Gemini 3.8 Flash Text to speech

What it is

A text-to-speech node that generates conversation, not narration: a cast of speakers defined once, a script of dialogue turns, and one AUDIO object back containing the whole exchange. It calls google/gemini-3-8-flash-tts on KIE, which relays Google's Gemini TTS.

The use case is specific and worth stating: you have a scene with two or three characters and you want their lines performed with distinct voices. Doing that with a single-voice TTS means stitching files and hand-matching tone. This does it in one job.

How it works

The widget shape is the whole node:

• speakers - a JSON array of speaker configs. Each has speaker_id, voice_name (Gemini voices: Fenrir, Rasalgethi, Iapetus, Schedar…), and then audio_profile, accent, style, pace to describe delivery.
• dialogue_turns - a JSON array of {speaker_id, text}, performed in order.
• scene - the room and recording context, e.g. a quiet room with a fireplace crackling.
• sample_context - the overall tone and delivery reference, e.g. audiobook narration, gentle and inviting.
• temperature - 0 to 2. Tightens or loosens the performance.

The join key between the two arrays is speaker_id. Get it wrong and a line comes back in the wrong voice, or nothing comes back at all.

Under the hood it's an async KIE market task: submitted to /api/v1/jobs/createTask, polled until complete or until timeout_seconds (default 1200 s) elapses. Hence the task_id output, and hence the fact that a job can finish on KIE's side after your node has stopped waiting.

First thing to do: clear the placeholder defaults. The speakers and dialogue_turns values that ship in the widget are copied from KIE's docs example - two speakers with lorem-ipsum lines and a British (Brixton) accent. The node doesn't care that it's nonsense. It will synthesise it and charge you.

Inputs and outputs

Required: speakers, dialogue_turns. Optional: temperature, scene, sample_context, timeout_seconds, callback_url (leave empty in ComfyUI; it's for server-side callbacks).

Outputs: audio, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. The native audio output is what you wire onward - give it to KIE • Assemble Video + Music as dialogue_audio and the music bed ducks under it, or to KIE • Save Video's chain via a mux step. Keep task_id for KIE • Reuse Completed Audio so a re-render doesn't cost twice, and log the whole thing with KIE • Save Generation Recipe.

Chain the writing too: an LLM node (Gemini 3.8 Flash, or Claude Opus 5.5 for heavier work) can emit the dialogue_turns JSON, so your script and your cast stay in one graph.

Install

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

Then Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. Paid node: every run submits a job.

Where people get burned

• Lite vs full. The pack also ships Gemini 3.8 Flash Lite TTS. The two nodes are identical apart from the model ID. Lite is the lighter tier - audition both on one real line and keep whichever reads your script better; don't assume the bigger name is automatically right for a short line.
• Long scripts time out. Bump timeout_seconds past 1200 for anything substantial, or recover the finished job with the task_id rather than re-running it.
• No word-level timings. You get a single audio file. Syncing dialogue to picture is your job, and it's the step that eats the afternoon.
• JSON discipline. Both arrays must parse. Edit them in a text editor, not by hand-juggling quotes inside the widget.
• It's hosted, so it filters. Content policy is Google's, and nothing in the node can route around it - which is the standing trade for every API node in ComfyUI.