Google's music model for full songs with vocals and your own lyrics, or instrumentals, up to about three minutes.
Lyria 3.5 is Google's current music generation model, built by Google DeepMind. It writes a complete piece of music from a description: a genre, a mood, the instruments, a tempo and, when you want one, a singer. It is the model on this platform for anything with a voice in it, because it sings, and it sings the lyrics you write rather than lyrics it invents.
Lyrics take section tags such as [Verse], [Chorus] and [Bridge], and that is how you shape the song: the tags tell the model where the hook lands and when it comes back. It is the right choice for jingles that repeat a brand line, launch anthems, channel themes and any track where a sung hook does the selling. Ask for instrumental and it makes themes and beds as well.
Length is steered by the prompt rather than set to the second, roughly up to three minutes. Output is 44.1 kHz stereo, and every track carries Google's inaudible SynthID watermark, so it can always be identified as AI-generated. Google's filters refuse requests for a specific artist's voice or for copyrighted lyrics, which is why the reliable route to a sound you already like is to read a reference track with Audio to Prompt and make music from its description.
Make it with AI Music Generator, which runs Lyria 3.5 and shows the price before you spend anything. To start from a track, a voiceover or a sound that already works, read it with Audio to Prompt first: it describes the sound without naming anyone and hands that description to the generator.
An upbeat pop jingle for a bike shop called Spoke, bright female vocal, handclaps. Lyrics: [Verse] Flat tyre, Monday, we have got you covered [Chorus] Roll on in, roll on out, Spoke on Market Street
Section tags tell Lyria where the hook goes and when to repeat it. Keep the chorus short and put the brand name in it: that is the line people remember.
Late 1970s disco: four-on-the-floor kick, octave bass, bright strings, wah guitar, a joyful female lead vocal, 118 BPM, about 60 seconds.
Naming a singer or a band gets refused. Era, instruments, tempo and vocal tone get you the sound without asking for anyone in particular.
Read the reference in Audio to Prompt, then use its description here: genre, instruments, tempo, key and structure, with your own lyrics added.
A reference that already performs is the fastest brief there is. Audio to Prompt describes its qualities without naming anyone, which is exactly the kind of prompt Lyria accepts.
Cost on Prompt Reverse, per take: 10 AI Credits, two takes in one run: 20 AI Credits. AI Credits come from a one-time credit pack and never expire, with no subscription and no separate provider account.
Lyria 3.5 is Google's music generation model, from Google DeepMind. It makes full songs with vocals and your own lyrics, or instrumentals, up to about three minutes long, in 44.1 kHz stereo.
Yes. Write them with section tags like [Verse] and [Chorus] and Lyria sings them. Choose instrumental instead for a track with no voice.
Yes, in ads, videos, podcasts and social posts, online and offline. Every file is tagged as AI-generated in its metadata and carries Google's inaudible SynthID watermark. It is not cleared for film, TV, radio broadcast or games distribution.
No. Google's filters refuse requests for a named artist's voice or for copyrighted lyrics. Describe the era, genre, instruments and vocal tone instead, or start from a reference track with Audio to Prompt, which reads its qualities without naming anyone.
Roughly up to three minutes. Length is steered by the prompt rather than set to the second, so for a bed that has to match an edit exactly, use Stable Audio 3.
On Prompt Reverse it runs on never-expiring AI Credits from a one-time credit pack, with no subscription. The price of each take is shown before you run it, and a take that fails is refunded.