google / lyria 3

Overview

Lyria 3 by Google is an AI music generator that composes short, coherent tracks from your creative direction. It uses latent diffusion over compact audio representations to synthesize full-fidelity music, not loop stitching. You describe genre, instruments, tempo, mood, and vocals, and it returns a finished 30-second clip that you can drop into a project.

What you can build

• Background music for short videos, reels, and promos
• Mood sketches for a longer song you plan to produce later
• Instrumental beds for podcasts, ads, and narration
• Quick theme variations for games, apps, and UI moments
• Sonic branding drafts like stingers, bumpers, and outros
• Lyric-and-vocal ideas when you want a sung hook in 30 seconds

Inputs

• A detailed music request written in plain language. Include genre, era, mood, tempo (BPM), key, and instruments. Add vocal notes if you want singing.
• Optional short lyrics you want performed. Put "Lyrics:" before the lines you want sung. Keep lyrics brief for a 30-second clip.
• Optional instruction to avoid vocals. Say "Instrumental only, no vocals" when you want pure music.
• Optional inspiration image to steer the vibe. Upload 1 image as JPG, JPEG, PNG, or WebP. Use clear, high-contrast images for stronger style cues.

Outputs

• One downloadable MP3 music clip.
• Fixed duration: 30 seconds.
• Audio format: 48 kHz stereo.
• The clip can include vocals and performed lyrics if you ask for them.
• The audio includes a SynthID watermark to mark it as AI-generated.

Limitations

• Clip length is fixed at 30 seconds. It won’t generate longer songs in this variant.
• Generation is single-shot. You can’t edit specific bars or stems inside the same run.
• Results can change between runs, even with the same description.
• Image guidance is suggestive, not exact. It may not match specific colors or objects.
• Low-quality inputs reduce control. Blurry, tiny, heavily compressed, or noisy images can weaken steering.
• Some requests are blocked by safety filters, including disallowed or harmful content.

Safety & compliance

Google applies safety filtering to prompts and generations in supported products. Don’t request hateful, violent, sexually explicit, or otherwise harmful content. Don’t ask for misleading attribution, impersonation, or content that violates rights, including copyright and privacy. All generated audio includes SynthID watermarking for provenance and detection.