AI sings on your beat
You can make a beat but would rather not sing on it? Upload the beat, paste your lyrics and pick a voice. The AI listens to the beat, sings your words in time and in key, then mixes the vocal onto that exact beat. Pick My voice and the song comes out in your own timbre, even if singing is not your thing.
- Your beat stays untouched
- Male, female or your own voice
- Download the vocal on its own
Your beat
Drop your beat here
MP3, M4A, WAV, FLAC or OGG · up to 20 MB · 15 seconds to 5 minutes · instrumental only, no vocals
Leave empty and the machine finds where the intro ends. If your beat has a long intro or an unusual instrumental part, type the second the vocals should start.
Plain lyrics are fine: the machine treats repeated stanzas as the chorus and leaves room for instrumental breaks. To be sure, mark sections yourself with [Verse], [Chorus], [Bridge].
0/3000 charactersThe AI writes a new melody for your lyrics on this beat. To keep the exact melody of an existing song, use AI Cover instead: upload the original song and the machine keeps its melody, changing only the voice.
Voice
A female voice the AI creates as it sings. It is not the voice of any real person.
What makes the AI sing best
- Instrumental only. If the beat still has vocals or shouts, the AI sings on top of them. Beat has vocals? Use the Vocal Remover to pull out the music first.
- Lyrics that fit the beat. A three-minute beat usually holds two verses and two or three choruses. Keep each line about one musical phrase long.
- Mark sections if you want to be sure. [Verse] sits lower and tells the story, [Chorus] lifts and repeats. With plain lyrics the machine guesses: repeated stanzas are the chorus, and it leaves an instrumental break after the chorus.
- Pick the language of your lyrics. Lyrics in one language sung as another come out badly mispronounced.
- Run it a few times. Each run is a new take on the same beat and words. Keep the one you like best.
Three steps
- Upload your beat (music only, no vocals).
- Paste your lyrics, pick the language and the voice.
- Press Let the AI sing, wait a few minutes, listen to Full song / Beat / Vocal and download.
Want to mix it yourself? Download the vocal on its own and drop it into Auto Mix with your original beat: you can adjust reverb, EQ and vocal level like in a studio.
Frequently asked questions
- Does the AI sing in time with my beat?
- Yes. The machine measures the tempo and key of your beat, finds where the intro ends, then lets the AI listen to the beat as it sings: lines sit on the beat, the chorus lands where the beat builds up, and instrumental breaks are left open. It then filters the result so only the singing is kept, with no guitar or drums bleeding in, before mixing it onto your original beat. Your beat stays exactly as it is.
- I pasted the lyrics of an existing song onto its own beat. Why does the AI sing a different melody?
- Because this tool writes a new melody. A beat carries the tempo, key and chords but not the singer's melody, so the AI composes one that fits the beat for the lyrics you give it. To keep the original melody, use AI Cover: upload the original song with its vocals, and the machine keeps the melody and changes only the voice.
- I am a bad singer. Can I still use my own voice?
- Yes, that is what My voice is for. The AI sings first in its own voice, then converts the timbre to yours from a 20 to 40 second sample. The phrasing, timing and pitch come from the AI; the voice colour is yours.
- Is it free?
- Yes. No sign-up and no daily limit. Each person runs one song at a time: when it finishes you can start the next one right away. Each song runs on a graphics card for a few minutes, so sometimes there is a queue.
- Do you keep my beat and lyrics?
- The beat and voice sample are deleted as soon as the job finishes, and the song deletes itself after 24 hours. If you use your own voice, the voice model learned from your sample is kept for 24 hours after its last use, then deleted. No public page links to your song, and we block search engines from indexing it.
- Who owns the finished song?
- The beat and lyrics are yours. The AI vocal is not the voice of any singer, and you can use it in your own song. If you plan to release commercially, check what your distribution platform says about music made with AI.
- Can I pick a famous singer’s voice?
- No. There are only AI voices that belong to no one, or your own voice. Using someone else’s voice without their consent violates their rights to their voice.
What this tool does
You give it an instrumental beat and your lyrics, then pick an AI male voice, an AI female voice or My voice. The singing model listens to the beat, writes a vocal melody for your words and sings it in the tempo and key of that beat. The vocal is then mixed onto your original beat and you get three files: the finished song, the beat, and the vocal on its own. The music is not regenerated: the beat you upload is the beat you hear in the song.
It suits producers who make beats but do not want to sing, and writers who have lyrics and want a demo to hear how the song could sound. No beat yet? The song generator builds a whole track from scratch.
Five steps before the song exists
- Tempo and vocal entry. The tool measures BPM, finds bar lines and looks for the end of the intro. The main rule: never bring the voice in where the music gets louder, because that is usually an instrumental hook, not the first verse. You can also type the entry yourself in the Vocal entry box, for example 0:32.
