AI Song from a Hum

Your recording, and the worked example

The seed is your voice: a voice message, a hum into the phone, a line you sang or rapped — a few seconds is enough. If you have no recording, an idea works too (a song from an idea). Everything below is a method with slots: your words, your melody, your vibe.

Example run — the owner’s voice message. An 8.3-second voice message in Russian: a laid-back rap line about flying above the ground. We took it through every step on 9 October 2026 — first by hand through the kie.ai API, then in a source chat prototype, where every creative choice was a button.

Two listeners, two jobs. ElevenLabs Scribe v2 writes down the words with a time for each — it is the one to trust for words. Gemini 3.8 Flash (through kie.ai) listens to the whole thing and says what it is: rap, singing or humming, the tempo, the key, the delivery, three genres it fits, the mood and what the words are about. It also writes the words — and where the two listeners disagree, that is the word to ask about.

• Three of four lines agreed; one word did not (3.6–4.0 s): Scribe heard «блики, бе», Gemini «блинги, бэ», and Scribe v1 on the same file a third version. The person wrote their own reading on the card — that spelling went into the lyrics.
• About 15 seconds for the transcription and sound description together.

“Cloud rap” means little until you hear who it is. So every vibe comes with three artists you know — taken from the MusicBrainz database (the full dump of 7 October 2026: 223 thousand artists with genres or tag votes), ranked by how known they are, not by how often a tag was clicked. The same data gives the style line for Suno: the genre and its neighbours from the genre list, never an artist’s name — Suno doesn’t take names in the style, and the sound is what you want anyway.

Above: the vibes Yaro proposed for an idea with no recording — “a song about night-time Moscow in the rain, sad but beautiful”. Genre names from a model are matched to MusicBrainz (“Russian Trap” → trap). If the person types an artist instead, the style comes from that artist’s genres and tags.

The seed’s words are the chorus, word for word. The rest is written for this song, in the language of the seed — two verses with concrete pictures from its theme, a bridge, an outro, structure tags on their own lines. Never the words of an existing song. The style line starts from the picked vibe and adds the voice, the tempo from listen, two to four instruments and the mood.

• A bare [Chorus] tag is sung as the chorus again. The draft had the chorus written once and two empty tags; Scribe on the finished vocal stem found it sung four times. Write a repeat out only when it must change.

The song is a cover of the seed: Suno V6 takes the melody and the flow from your recording and builds a song on the lyrics and the style. audio_weight sets how close it stays — 0.65 by default, 0.75 to stay closer to the hum, 0.5 to rework it. Every request returns two takes.

• Both takes took 121 seconds — the same job took 56–67 s of generation earlier in the day; plan for up to three minutes.
• The hook is the seed’s own words: Scribe on the take’s vocal stem finds «я лечу над землёй» six times and the seed’s first line in all four choruses.

Pick a take, then what to do with it. Stems split it into the vocal and the instrumental — that is the default; up to twelve stems or one instrument are there on request. Longer extends the take from a second you choose with new lines. Suno also draws a cover picture for every take.

• Stems took about a minute. Extension took about a minute; the take continued from 50 s to 1:50.
• Extend needs the take’s Suno id, so it works on takes made in the same place; for any other file make a new cover.

No voice to start from? Say what the song is about. The vibe card comes from MusicBrainz the same way, the lyrics are written for the idea, and Suno makes a new song with no source. A 30-second test: “a song about a dog that steals socks”.

In the source chat prototype, a voice message led to cards for the uncertain word, vibe, lyrics, “Make it?”, which take and what next. The user could type their own answer at each step. This illustrates the sequence of creative choices, not an installed or integrated FlashBoards feature.

• Your voice is the best prompt for a melody. A cover of your recording keeps the tune and the flow; a text prompt can’t.
• Always give it lyrics — your line as the hook, the rest written for the song; without them you get an instrumental.
• Two listeners for the words. Scribe for the text, Gemini for the sound; ask the person where they differ.
• Pick the vibe by ear, with names you know — and keep the names out of the style line.
• Only your own audio as the source. A hit’s recording is refused; a new song in that style is not.
• Two takes — make another round rather than tweak; stems and a longer version are about a minute.