Open Music Models Now Generate Sound From Text
Meta's AudioCraft family ships under an MIT license, with up to 3.3 billion parameters for music, extending its open-model strategy into audio.
Updated on Nov. 3, 2026 in Artificial Intelligence
Meta has open sourced AudioCraft, a family of three audio-generation models: MusicGen for music from text prompts, AudioGen for sound effects, and EnCodec as the audio codec underneath both. The models, weights and code are available under an MIT license, making this a shipped, publicly downloadable release rather than a research preview.
Why it matters
Generating audio directly from raw signals requires modeling extremely long sequences, which is why the team converts sound into discrete tokens first. Meta says open-sourcing the models could help the wider community address dataset bias, since the training data is heavily weighted toward Western-style music with English text.
The release spans model sizes from 285 million to 3.3 billion parameters across the two generators, with the improved EnCodec decoder reducing artifacts in music output. The stub provides no quality benchmark scores against prior systems, so the performance gains remain described qualitatively rather than measured.
The players
Meta
Meta is the social media and AI company behind the Llama 2 language model and the AudioCraft audio-generation family released as open source.
The details
AudioCraft works in two stages. First, the EnCodec neural audio codec, a compression model that turns raw sound into compact units, converts audio into discrete tokens, creating a fixed vocabulary of music samples. Second, autoregressive language models, which predict each next token from the ones before it, are trained over those tokens to generate new sequences, and the EnCodec decoder converts the output back into audio.
Timeline
Mid-July: Meta open-sources Llama 2, its large language model, for research and commercial use.
July 28: Meta reports more than 150,000 Llama 2 download requests in its first week.
The Tech Race
Meta's release follows its mid-July open-sourcing of Llama 2, applying the same open-weights strategy to generative audio. The stub does not name competing audio-generation products or research programs, so the comparison with rivals cannot be drawn from the source.
The models and code are available now under an MIT license, so developers and researchers can download and build on them without a subscription. The stub does not specify hardware requirements or a timeline for downstream products built on AudioCraft.
The takeaway
AudioCraft shows open-weight releases extending from text into music and sound, and the bias-reduction claim is worth watching once outside researchers test the models against more diverse datasets. Track whether independent evaluations confirm the quality improvements Meta attributes to the upgraded EnCodec decoder.
Further reading
For broader context on the open-model movement behind this release, explore Artificial Intelligence coverage on We are the Frontier.
Source note: This article includes information reported by AI Business.
