Seed Audio 1.0 · Text to Audio

Most text-to-speech nodes give you one voice reading one line. Seed Audio 1.0 is a different shape of tool: you describe a scene - who speaks, how, what they say in quotes, plus music, ambience and effects - and it returns up to about two minutes of finished audio with the whole thing mixed. $0.36 per run, flat, regardless of how long it runs.

That flat price is the detail that decides whether this node is useful to you. A 30-second clip and a 110-second clip cost the same, so once you've paid the $0.36 you should be getting your money's worth. Sketch a full scene, not a sentence.

For context on where it sits: the community's TTS landscape is dominated by local options - Kokoro for CPU-only work, Piper for tiny and fast, Chatterbox and F5-TTS for cloning - with the API winning on exactly two jobs, best-in-class multilingual TTS and song generation. A multi-speaker scene with music and Foley baked into one file is closer to that second job than to "read this sentence aloud", and that's the niche this node fills.

Writing the prompt

prompt is multiline and it is the entire creative control. The author's own description is the best instruction: describe the whole audio scene - who speaks, how, what they say in quotes, and any music, ambience or sound effects.

The quotes matter. Put dialogue inside quotation marks or the model narrates a description of talking rather than performing the line. Name your speakers so the model can keep them distinct. And because effects are in scope, a scene can include the room tone and the sound of a door as part of the same render.

The knobs

• speed - a float from 0.5 to 2, default 1. Normal speech rate.
• pitch - an integer from -12 to 12 semitones; 0 keeps the original pitch. Small values (1 or 2) read as a different speaker without sounding like a chipmunk.
• volume - 0.5 to 2, default 1. A loudness multiplier.
• output_format - mp3 by default, with wav, opus and pcm available. The author flags pcm as headerless 16-bit little-endian stereo - fine if your pipeline expects raw samples, annoying if you expected a playable file.
• sample_rate - a dropdown including model default (the model uses 24000 Hz when it's omitted) plus 8000, 16000, 24000, 32000, 44100 and 48000. Leave it on model default unless something downstream insists.

The two optional inputs, and their conflict

reference_audio_urls is an autogrow group taking up to three reference clips, addressed in the prompt as @Audio1 through @Audio3 - that's how you make a voice recur across a scene, or hold a music bed to a specific feel. image_url takes an IMAGE socket and builds the audio around a picture.

They're mutually exclusive. The tooltips say it plainly: reference audio is "not with image_url", and the image is "not with reference_audio_urls". Pick your anchor.

Outputs

audio is a standard ComfyUI AUDIO value - Save Audio takes it. Note the tooltip: several results come back as one batch and Save Audio writes a file per item, so a multi-result run isn't lost. task id is a string for your console or a support ticket.

If you want the audio paired with a face, the same pack carries Lip Sync and Talking Avatar model nodes that take audio in.

Install

Manager → search SpicyAPI → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart, add a key under Settings > SpicyAPI (or set SPICY_API_KEY), restart again. No dependencies, no downloads - the model runs server-side, so audio generation costs you nothing but the flat $0.36.

Practical notes

Set Max cost per run in Settings anyway; it's per-run and this node's run price is fixed, so the cap mostly protects you from queueing the same scene ten times. Reference clips are uploaded for you, and the upload caps sit at 90 MiB for audio. Interrupting a run doesn't cancel it - it finishes on SpicyAPI, gets charged if it succeeds, and the result stays in your console for up to 14 days. And no key means "No SpicyAPI key is set", not a silent failure.