Generate AI vocals that sing your lyrics
Lyro’s AI vocal generator sings the lyrics you give it, in 12 languages and in voice types such as soulful female, deep male, rap or duet. It can sing inside a full generated song, or over a beat you upload, in which case the vocal comes back as an isolated stem. Sung lyrics go to Google’s Lyria 3 Pro by default, which scored 4.67 out of 5 for Turkish pronunciation in our one small internal blind test.
How it goesThree steps
- Write or draft the lyricsBring your own words or let the lyric writer draft them. Mark the sections with tags such as [Verse] and [Chorus].
- Choose the voice and the languagePick a voice type and one of 12 languages. Lyro recommends Lyria 3 Pro for sung lyrics, and tells you where the other model is weak or untested in your language.
- Generate and keep the vocalGet a full song, or upload a beat first and get the vocal back as an isolated stem that opens in the mixer beside it.
The details
Four kinds of vocal you can get
- A sung song. Lyrics, voice and music generated together. This is the AI song generator.
- An isolated vocal over your beat. Lyro sings your lyrics at your beat’s tempo and key and returns only the vocal stem. See add vocals to a beat.
- An improvised topline. With no lyrics, the model listens to your beat and improvises a vocal melody over it.
- Vocal chops. Short, pitched vocal phrases used as a rhythmic hook, common in house and afro house.
How the model is chosen
The question Lyro asks is how clearly the words have to be understood. Where they matter, the request goes to Google’s Lyria 3 Pro. In our one internal blind test, on 21 September 2026, it scored 4.67 out of 5 for Turkish pronunciation, against 2.00 for ACE-Step 1.5 and 2.17 for MiniMax Music 2.6, which we then retired. Where the words do not need to be understood, as with vocal chops and improvised toplines, ACE-Step 1.5 does the job: it won the vocal-chop brief, and because it is priced by length, a short hook costs less on it.
The test was small and covered Turkish only: one rater, seven briefs, three models, names hidden. The other 11 languages are supported by the models, but we have not tested them, so judge them by ear. See how we test.
Where the limits are
- AI vocals can mispronounce words. Names, slang, abbreviations and rare words are the usual problems.
- You choose a voice type, not a specific singer, so the voice will not be identical from one song to the next.
- Lyro has no voice cloning, and you cannot ask for a real artist’s voice or name.
- You cannot dictate the melody note by note. You steer it with the lyrics, the style prompt and the voice type.
- Improvised toplines do not follow written lyrics.
- An isolated vocal stem comes from stem separation, so it can carry faint artefacts from the music it was sung over.
Formats, credits and rights
A sung song on Lyria 3 Pro costs 4 credits. Singing your lyrics over your own beat costs that generation plus the separation (one credit plus two per started minute), which comes to 11 credits for a three-minute beat. An improvised topline over your beat costs one credit per started minute, with a minimum of 2. Vocal chops run on ACE-Step 1.5, which is priced by length: one credit per 30 seconds with the detailed pass, or one credit per minute with the fast pass. Bringing your own beat needs the Creator plan.
Vocals made with Lyria 3 Pro carry Google’s SynthID watermark, and every download carries an ID3 tag declaring it AI-generated. Every paid plan includes a commercial licence. We do not promise copyright protection for AI-generated audio. See music rights.
Tips for clearer vocals
- Keep lines short and similar in length, so the phrasing stays even.
- Respell a word the way it sounds if the model keeps getting it wrong.
- Write numbers and abbreviations out as words.
- Keep the chorus identical each time it returns.
- Set the language yourself instead of leaving it to be detected, especially if you type without accents.
Questions
Can the AI vocal sing in my language?
Lyro supports English, Turkish, Spanish, Portuguese, French, German, Italian, Arabic, Hindi, Japanese, Korean and Russian. Pronunciation quality differs by language and model. Turkish is the only language we have blind-tested, and the model we route it to scored 4.67 out of 5 for pronunciation in that one small test.
Can I get the vocal without the music?
Yes. If you upload a beat, the vocal comes back as an isolated stem. If you generated a full song, the stem splitter can separate its vocal, at one credit plus two per started minute. Both need the Creator plan, and separated vocals can carry faint artefacts.
Can it sound like a specific famous singer?
No. Lyro has no voice cloning, and our rules do not allow naming a real artist as a style or imitating a real person’s voice without their documented consent. Spotify, for example, bans impersonation. Describe the voice instead: gender, tone, delivery and language. See the acceptable use policy.
Why did the vocal mispronounce a word?
Models guess pronunciation from spelling, and they guess wrong most often on names, slang, abbreviations and rare words. Respell the word the way it sounds, write numbers out in full, and generate again. A new generation uses credits, so fix all the problem words in one pass.
Can I use my own voice instead?
Yes. On the Creator plan you can upload a vocal or record one in the browser, tidy it with the vocal tools (presets, pitch shift and time-stretch), and have Lyro write a beat around it. See add a beat to your vocals.
Related: Add vocals to a beat, AI lyrics generator, AI song generator