Turn a hum into music with AI: what Sume takes as input

Can you turn a hummed tune into a finished track with Sume? Not directly. Sume's Music Router takes a text prompt of 1 to 5000 characters and one optional public HTTPS image_url. There is no audio input field, so you cannot upload a recording of yourself humming and have the engine follow it.

The idea is in the news because Stability AI's Sean Parker said, as reported by TechCrunch on October 2, 2026, that an upcoming update will let users hum a melody or beatbox a drum pattern to steer its music tools. That is described as upcoming, not shipped. Until a Sume engine offers audio input, the workable route is to write the melody down in words and notes and let the prompt carry it.

Which inputs does each tool take?

Sources differ on what each tool accepts. Sume's side comes from its docs, and Google's side from its Lyria page, which lists text and images and does not mention audio input.

How do I describe a hummed melody in a prompt?

A prompt cannot guarantee notes, but it can steer the shape of a tune. The Music docs say the axes in a brief are creative directions, not guaranteed output settings, so listen to what comes back. A hummed phrase has five things you can say in words:

• Contour: say where it rises, falls and rests, for example "opens on a held note, climbs a fourth, then steps down in three beats".
• Rhythm: count it. "Two short pickup notes, then a long one" says more than "catchy".
• Key and tempo as numbers: the docs' own example brief uses a BPM and a key such as D minor.
• Note names, if you know them: "D, G, F, E, D" is a clear hint, though not a guarantee that those pitches come back.
• Who plays it: pick one lead instrument so the line is not buried, for instance a muted trumpet or a Rhodes.

What does the request look like?

Here is a submit that uses all five. The length goes in the prompt, because Sume rejects duration fields.

curl -X POST https://api.sume.com/v1/music-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: hum-brief-001" \
  -d '{
    "prompt": "Gentle chamber-folk, 76 BPM, D major. The lead is a solo clarinet playing a simple tune: two short pickup notes, then a long held note, climbing a fourth, then stepping down D-G-F-E-D. Fingerpicked guitar underneath, brushed snare. The tune returns at 0:20 with strings doubling it. A 30-second track. Instrumental, no vocals."
  }'

How do I get the result and refine it?

Submit, poll GET /v1/jobs/{id}/status until it is terminal, then read GET /v1/jobs/{id}/result and take the artifact with type audio. The Jobs and results page gives the loop.

Expect to run it more than once. The docs describe no seed and no guidance control on the Music Router, so each submit is a fresh take, and each accepted generation is billed at the fixed Music price. Change one thing per attempt, usually the contour line, and keep the rest of the brief fixed so you can hear what moved.

What an image can and cannot do

An image is the only other input. If your tune goes with a place or a mood, pass a still as image_url and the engine conditions on it, as covered in music for a scene. It will not carry your melody, though. That is the gap a hum input would close.

If you already have the hummed recording and need it kept note for note, Sume cannot do that today. Use a notation or DAW tool to produce the line yourself, and bring Sume in later for a backing bed under it. When an engine on the Music Router adds audio input, the catalog at GET /v1/music-router/models is where a new id would show up.

How do I check the take against my hum?

Judge each take by ear against your hummed original, not by whether the prompt looks right. Three quick checks catch most misses: does the opening gesture match, does the tune come back where you asked, and is the lead instrument audible over the bed. If the tune is right but the arrangement is wrong, keep the melody lines of the prompt and rewrite only the instrument and production lines.

A prompt that names the hummed tune's source song will not help and may be rejected. The docs say that on a policy rejection you revise the flagged content while keeping the musical brief, so describe the shape of the tune and leave the title out.

What to watch

Check Stability's announcement again when it ships. This post reflects reporting dated October 2, 2026 and Sume's docs on the same read date.