- Key. Three different key detectors vote. The key is passed to the singing model only when enough of them agree; otherwise the model listens to the beat itself instead of being told a key that may be wrong.
- Lyric structure. Plain pasted lyrics are labelled automatically: the most repeated block of lines becomes the chorus, the rest becomes verses, and a break is left for the instrumental after a chorus. A stanza made only of vocalising such as hu hu hoh becomes an instrumental interlude. Labels you wrote yourself are kept.
- Sing, clean, score. The model sings over the beat trimmed from the entry point, and the vocal is placed back at the right second. The vocal track goes through a vocal separator to remove guitar and drum bleed. Every take is scored; a take that comes in late is sung again, at most twice, and the best one is delivered.
- Mix. The vocal sits 2.2 LU below the beat, the median measured on a public stem dataset of 144 songs, then the song is mastered to about −14 LUFS with a −1.3 dBTP true peak.
What we measured
Each step has a test bench where the right answer is known in advance. The main numbers:
| Test | Result |
|---|---|
| Vocal entry, 57 songs with known answers | Singing from second 0 covers the intro in 79% of songs. Current detector: within ±1 bar 64.9%, median error 0.5 bar, intro covered 1.8% |
| Key detection, 48 songs with known keys | 46/48 correct (previous version 40/48) |
| Lyric structure, 300 labelled songs | Structure found 96%, chorus lines identified 87%, verse not mistaken for chorus 92% |
| Rhythm match (1 equals the beat's original singer) | Pop 99 BPM +0.89; lo-fi 144 BPM +0.66. A vocal from a different song on the same beat scores only +0.06 to +0.23 |
| Vietnamese lyrics heard back, with tone marks, per line | Pop 0.938; ballad 0.860; lo-fi 0.688 (first two lines lost) |
| Beat bleed in the vocal track | +0.736 before cleaning, +0.257 after; real singers median +0.254 |
| Late entry, a user's song pasted as 48 unbroken lines | 16.1 seconds before the fix; 1.1 and 0.0 seconds on two reruns |
One more measure is sustain: the share of voiced time spent on notes held steady for 0.6 seconds or longer. On real vocals, rap sits at 0.11 to 0.19 and ballads at 0.22 to 0.43. A ballad below 0.2 sounds closer to speaking than singing.
How long it takes
For a 75 second beat, measured through the live site at launch: 78 seconds with an AI voice and 173 seconds with My voice. After the cleaning and scoring steps were added, a 100 second beat with 12 lines of lyrics and an AI female voice took 219 seconds end to end. Queue time on busy hours comes on top. My voice takes longer because the AI sings first and the voice changer then moves the timbre to yours; the first time a voice sample is used, the tool also has to learn that voice.
Limitations
- The melody is new. A beat, even a karaoke track, does not contain the singer's melody. The singing model writes a new tune for your lyrics, so old lyrics will not come back on their old melody. To keep the melody of a recording you have the right to use, use AI cover.
- Ballads sometimes hold notes too little. Sung as one full take, a ballad can sustain only about 0.18, which is rap territory. Singing each section of about 100 seconds separately raises the average to 0.30, but two separately generated sections sound like two different singers, so this is not switched on yet.
- EDM beats with drops. The tool stays silent through the intro but may still sing over drops and breaks.
- Key names can be one step off. A song written in D minor was once reported as F major. The two keys share the same notes, so the singing is not affected, but the name on the result page differs from the original score.
- Slow ballads may land off the grid. A 71 BPM ballad scored −0.36 on the beat and +0.63 when shifted by a sixteenth. Whether that phrasing sounds good is still open.
- Every run sings differently. Same beat, same lyrics, a new performance each time.
Frequently asked questions
My beat still has vocals or shouts in it. What happens?
The AI will sing on top of them. Use Stem splitter to pull out the instrumental first and upload that.
The vocal came in at the wrong place. How do I fix it?
The result page shows the second the vocal starts and whether the tool detected it or you chose it. If it is wrong, run again and type the start time in the Vocal entry box, for example 32 or 0:32. Long intros and unusual instrumental hooks are where detection fails most often.
Do I need to add [Verse] and [Chorus] labels?
No. Before automatic labelling, a song pasted as 48 unbroken lines became a single verse, came in 16 seconds late and sounded spoken. The tool now finds repeated stanzas itself, and on those same lyrics it produced exactly the labels the writer had added by hand. If you want certainty, label anyway; your labels are kept.
Why does the AI repeat my lyrics?
When the lyrics are shorter than the beat, the singing model repeats them to fill the remaining music. A three minute beat usually needs two verses and two or three choruses. Write more lines or trim the beat to avoid it.
Related research
No beat yet?
Make a whole song with AI in a few minutes, or pull the beat out of an existing song.