audiocraft-audio-generation

Load a pretrained model with MusicGen.get_pretrained('facebook/musicgen-small'), set generation parameters like duration and temperature, then call model.generate with a list of text descriptions. Save the output tensor as a WAV file with torchaudio at 32kHz.

MusicGen generates music from text prompts and supports melody conditioning and stereo output at 32kHz. AudioGen generates sound effects and environmental audio from text at 16kHz. Both are part of the AudioCraft library.

Yes, MusicGen is available in Transformers as MusicgenForConditionalGeneration. Load it with AutoProcessor and generate audio using do_sample, guidance_scale, and max_new_tokens parameters instead of the native AudioCraft API.

MusicGen-small needs about 4GB VRAM in FP32 or 2GB in FP16, medium needs 8GB or 4GB, and large needs 16GB or 8GB. Reduce duration or use the small model if you hit CUDA out-of-memory errors.

Silent output usually requires normalizing the waveform amplitude. Poor quality improves with a larger model, higher cfg_coef for text adherence, lower temperature, and more specific prompts describing genre, instruments, and tempo.

MusicGen typically generates up to 30 seconds of audio per call, so it is not suited for full-length songs. For longer commercial music, alternatives like Stable Audio are recommended, and Bark is better for speech with music.