🌸 OVERTLI Pollinations Speech to Text

GZ_SpeechToText is the Overtli Studio Suite's listening node, and it's the one that makes the suite feel like a loop rather than a bunch of one-way pipes. Feed it an AUDIO clip and it sends the audio to Pollinations' transcription API, returning what was said as a STRING - or, if you ask nicely, as srt or vtt subtitles you can attach straight to a video. In a suite built mostly around writing prompts, this is the node that reads the room (literally) and hands text back.

It's also the natural partner to the pack's own GZ_TextToSpeech: speak a draft, transcribe it back, feed the transcript into an enhancer - suddenly you have a full speech-to-prompt round trip that never left ComfyUI. For captioning an existing video or turning a voice memo into prompt material, it's the same idea in one node.

How it works

The node takes an AUDIO input - and it's required in practice, even though it sits in the optional slot. The README's troubleshooting section leads with this exact mistake: "GZ_SpeechToText requires an AUDIO input connection to function." There's no "upload a file" button; you need an audio tensor in your graph from a load-audio node or a generation node like the pack's TTS. Feed it in, pick a transcription-capable model from the live catalog (43+ entries tagged [stt]; the default is a free Gemini STT entry), and run.

mode_preset is where this gets interesting - the transcription doesn't have to come back raw. Presets like Clean Transcript (readability cleanup while preserving meaning), Punctuation and Casing, and Structured Notes (concise key points and actions) post-process the raw transcription, and stt_style_preset adds delivery-aware shaping (Verbatim Accuracy, Meeting Notes, Podcast Chapters…). That's genuinely useful: raw STT output is wall-of-words; these turn it into something you can wire into a prompt.

Inputs that matter

• audio - the clip to transcribe. The critical one.
• model - STT-capable model from the catalog. If the default refuses or errors, [free] entries vary in language coverage.
• response_format - text (default), json, verbose_json, srt, or vtt. Pick srt/vtt for captions.
• mode_preset and stt_style_preset - the cleanup/structuring layer.
• input_language - defaults to auto; pin it (en, es, ja…) if auto-detection fumbles a strong accent or mixed speech.
• prompt and custom_instructions - a priming prompt can nudge the model toward domain vocabulary.
• temperature - default 0.5; lower it for verbatim, higher for looser cleanup.
• api_key - optional on the free tier.

Output

One transcript STRING. In srt/vtt mode that string is the subtitle file content - write it to a file or wire it into a captioning workflow. In plain text mode it feeds any text input, including the pack's enhancer nodes.

Install and gotchas

Same pack install: search "Overtli" in ComfyUI Manager, or clone into custom_nodes + pip install -r requirements.txt, restart. Nothing heavy to download - the audio goes up to the provider, the transcript comes back.

Where people trip: the missing-AUDIO-input error above is the #1 footgun, so check your wiring before suspecting the node. Beyond that, the usual hosted-service realities apply - free-tier transcription can be slow on long clips and accuracy is model-dependent, so a clip with heavy background music or overlapping speakers is where you'll want to pin input_language and maybe reach for a [paid] model. And everything you transcribe leaves the machine, so don't send anything you need kept private through a free cloud STT route.