AudioCraft

AudioCraft is a PyTorch-based research library released by Meta AI on August 2, 2023. It is not a hosted product or consumer app but a self-hosted framework that bundles three distinct generative audio models: MusicGen, which generates instrumental music from text prompts or melody references; AudioGen, which generates environmental sounds and sound effects from text; and EnCodec, a neural audio codec that underpins both models as a high-fidelity compressor and tokenizer. The code is MIT-licensed and the pre-trained model weights are released under CC-BY-NC 4.0, meaning non-commercial use only unless a separate commercial agreement is reached with Meta.

MusicGen ships in small (~300M parameters), medium (~1.5B), and large (~3.3B) variants. Its melodic conditioning feature accepts a reference audio file, extracts a chromagram representation using Meta's HT-Demucs source separation, and uses that harmonic fingerprint to guide the arrangement, letting producers transform a hummed melody or piano sketch into a full instrumental track. AudioGen handles ambient beds and foley-style sound design at 16kHz. Two newer additions expand the framework: MAGNeT, a non-autoregressive model added in May 2024 offering faster inference, and JASCO, added in January 2025, which introduces fine-grained chord and beat controls via Flow Matching. All four models run locally on GPU hardware.

What AudioCraft produces in April 2026

The core MusicGen models generate up to 30 seconds of instrumental audio per pass at 32kHz mono, or stereo with the stereo variants released in January 2024. The medium and large models produce noticeably more coherent musical structure than the small model, but both require at least 16GB of GPU VRAM for comfortable inference. The small model runs on 8GB cards with quality trade-offs. AudioGen generates mono 16kHz sound effects, and its outputs are more reliable for short ambient clips than for precise foley timing.

JASCO, the newest model, operates in 400M and 1B parameter variants with two conditioning modes: text plus chords and drums, or text plus chords, drums, and melody. The 1B variant with full conditioning is the most controllable music generation model in the framework. It is also the least documented and, like all AudioCraft weights, non-commercial only. The metademolab.com demo pages provide Gradio interfaces for testing MusicGen and AudioGen without local installation, though the public demos cap generation at 12 seconds.

EnCodec, independent of the generation models, is a fully real-time neural codec supporting mono and stereo audio at multiple bitrates. It has seen adoption outside AudioCraft, including in voice applications and speech research, and is arguably the most practically useful component of the library for production engineers who need high-quality audio compression.

Where AudioCraft sits versus Stable Audio Open and Suno

The two most relevant comparisons are Stable Audio Open (open weights, active development) and Suno (closed, commercial, vocal-capable).

Stable Audio Open (Stability AI, July 2024) uses a latent diffusion architecture rather than an autoregressive transformer. Its core model is a 1.057B-parameter diffusion transformer (DiT) conditioned by a 109M-parameter T5 text encoder, plus a 156M-parameter autoencoder that compresses stereo 44.1kHz waveforms into latent space. Output runs up to 47 seconds of stereo audio at 44.1kHz, compared to AudioCraft's 32kHz mono ceiling. Stable Audio Open was trained on 7,300 hours of Creative Commons audio, which gives it better coverage of environmental and percussive sounds. AudioCraft's MusicGen trained on 20,000 hours of licensed and Meta-owned music, which gives it an edge for melodic and harmonic musical structure. Practically: Stable Audio Open performs better for sound design and short ambient clips; MusicGen performs better for longer instrumental compositions with musical coherence. Stable Audio Open has also received more active maintenance in 2024-2025, including a 341M-parameter mobile variant released in partnership with Arm.

Suno (commercial, closed) takes a completely different approach. Its hybrid architecture combines Bark, which handles realistic vocal synthesis and lyric generation, with Chirp, which handles instrumentation. The output is a full produced song with lead vocals, harmonies, and music bed, up to 4 minutes per clip. AudioCraft cannot generate vocals at all. Meta deliberately stripped vocals from MusicGen's training data using HT-Demucs source separation precisely to avoid copyright issues with singing. Suno v5 (2025) posts an ELO score of 1,293 in independent audio quality evaluations, placing it ahead of every publicly tested competitor. However, in July 2024 the RIAA filed lawsuits against Suno and Udio alleging unlicensed copyrighted music in training, a legal shadow that AudioCraft does not share given Meta's use of licensed data. Suno charges via subscription tiers starting at approximately $8/month; AudioCraft is free.

"couldn't sell this output for a free mobile clicker game", anonymous commenter, Hacker News, August 2023, discussing AudioCraft's launch demo audio

That reaction, while harsh, illustrates what AudioCraft is and isn't. At launch, the outputs were research-grade, not production-grade. The gap between AudioCraft and commercial music generators has widened since 2023 as Suno and Udio iterated aggressively while Meta's core MusicGen models have remained largely static.

