What is Lyria? Choosing songs, short clips and real-time music
Lyria is Google DeepMind’s music generation family. Learn how text or images guide music and how Lyria 3.5, short clips and real-time generation differ in lyrics, audio output and editing limits.
Meet this model
Lyria is Google DeepMind’s music generation family. Creative descriptions guide music, and some models also accept images to suggest mood. Lyria 3.5 can create songs with vocals and arrangements, while short clips and real-time music have separate options. Unlike text-to-speech, its main task is musical creation.
Song structure conveys more than a genre label
For travel footage, describe a light opening, increased rhythm on the mountain road and room for a fade-out. Lyria 3.5 accepts structures such as verses and choruses and can sing supplied lyrics. Separate lyrics from production instructions so directions are not sung into the track.
An image such as a night street can suggest mood, but does not specify every beat. For precise synchronization, describe important timed changes and listen to the track within the edit rather than review only its first few seconds.
Short clips and complete songs have different conditions
The current Gemini API guide distinguishes fixed 30-second Lyria 3 Clip output from fuller Lyria 3.5 songs. Prompts can influence 3.5 duration, but do not guarantee any exact second count. For background loops, check whether the ending joins naturally to the start.
Use the dedicated real-time route for interactive music. Playing a track in a webpage does not establish continuous interactive control. A generated file and an audio stream receiving ongoing controls require different integrations; follow the selected model’s documentation.
Check lyrics and channels in the output track
Lyria 3.5 documentation specifies 44.1 kHz stereo, default MP3 and supported WAV output. Sample rate describes the file, not lyric accuracy or mixing suitability. Compare each sung line with the intended lyrics, especially names and foreign-language words.
The model can vary arrangement and vocal delivery, but results can still depart from instructions. For narration backgrounds, play the music with the voice to check for masked words. This article generated no tracks and does not substitute official examples for loudness or listening tests.
The current version does not directly edit a track over multiple turns
The 3.5 guide describes single-turn generation and does not support refining a completed track through multiple prompts. Regenerate candidates or edit in audio software when changing a chorus. Do not assume image-style conversational editing precisely updates an existing audio file.
Generated audio includes SynthID, and requests for specific artists’ voices or copyrighted lyrics are filtered. Lyria models and app interfaces such as Gemini have separate access conditions. This article uses official model and API documentation and does not promise local weights or an open license.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.