Verified Agent Skill record

Audiocraft Audio Generation

Guide to Meta's AudioCraft PyTorch library for audio generation, covering MusicGen for text-to-music (including melody-conditioned, stereo, and style-transfer variants) and AudioGen for text-to-sound generation, plus EnCodec for neural audio compression.

Tasks

Generate music from text descriptions using MusicGen (facebook/musicgen-small/medium/large)
Create sound effects from text prompts using AudioGen (facebook/audiogen-medium)
Perform melody-conditioned music generation with generate_with_chroma on musicgen-melody
Generate stereo audio with musicgen-stereo model variants
Apply reference-based style transfer with musicgen-style and generate_with_style
Compress and reconstruct audio with EnCodec via CompressionModel

Inputs

Text descriptions/prompts for music or sound generation
Melody audio files (e.g., melody.wav) for melody conditioning
Style reference audio files (e.g., reference_style.wav) for style conditioning
Existing audio files (e.g., intro.wav, audio.wav) for continuation or EnCodec encoding
Generation parameters: duration, top_k, top_p, temperature, cfg_coef

Outputs

Waveform tensors shaped [batch, channels, samples]
Saved WAV files (e.g., output.wav, music_0.wav, sound_effects/*.wav) at 32000 Hz for MusicGen or 16000 Hz for AudioGen
EnCodec audio codes and reconstructed audio

Limitations and checks

Documented duration range is 1-120 seconds, with examples using up to 30 seconds
The guide suggests alternatives (Stable Audio, Bark, Riffusion, OpenAI Jukebox) for longer commercial music, TTS, spectrogram-based, and lyrics-based generation
Reference files mentioned in the guide (references/advanced-usage.md, references/troubleshooting.md) are not included in this document
Confirm the correct model variant is selected (e.g., musicgen-stereo-* for stereo output, musicgen-melody for melody conditioning) since features depend on the variant
Verify output sample rate matches the model: 32000 Hz for MusicGen, 16000 Hz for AudioGen, per the documented examples
Monitor VRAM against the documented table (musicgen-small ~4GB FP32, medium ~8GB, large ~16GB) and reduce model size or duration on CUDA OOM
Resample input audio to 32 kHz before EnCodec encoding, as shown in the guide
Set generation parameters (duration, temperature, cfg_coef) before calling generate, as set_generation_params is separate from generation