Split a Song Into Vocals, Drums and Bass With AI

Say you want to rebuild a track with new drums, learn a bass line by ear, or lower a singer under your voiceover. All three need the mix pulled apart. On Kadraris that job belongs to Multistem Splitter: you send a single audio file, and one task gives you three files back, the vocal, the drums and the bass. The price is $1.50 per minute of audio for the whole set, counted per started minute, so a four-minute song costs 4 × $1.50 per minute.

Is three stems what you need?

Pick the tool by what happens to the files next:

Compare the two routes on a four-minute song. Stems: 4 × $1.50 per minute. Instrumental plus a cappella from the two single tools: 4 × $0.50 per minute + 4 × $0.50 per minute. When you do not need drums and bass as separate files, the simpler pair is cheaper. That route is explained in how to remove vocals from a song, and the two tools are compared in Multistem Splitter vs Vocal Remover.

The three files and what is missing

The set on Kadraris is fixed:

• vocals: the lead and any backing voices in one file,
• drums: the kit plus percussion,
• bass: the bass part.

The underlying engine can separate up to six stems, adding piano, electric and acoustic guitar. Kadraris does not yet let you pick the stem list, so every job returns the default trio. Extraction strength, engine choice and dereverb are not available either.

Plan around that. Guitars, keys, strings and pads do not get files of their own; they are simply not in the output. If you need "everything except the singer" as one file, Music Extractor is the right tool, since it returns the complete instrumental.

What people use the stems for

• Reworks. Keep the vocal, write new drums and bass. Without the other instruments as separate files, this suits a fresh production more than a faithful remix.
• Practice. A bassist transcribes the isolated line; a drummer studies the groove without the band.
• Narration over music. A sung vocal competes with speech. With stems you can dip or mute the singer and keep the rhythm going.
• Sampling your own material. A drum loop or bass riff from your recordings becomes a building block.
• Mixing an AI song yourself. Create a track in Mureka V9, split it, and balance it in your DAW. Generation is covered in the Mureka V9 guide.

How billing works, minute by minute

$1.50 per minute already covers all three stems. The vendor counts each minute once for every stem it returns, and the Kadraris price is that total, not a per-stem figure. Minutes round up:

Rows one and two show how a few seconds move you into the next minute. Cut silence and long intros before uploading. You pay from a prepaid USD balance with no subscription and no expiry, and failed jobs are free.

Running a split in Studio

1. Create an account and put at least $10 on the balance.
2. Go to the Audio tab and select Multistem Splitter.
3. In Source audio (URL) either press Upload file (MP3 or WAV, 25 MB max) or paste a public link.
4. Read the Cost line and click Generate.
5. The three files show up in the gallery. Download them now and rename them after listening.

About that Cost line: it is a reservation. Kadraris cannot measure the song before downloading it, so it holds the price of ten minutes, 10 × $1.50 per minute. Once the job ends, you are charged only for the minutes actually started and the remainder returns to your balance at once. Because the full hold has to fit within your available funds, the smallest top-up may not be enough to start a job here.

Running it from code

There is a single input field, audio_url, which must be a public http(s) link to the file:

curl https://kadraris.com/v1/generate \
  -H "Authorization: Bearer $KADRARIS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "multistem-splitter",
    "input": { "audio_url": "https://your-server.com/audio/track.wav" }
  }'

You get 202 with an id and estimated_cost, which is the hold. Poll GET /v1/tasks/{id}; when it succeeds, every stem is a separate item in output with type: "audio".

A Python version that submits, waits and writes the files:

import os, time, requests

API = "https://kadraris.com/v1"
KEY = os.environ["KADRARIS_API_KEY"]
HEADERS = {"Authorization": f"Bearer {KEY}"}

job = requests.post(f"{API}/generate", headers=HEADERS,
                    timeout=30, json={
    "model": "multistem-splitter",
    "input": {"audio_url": "https://your-server.com/audio/track.wav"},
}).json()
print("reserved:", job["estimated_cost"])

# check every 5 seconds until the split is done
while job["status"] not in ("succeeded", "failed"):
    time.sleep(5)
    job = requests.get(f"{API}/tasks/{job['id']}",
                       headers=HEADERS, timeout=30).json()

if job["status"] == "failed":
    raise SystemExit(job["error"])

for n, stem in enumerate(job["output"], start=1):
    link = stem["url"]
    ext = link.split("?")[0].rsplit(".", 1)[-1]
    name = f"track-stem-{n}.{ext}"
    with open(name, "wb") as f:
        f.write(requests.get(link, timeout=120).content)
    print("wrote", name)

print("charged:", job["cost"])

Do not assume a fixed order in the array. Listen to the first results and label them by ear. Links expire, so the script downloads immediately. For many tracks, add webhook_url so Kadraris notifies your server when each one is ready; see webhooks and the general pattern in async AI video with webhooks.

Getting cleaner stems

Separation works on a finished mix, so the source decides a lot:

• Expect some bleed. In busy arrangements, bits of one instrument leak into another stem: cymbals in the vocal, kick in the bass. Listen to each file before using it.
• Feed it the best master you have. WAV or FLAC separates more cleanly than a crushed MP3. FLAC works through the API as a link; Studio uploads accept MP3 or WAV up to 25 MB.
• Link to the file itself. audio_url cannot be a streaming page; it has to resolve to audio.
• No control over stems or format for now: you always get vocals, drums and bass in the vendor's default format.

Splitting a song does not change who owns it. Releasing a remix of another artist's track needs their permission; your own recordings and AI songs you hold rights to are the simple case. To put a finished song to pictures, read the AI music video guide. If you are comparing music services, see the Suno API alternative.

Questions

Why do I hear hi-hats in the vocal file?

That is bleed, and it is normal for dense mixes. A higher quality source (WAV or FLAC rather than MP3) usually reduces it. For small leaks, an EQ or gate in your DAW cleans up the rest.

Can I split a song that is longer than ten minutes?

The ten-minute figure is only the size of the hold Studio reserves before it knows the real length. Your balance needs to cover that reservation; the final bill follows the actual minutes.

Will I get piano or guitar stems?

Not yet. The engine can do it, but Kadraris does not expose stem selection, so every job returns vocals, drums and bass. For "everything but the voice" in one file, use Music Extractor.

Can I try it for free?

No. There is no free tier or welcome credit. You can start with a $10 top-up, failed jobs cost nothing, and a 50-second excerpt is billed as one minute, which makes a cheap first test.