Lyria 3
Google's music and ambient-audio model family. Two variants share the same prompt surface but trade off length for runtime:
- Lyria 3 Clip — short clips, loops, and previews up to 30 seconds (~30s runtime).
- Lyria 3 Pro — full-length songs with verses, choruses, and bridges up to ~2 minutes (~1min runtime).
Both produce 48 kHz stereo audio.
Capabilities
| Feature | Lyria 3 Clip | Lyria 3 Pro |
|---|---|---|
| Text-to-Audio | Yes | Yes |
| Max Duration | 30 seconds | ~2 minutes |
| Output | 48 kHz stereo (MP3 / WAV) | 48 kHz stereo (MP3 / WAV) |
| Vocals + instruments | Yes | Yes |
| Reference audio | No | No |
| Negative prompt | No | No |
| Reference image | Yes | Yes |
When to use which
- Lyria 3 Clip — fast iteration on a single musical idea, SFX-adjacent loops, short cues for stings or transitions.
- Lyria 3 Pro — production-ready music with full song structure (intro → verse → chorus → bridge → outro), or longer ambient pads / soundscapes.
Prompt phrasing carries over between the two — start with Clip to lock the genre, instrumentation, and mood, then re-run on Pro for the full piece.
Prompting Tips
- Lead with genre and mood. "Cinematic orchestral, slow tempo, melancholy" beats "sad music".
- Name instruments explicitly. "Solo piano with light reverb and distant cello" gives the model anchors to work from.
- Describe structure for Pro. "Slow piano intro → vocal verse → string-driven chorus → return to piano" shapes the arrangement directly.
- Use the prompt to control vocals. Say "instrumental only" or "with vocal hook" — the model honors the cue.
Limitations
- No reference-audio input — drive the model with a prompt and an optional reference image
- No seed / determinism — each generation is fresh
- No negative prompt — describe what you want, not what to avoid
- Output is finalized as a single take; can't be re-rolled in place