🌸 OVERTLI Pollinations Text to Speech

If you've ever wanted a voiceover on your video without training a voice model or babysitting a separate TTS app, GZ_TextToSpeech is the node that closes the loop inside ComfyUI. It's the Overtli Studio Suite's speech route: text goes to Pollinations' audio API, and a spoken clip comes back as a native ComfyUI AUDIO output you can wire into a video or preview directly. Text-to-speech, no local model, no VRAM, no key needed on the free tier.

The genuinely clever part is what the node calls the script path. Its default model (openai-audio-large) is tagged for both TTS and music, and the node has a whole preset family - Voiceover Script, Character Voice, Podcast Intro, Dramatic Reading - that turn your rough text into a script first and then read that script. That's why the node has two outputs instead of one.

The inputs that matter

• text - what you want spoken.
• mode_preset - off by default. Turn on something like Voiceover Script and the text gets rewritten into narration-friendly pacing before synthesis; format_script (optional) further shapes the output.
• model - audio-capable models from Pollinations' catalog, tagged [tts], [music] or both, [free]/[paid]. Defaults to a free one.
• voice - the character. OpenAI-style voices (nova, alloy, echo, fable, onyx, shimmer, plus newer ones like verse, ballad, sage). nova is the default and the safe pick.
• tts_style_preset - delivery styles (Natural Narration, Podcast Host, Broadcast News…) layered on top of the raw voice.
• custom_instructions - free-form direction for the model.

There are no speed or format knobs on this node's schema - the sibling GZ_AdvancedTextEnhancer exposes audio_speed and audio_response_format if you need that granularity - so this one stays intentionally simple.

The outputs

• audio - the native ComfyUI AUDIO output. Wire it to an audio-preview or combine it with your video track.
• script_text - the STRING that was actually spoken. This is the sleeper output: when a mode preset rewrote your text, you get the polished script back, which means you can save the narration copy without re-typing it. It also pairs with GZ_SpeechToText if you ever want to verify what the model actually said.

Install and expectations

Standard pack install - search "Overtli" in ComfyUI Manager, or clone the repo into custom_nodes and pip install -r requirements.txt, restart. Deps are light (requests, Pillow, numpy); the node converts returned audio through ComfyUI's own audio helpers, so a reasonably current ComfyUI build is worth having.

The honest framing: this is a convenience TTS, not a voice-cloning studio. The community's deep TTS work - Kokoro-class lightweight local models, Chatterbox, real cloning - lives on the local side (audio-generation.md); a hosted API gives you a good generic voice in a dozen styles with zero setup, and the free tier keeps it costless for drafts. Where it shines is pipeline speed: a script, a voice, and a narration clip that lands in your graph in one run. Where it won't help: if you need a specific cloned voice, or something the provider's moderation won't speak - the audio API's filter applies to speech like any hosted model, and there's no local patch.