Gemini 3.1 Flash TTS

Gemini 3.1 Flash TTS is Google's text-to-speech model behind a Runware node - the sibling of the pack's text Gemini models, but it returns audio instead of strings. You type the script, pick a voice, and get a native AUDIO output that drops straight into a player or save node, with no local TTS stack, no model download, and no per-voice setup. It's the pick when you want Google-quality voices without opening a Google Cloud account.

Compared to the pack's own Dia models, this one is the Google-flavored option: fewer creative dials, and the voice selection happens in the Runware Speech builder rather than on this node. It's the "it just needs to sound good" option - you hand it text, it speaks. Where the Dia nodes reward tweaking, this one rewards not fiddling.

What you set

• speech (required) - from the Runware Speech builder: text (the script), voice (default Zephyr), plus speed, pitch, volume, and language. All the actual voice character lives here.
• audioSettings - optional, from the audio settings builder: channels (1 or 2), and gated bitrate / sampleRate overrides.
• settings.temperature - 0–2, default 1. Lower for consistent, flatter takes.
• seed - reproducible reads.
• outputFormat - MP3 / WAV / FLAC / OGG.

Output is audio (AUDIO) - the native dict with waveform and sample rate.

How it works

Standard pack flow: taskType: audioInference over REST through the Runware SDK, and the returned audio file is downloaded and decoded with soundfile (libsndfile) into an AUDIO tensor, falling back to torchaudio and finally to a saved-file warning if decode fails. Cost prints on the title bar. Because it's a hosted call, each read is a fresh inference - iterate on takes freely; short voice clips are cheap. It's also worth knowing the voice model is just one of the pack's Google nodes: the same voice contract feeds both this node and the text Gemini models, so a script you've written for one works in the other.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

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

No model downloads; deps are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Key-not-set errors first, as always. Robotic or clipped audio usually means the chosen voice doesn't fit the language - set language in the speech builder to match the script. If a take sounds too flat, nudge temperature up a touch (and remember it's capped at 2, not the 5 the Dia nodes allow). Want more granular control over the voice and its audio tokens instead of Google's defaults? That's the Dia2 2B node's job.