Suno Voice Create Custom Voice API

This is the node that actually mints a Suno voice - the thing you'll later drop into a music generation node when you want a specific singer instead of whatever the model invents. And yes, it's a bit of a ritual: KIE asks you to record a phrase back, and you hand the recording to this node.

Worth setting expectations about the pipeline first. Suno still wins the song-generation job - the open alternative, ACE-Step, gets genuinely close on instrumental work and then falls over on vocals and lyrics - and voice cloning is the part of that job where the API has no serious local rival yet. So if you want "my voice, but on a jazz track", this is the door.

The order of operations

Three nodes, used in sequence:

1. Suno Voice Generate Verification Phrase API - point it at your source recording and a slice of it. KIE returns a short phrase.
2. Suno Voice Create Custom Voice API (this node) - you record the phrase, then submit that recording here.
3. Suno Voice Check Availability API - poll until your voice exists.

Skip step 1 and this node has no task_id to work from, so it can't do anything.

Inputs that matter

task_id is required and comes from the verification-phrase step. The default in the widget is a bare "12", which will fail - paste a real ID or wire the socket.

verify_url is required and wants a URL to the audio of you performing the returned phrase. The author's tooltip adds a detail most people miss: sing it rather than speak it. It is a voice model, and a sung take gives it better material than a monotone read. This is a URL field, not an audio socket, so the practical path is KIE • Upload Audio → take its url output → wire it into verify_url. That node converts a ComfyUI AUDIO object to 16-bit PCM WAV and hosts it for you.

Then the descriptive fields, all optional but collectively what makes the voice reusable and findable: voice_name (label it something you'll recognise in a dropdown three months from now), description and style as free text, and singer_skill_level - a four-way choice of beginner, intermediate, advanced or professional. Professional is the default, and for most people it's the right one; the others are for matching a deliberately rougher performance.

timeout_seconds runs on the standard 30–7200 second range with a 1200-second default, and callback_url should stay empty for normal ComfyUI use.

Outputs are the pack's usual set: audio, url, all_urls_json, task_id, raw_json, plus credits_consumed and credits_left. The task_id is the one you actually need downstream - save it somewhere, because you'll paste it into the availability check. Everything your voice ends up being worth is in raw_json, and you'll want the voice ID from there to use later as persona_id with persona_model set to voice_persona on a music generation node.

Install and key

One-time setup, shared by every node in the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r requirements.txt

Restart ComfyUI, then Settings → KIE.ai Nodes Next → Connection, paste your API key, and confirm the CONNECTED status. Nothing else to configure; there are no local models. ComfyUI Manager can do the install for you if you'd rather search "KIE.ai Nodes Next" and click install.

Practical notes

Cloning voices of real people without permission is a bad idea and, in a growing pile of jurisdictions, an illegal one. The pack is careful about that elsewhere - its character packs ask you to confirm permission when you mark a likeness as a real person's - and the same logic applies here.

Two mechanics that will bite you eventually: verify_url must be publicly reachable, which is exactly why the upload node exists; and any link KIE hands back is temporary, so don't build a workflow that assumes a stored voice URL still resolves six months later. Look the voice up again rather than hardcoding.

Finally, because this pack syncs its node list from KIE's live docs, expect the odd unannounced reshape of these voice endpoints. When something that worked last week returns a schema error, run a catalog refresh before you start editing your graph.