Gemini Text-to-Speech

Text-to-speech in ComfyUI usually means a local model - Kokoro for tiny-and-fast, Chatterbox for quality - and those are genuinely good now. So why would you call Gemini's API for the same job? Because this is the "make a video clip actually talk" node, and Gemini's TTS voices land somewhere between ElevenLabs-class naturalness and the zero-setup convenience of a hosted call. Thirty prebuilt voices (Zephyr, Puck, Kore, Fenrir, and friends), three models, no voice files to manage, no VRAM. Paste text, get audio.

The KB's audio-generation.md frames it honestly: local open TTS is competitive, and for cloning your own voice, local is now the privacy-preserving default. But for grabbing a polished, expressive narrator voice for a video render without installing anything, this node is the fastest on-ramp in the pack. It also does something neat that local TTS usually needs extra tooling for: multi-speaker dialogue by typing speaker tags right in the text.

How it works

Gemini TTS runs through the same generate_content endpoint as the text nodes, but with the response modality set to audio and a speech_config carrying your chosen voice. The text goes in, and instead of tokens you get back a raw audio blob - specifically 24 kHz signed 16-bit PCM, which the node converts into a standard ComfyUI AUDIO dict. The style_prompt input is prepended to your text ("Say cheerfully: …") to steer emotion and pacing without changing the content.

Inputs and outputs that matter

• text - what to say. The multi-speaker trick: use speaker tags like Alice: Hi there. Bob: Hello! and the model assigns voices to each speaker. That's real capability for dialogue scenes.
• voice - 30 prebuilt names. Each has a character; Kore and Fenrir are the defaults people tend to land on. Try a few - they're noticeably distinct.
• model - three TTS models: gemini-2.5-flash-preview-tts (default, fast), gemini-2.5-pro-preview-tts (higher quality), and the newer gemini-3.1-flash-tts-preview.
• style_prompt - the emotion dial. "Say it slowly and somberly:" changes the delivery.

One output: audio (AUDIO), ready for a save/preview node or to pair with a lip-synced video.

How to install it

Pack-level install:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-API-Toolkit
cd ComfyUI-API-Toolkit
pip install -r requirements.txt

or "API Toolkit" in ComfyUI Manager. Needs google-genai>=0.8.0 and a Gemini API key. Nothing to download - voices are server-side.

Common issues

The failure you'll hit first is "Gemini TTS returned no audio data," which is the node's way of saying the call came back without audio - usually a prompt that got filtered or a model ID that's rotated out (the TTS previews change; check the pack's List Models node and use custom_model if needed).

Beyond that, manage expectations: Gemini voices are great but they are what they are - you can't clone a specific voice with this node, so if you need your voice, that's a job for local cloning or ElevenLabs. And every queue is a paid call (the pack's API nodes always re-execute), so a long narration is one request, but re-Queueing to tweak one line bills the whole clip again. Keep the text tight while you iterate.