Gemini 3.8 Flash-Lite TTS · Text to Speech

Cheapest per second of any audio node here: $0.00021, against $0.00032 for Flash and $0.00045 for 2.5 Pro. Same 30 voices, same 24 kHz mono WAV, usually back within seconds. For long-form narration, audiobook drafts, or anything where you're generating a lot of speech and only some of it survives, this is the node to reach for.

The caveat is in the node's own tooltip, and it's worth reading twice: spell out sounds such as sighs or laughter; bracketed cues may be read aloud. That's the tell that this tier is the least theatrical of the three Google TTS endpoints. Flash acts out [whispering]; Lite might just say "whispering" at you. If you need the performance, pay the extra $0.00011 a second and use Flash. If you need the words read clearly and cheaply, this is the one.

How it runs

Your script and settings go to SpicyAPI, the model returns audio, and the pack gives it back as a ComfyUI AUDIO - the same AUDIO type any ComfyUI audio node produces, so it drops into Save Audio or a video-sync workflow with no conversion. No local model, no VRAM, no download. One quirk worth knowing before you wire it up: several results can return as a batch, and Save Audio writes one file per result.

The inputs

• text - spoken exactly as written, and the field that determines the bill.
• voice_id - 30 voices, Kore by default, each with a fixed character.
• style, pace, accent - all left to the model's defaults on this node. There's real room here: a slower pace and a tighter style do more for a narration than any amount of punctuation tweaking.
• instructions - mood, mic distance, who's listening. Free text, not spoken, not billed.
• scene - where and to whom it's spoken. Same deal: free, and it shapes delivery.
• temperature - −1 to 2, default −1 (model's choice).
• dialogue + speakers - JSON for a two-person conversation, used instead of text. Leave text blank if you go this way.

Speakers carry a name plus an optional pace, style, accent, profile and required voice_id; each dialogue line names its speaker:

{"speakers": [{"name": "Host", "voice_id": "Kore"}],
 "dialogue": [{"speaker": "Host", "text": "Welcome back to the show."}]}

Outputs: audio and task id, a STRING identifying the SpicyAPI request.

Install

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

Restart ComfyUI, then Settings > SpicyAPI and paste a key from spicyapi.ai/console/keys - or export SPICY_API_KEY on a server. Restart once more so the node list is rebuilt from your account. ComfyUI Manager users: search SpicyAPI. No dependencies and no model files; the pack ships nothing but Python that talks to the API.

Cost

$0.00021 per second of audio - a rate so low that the real cost of a bad read is your time, not your balance. Rate shown on the node, quote taken before submission, and Settings > SpicyAPI > Max cost per run stops anything above your ceiling before it's charged. The pack forces seeds to fixed, so re-queueing reuses the cached result rather than paying for a second read; set control after generate to randomize when you want a different take.

Errors are uniform across the pack: No SpicyAPI key is set, 40310 for an unverified account email, 40004 when no deployment serves the combination (the message names the field, nothing charged), and request_id ... errors to pass to support. Invalid JSON in dialogue or speakers is rejected as an input error. Cancelling a run stops your wait, not the server-side job.