions using AudioGen, a variant of the MusicGen architecture adapted for non-musical audio. Operates through the same tokenization-generation-decoding pipeline but trained on sound effect datasets with different conditioning strategies optimized for environmental and synthetic sounds.

Provides a modular configuration system enabling composition of different components (compression models, language models, conditioning systems) into custom audio generation pipelines. Models are defined through YAML/JSON configs that specify architecture, hyperparameters, and component connections. Enables swapping components (e.g., using different encoders or decoders) without code changes.

Provides utilities for audio loading, resampling, normalization, and feature extraction (spectrograms, mel-spectrograms, MFCC, chroma features). Includes wrappers around librosa and torchaudio for efficient batch processing. Enables preprocessing of audio for training and inference, and extraction of audio features for analysis or conditioning.

Provides unified inference API for loading and using pre-trained AudioCraft models (MusicGen, AudioGen, MAGNeT, JASCO, etc.) with automatic model downloading, caching, and device management. Abstracts away model-specific implementation details, providing consistent interface across different generation models. Handles model loading, GPU memory management, and inference batching.

Compresses audio to discrete token sequences using EnCodec, a neural codec that learns to represent audio as quantized embeddings across multiple codebooks. The codec operates as an autoencoder with a residual vector quantizer, enabling variable bitrate compression (1.5-24 kbps) while maintaining perceptual quality. Serves as the tokenizer for all downstream generation models in AudioCraft.

Generates music from text descriptions while conditioning on a reference audio style using MusicGen-Style. The model extends MusicGen with dual conditioning: text embeddings for semantic content and audio embeddings extracted from a reference track for stylistic characteristics. Style embeddings are computed via a separate audio encoder, then jointly processed with text through the transformer decoder.

Generates music and sound effects using MAGNeT, a non-autoregressive transformer that predicts all tokens in parallel rather than sequentially. Uses iterative refinement with confidence-based masking: initially predicts all tokens, then iteratively refines low-confidence predictions in subsequent passes. Achieves faster inference than autoregressive models at the cost of potential quality trade-offs.

Generates music conditioned on explicit musical structure using JASCO (Joint Audio-Symbolic Conditioning), which accepts text descriptions alongside chord progressions, melody contours, and drum patterns. The model processes symbolic music inputs (represented as token sequences) through dedicated conditioning encoders, then jointly fuses them with text embeddings in the generation transformer. Enables fine-grained control over harmonic and rhythmic structure.

Enhances audio quality by applying diffusion-based decoding as a post-processing step after EnCodec reconstruction. MultiBand Diffusion operates on frequency bands independently, using a diffusion model to refine reconstructed audio and reduce compression artifacts. Can be used as a drop-in replacement for the standard EnCodec decoder or applied to any compressed audio.

Embeds imperceptible watermarks into generated audio using AudioSeal, a watermarking system that adds inaudible signals to audio while preserving quality. The watermark encodes metadata (e.g., generation timestamp, model version) and is designed to survive common audio transformations (compression, resampling, time-stretching). Enables detection and attribution of AI-generated audio.

Enables training of large audio generation models across multiple GPUs and nodes using Fully Sharded Data Parallel (FSDP) and gradient checkpointing. The framework automatically distributes model parameters, activations, and gradients across devices, reducing per-GPU memory requirements. Gradient checkpointing trades computation for memory by recomputing activations during backpropagation rather than storing them.

Generates audio in streaming fashion using a streaming transformer architecture that processes audio in chunks with limited context window, enabling generation of audio longer than typical 30-second limits. The model maintains a rolling cache of key-value pairs from previous chunks, allowing efficient incremental generation without reprocessing entire sequences.