Turn a description into a finished song

How the model is chosen

Lyro is multi-model. It does not push every request through one engine. The routing follows the one blind listening test we have run so far: one rater, seven Turkish-language briefs and three models, on 21 September 2026, with the model names hidden from the person scoring.

• Sung lyrics go to Google’s Lyria 3 Pro by default. In that test it scored 4.67 out of 5 for Turkish pronunciation and 4.57 overall.
• Vocal chops and tempo-locked beats go to ACE-Step 1.5, an open model under the Apache-2.0 licence. It scored 3.29 overall but won the vocal-chop brief, and it holds an exact tempo and key.

We have not tested the other languages, and we have not run a test for each genre. Outside sung Turkish the routing is a working rule, not a measurement.

Both models are reached through fal.ai. The model is shown on each track, and you can change the recommendation before you generate. The method is in How we test AI music models.

What you control

You control the brief, the words and the voice. You can leave everything else to Lyro.

• Genre and mood, or a free-text style prompt. The prompt generator and the prompt library give you wording that models follow.
• Lyrics, with section tags such as [Verse] and [Chorus], in any of 12 languages.
• The kind of vocal: a voice type such as soulful female, deep male, rap or duet, or vocal chops, or none.
• Tempo in BPM and musical key, if you want to set them.

Where the limits are

AI song generation is good, and it is not reliable in the way an instrument is.

• AI vocals can mispronounce words, especially names, slang and rare words. Respelling a word the way it sounds usually fixes it.
• The same prompt gives a different song every time. Plan on a few generations to get one you keep, and budget credits for that.
• The result is one mixed stereo file, not separate tracks. To rebalance it you need the stem splitter, which leaves some artefacts.
• Only the tempo-locked route guarantees an exact BPM and key.
• You cannot ask for a real artist’s voice or name. Describe the sound instead.

Formats, credits and rights

A Lyria 3 Pro song costs 4 credits, whatever its length. ACE-Step 1.5 is priced by length: one credit per 30 seconds with the detailed pass, or one credit per minute with the fast pass. Cover art is made automatically, at no charge, for any generation of 2 credits or more, and new art on request costs 1 credit. If a generation fails or times out, every credit for it comes back automatically, with a notice in the app and a refund line in your credit history. MP3 and FLAC downloads are included on every plan, WAV export comes with Creator, and downloads are never counted or capped.

Every paid plan includes a commercial licence. As between you and us, you own whatever rights exist in your tracks. We do not promise that copyright protects them: in the United States, works generated entirely by AI are not copyrightable. Every download carries an ID3 tag declaring it AI-generated, and Lyria output also carries Google’s SynthID watermark. See music rights.

Tips for a better first result

• Give the tempo as a number, such as “122 BPM”, not a word such as “mid-tempo”.
• Name instruments and their role: “rolling log drum bassline”, “warm electric piano chords”.
• Keep chorus lines short and identical on every repeat.
• Write numbers and abbreviations out as words, the way they should be sung.
• Change one thing at a time between generations, so you can hear what it did.