MusicGen is a model developed by META and trained with 20,000 hours of music. Like all artificial intelligence models, it suffers from limitations and biases related to training data.
Try this
- Don't go overboard with customisation
- ✓ “A cheerful country-style Christmas song with guitar”
- ✕ “A cheerful, romantic Christmas song in country and metal style with drums and guitar”
- Try to be consistent
- ✓ “A festive 50s style Christmas song with violins”
- ✕ “A romantic Irish folk style Christmas song with castanets”
- Some cultures and musical styles are not well represented in the model, if the result does not convince you try changing the style
- Do not indicate voice-related customisations such as "choirs", "a cappella", "vocalisation"
- Do not indicate customisations linked to specific artists
- ✓ “An enchanting Christmas song, R&B style with trumpet”
- ✕ “A beatlesque Christmas song, Måneskin style with Solieri-like chords”
Limitations
- The model is unable to generate realistic voices
- The model does not work equally well for all musical styles and cultures
- The model sometimes generates the end of songs, which collapse into silence