google / lyria 3 pro

Overview

Lyria 3 Pro is an audio generation model from Google that creates full-length songs from a detailed music description. You guide the composition with section tags like [Verse] and [Chorus], plus timestamps for when parts should change. You can also add an optional reference image to influence the mood and tone. It’s useful when you need original music fast for demos, videos, and prototypes.

What you can build

• Background music for videos, trailers, and product demos
• Podcast intro and outro themes with a clear structure
• Game and app soundtracks for menus, cutscenes, and ambient scenes
• Song drafts with verses and choruses for songwriting exploration
• Multi-language songs by writing the prompt in the target language

Inputs

• A required song description written as plain text. Include genre, instruments, mood, tempo (BPM), key, and whether you want vocals.
• Optional lyrics written inside clear section blocks, such as [Verse] and [Chorus]. Add “Instrumental only, no vocals” if you want no singing.
• Optional timing control using timestamps like [0:00 - 0:20] to describe what should happen in each segment.
• Optional reference image to inspire the music’s feel. Upload a JPG, JPEG, PNG, or WebP file.

Outputs

• A generated song as an MP3 audio file.
• Stereo audio at 48 kHz.
• A single track that can be up to about 3 minutes long.
• Audio includes a SynthID watermark for AI-generated provenance.

Recommended settings

• Put the target length in the prompt, like “create a 2-minute song” or “about 180 seconds.”
• Use timestamps when you need clear transitions, like instrument entrances or a tempo shift.
• Use section tags for predictable structure, like [Intro], [Verse], [Chorus], and [Bridge].
• Keep lyrics and musical direction separated. Write lyrics in their own labeled blocks.

Limitations

• Exact duration isn’t guaranteed. The model tries to follow your requested length.
• Generation is single-shot. It doesn’t support step-by-step editing of an existing track.
• Results can vary between runs, even with the same description.
• Some prompts can be blocked by safety filters.
• Low-detail prompts often produce generic music. Give specific instruments and structure.
• Low-resolution, heavily compressed, or abstract images may have weak influence on the result.

Safety & compliance

• Don’t use prompts that request copying a specific artist’s signature sound.
• Follow applicable content, privacy, and intellectual property rules for your use case.
• Outputs are watermarked with SynthID to help identify AI-generated audio.
• Safety systems can block requests and filter outputs that violate usage policies.