The licensing and copyright reality

AudioCraft's code is MIT-licensed, which sounds maximally open. The model weights, however, are under Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0). That distinction matters significantly. You can freely read, modify, and redistribute the training and inference code. You cannot, under the weight license, use the generated audio or the models in any commercial product, service, or monetized workflow without a separate commercial agreement with Meta, which is not publicly offered.

Multiple developers at the August 2023 launch discussion on Hacker News flagged this framing: one noted bluntly that "you can't build a business on it" due to the noncommercial restriction. The gap between "open source" (code) and "open weights" (non-commercial) is a pattern that also appears in Meta's Llama releases, and it creates real friction for anyone hoping to build a commercial product on top of AudioCraft's outputs.

On the training data side, Meta stated that MusicGen was trained on roughly 400,000 recordings totaling 20,000 hours, using internal Meta-owned music and specifically licensed material from Shutterstock and Pond5. This positions AudioCraft more defensibly than Suno or Udio, both of which faced RIAA lawsuits in July 2024. AudioCraft's training data claims have not been legally challenged as of April 2026.

JASCO's weights carry the same CC-BY-NC 4.0 restriction as MusicGen and AudioGen. EnCodec's weights and code are both MIT-licensed, making it the only fully open component of the library.

Where AudioCraft reliably falls short

The clearest limitation is vocals. MusicGen cannot produce singing, speech, or any intelligible human voice in its output. This is a deliberate design choice, not a temporary gap. The training pipeline ran all audio through HT-Demucs to separate and remove the vocal stem before tokenization. Researchers who need vocal synthesis from the AudioCraft family are pointed toward EnCodec as a codec layer for other vocoder systems, not MusicGen itself.

Hardware requirements create a second barrier. The medium model, which is the minimum configuration for musically coherent output, requires 16GB of VRAM. Users with 8GB cards are limited to the small model, which produces noticeably simpler musical structure. A user on GitHub issue #435 reported in March 2024 that their 16GB VRAM NVIDIA laptop was showing only 5-10% GPU utilization during MusicGen inference while CPU cores ran at capacity, suggesting the inference pipeline does not consistently leverage GPU acceleration on all hardware configurations.

Fine-tuning on custom datasets is technically supported but practically difficult. A user who attempted to fine-tune a 33M-parameter custom model on 133 hours of music on a GTX 1060 reported after 20 hours of training that the generated samples "sounds like it's doing almost nothing," comparing unfavorably to other transformer architectures that show coherent progress within 20 minutes on equivalent hardware. Training requires substantial GPU resources that put serious fine-tuning out of reach for most individual researchers without cloud compute access.

Development velocity has also visibly slowed. The last substantive release was JASCO in January 2025. Before that, MAGNeT in May 2024. The core MusicGen and AudioGen models are effectively frozen at their 2023-2024 parameter counts while competitors like Suno and Stable Audio continue to release major version updates with quality improvements. With 362 open GitHub issues and Meta's research attention clearly shifting toward other projects, AudioCraft is in a maintenance state rather than active development.

"the model progresses absurdly slow, i've been training for the past 20 hours, and generated samples sounds like it's doing almost nothing". CDandRW, GitHub issue #388, January 15, 2024

Who AudioCraft is for

AudioCraft is the right tool for machine learning researchers and audio engineers who want to study or build on top of Meta's generative audio architecture. If you are writing a paper on text-conditioned music generation, need a baseline model for comparison, or want to train a custom model on a niche dataset (folk instruments, industrial sound design, game audio packs), AudioCraft's training code and EnCodec backbone are well-documented and genuinely useful. It is also the right tool for developers who need non-commercial generative audio inference on-premises with no usage cap, no API costs, and full control over the model.

AudioCraft is the wrong choice if you want to generate complete songs with lyrics and vocals, which requires Suno or Udio. It is also the wrong choice if you need fast, well-maintained open-source audio generation for non-music sound design, where Stable Audio Open offers more active development and better stereo quality at 44.1kHz. Skip AudioCraft if you need commercial licensing, since the CC-BY-NC weight restriction blocks any monetized deployment. And skip it if you lack a GPU with at least 8GB of VRAM or the comfort level to install Python dependencies and manage a PyTorch environment manually.

For indie game developers, student filmmakers, and academic audio researchers who are comfortable running local Python code and have a mid-range NVIDIA GPU, AudioCraft remains a capable and cost-free option for generating non-commercial instrumental music and sound effects. Its position as the foundational open-source audio AI framework that preceded everything else gives it continued citation value in the research literature even as newer tools surpass it on raw output quality.