Vocals on Your Own Beat: What the Engine Keeps and What It Re-records

The question producers actually ask

“Add vocals” sounds like an overdub: your beat, plus a singer. That is how it looks in the result, and for many uses it is all you need. But a producer who has spent a week on a drum sound, or an agency working from a client's cleared instrumental, needs a stronger promise: that the file under the vocal is the file they uploaded, sample for sample.

We did not want to answer that from the product description. We wanted to answer it from the audio.

The beat and the lyric

The beat is Rainy Desk, a 35-second lo-fi hip hop instrumental we made on Tunely as an instrumental (no vocals), which makes it a fair stand-in for a beat you would make in your own software. The lyric is eight lines we wrote for the test, and the voice brief was “lo-fi hip hop, soft intimate female vocal, close mic, gentle”. We ran Add vocals once, which returns two takes, then ran Get stems on take 1 to separate the vocal from the backing.

[Verse]
Rain on the window, coffee going cold
Every little plan I made is waiting to be told
[Chorus]
Stay a while, stay a while
Let the quiet find its way
Stay a while, stay a while
We can start again today

How we checked whether the beat survived

Two measurements, made with librosa on the exact files above, plus a control (an unrelated Tunely song) so we know what “no relation” scores.

• Waveform match. Line the two recordings up as well as possible (allowing up to three seconds of offset) and measure how well the actual audio agrees. 1.0 means the same audio; around 0 means unrelated. This is the decisive test: only the original file scores near 1.0.
• Chord-map distance. The same beat-by-beat pitch-class comparison we used in our language-version test. 0 is identical; higher is further away. It tells us whether the musical content is similar even when the audio is not.
• Why the instrumental stem matters most. Comparing the original beat to the full result is unfair: the vocal is in the way. So we separated the result into stems and compared the instrumental stem to the original beat. If the engine had kept our file, that stem would match it almost perfectly.

The numbers

Every row is a real file from the test.

Reading the table

• The beat was re-performed, not kept. The instrumental stem scored 0.11 against our original; the unrelated control scored −0.01. Both are “not the same audio”. Our drum hits, our exact sounds, are not in the result.
• The tempo and length were kept. 34.3 to 34.6 seconds against 35.2, and the same pulse. The takes started 0.8 to 2.9 seconds off our original's downbeat, which matters for the two-step method below.
• The musical content is close. 0.10 to 0.13 on the chord map, nearer than the control's 0.17. The gap is smaller than in our language test because a 35-second loop gives the measurement less to lock onto, but the feel, the chord movement and the lo-fi character are recognisably Rainy Desk.
• The vocal stem is clean. It scores like the control against the beat (−0.01), which is what you want from a separation: no beat left in it. That is the file you take into your DAW.

The two-step way to keep your exact beat

When the beat itself must not change, use Add vocals as a vocalist, not as a mixer:

• 1. Add vocals. From an instrumental in your library (song menu → Remix → Add vocals) or by uploading your beat in Remix Studio. Paste the lyric, describe the voice. 1 credit, two takes back.
• 2. Get stems on the take you like. Song menu → Remix → Get stems. 1 credit. You get the vocal on its own and the re-performed backing; keep the vocal.
• 3. Import your original beat and the vocal stem into your DAW. Both are at the same tempo, so once the first word sits on the right bar the rest stays aligned. Expect to nudge the vocal by one to three seconds; that is the offset we measured.
• 4. Mix. The stem is fairly dry. Add your own reverb, compression and level so it sits in your beat rather than on top of it.

Two credits in total, and your beat is untouched. Because the engine matched our tempo exactly in every take, we never had to time-stretch the vocal; the whole alignment was a nudge.

Exact-beat checklist: same tempo confirmed → vocal stem separated → first word on the bar → your original beat under it. Keep the re-performed backing only as a reference for how the vocal was meant to sit.

When the re-performed backing is exactly what you want

Plenty of the time, the one-step result is the right result. If the beat was itself made on Tunely, the re-performance is a sibling of the original and nobody is losing a bespoke drum sound. If you need a demo to pitch a topline, or you are going to re-produce the track around the vocal anyway, the finished mix that comes back is quicker than the DAW route and already sounds like a record. It also comes with its own commercial license certificate on Pro and Ultra, which the DAW route cannot give you for your own beat because your beat is yours already.

Rights, in one paragraph

Your instrumental stays yours; uploading it changes nothing about who owns it. On Pro and Ultra, the generated vocal and the mix that comes back are licensed to you for commercial use, with a certificate per result. The engine cannot tell whether you have the rights to a beat you upload, so that responsibility stays with you: a beat you bought under a lease, for instance, usually caps streams or requires credit, and adding a vocal does not change the terms of that lease. Details are in Can I use AI-generated songs commercially?.