TTS WebUI Parler TTS

Parler TTS is the Hugging Face model that inverts the usual TTS prompt. Instead of cloning from a reference clip, you describe the voice you want in plain English - "a young woman with a slight British accent, speaking quickly with a bit of emotion" - and the model synthesizes it from your words. It's the closest thing in this pack to prompting for a voice the way you'd prompt for an image. This node is the thin client for running it through the pack's TTS WebUI server.

How it works

Standard pack plumbing: the node POSTs an OpenAI-style body to http://127.0.0.1:7778/v1/audio/speech with model: "parler-tts", and the server does the work. Parler-TTS is a description-conditioned TTS: it takes your prose description as a conditioning signal and generates speech that matches it. The default model is parler-tts/parler-tts-mini-v1 - the "mini" matters, because it's the version that runs reasonably on a consumer GPU; the full parler-tts-large exists but is heavy.

The inputs that matter

• text - the line to be spoken.
• description - the voice prompt. Default is A neutral voice., and that's a fine starting point. This is the whole trick of Parler: get specific. "A female speaker with a very clear, slightly nasal voice and a soft tone" reads like prompt engineering because it basically is.
• model_name - the Hugging Face repo the server loads (parler-tts/parler-tts-mini-v1 by default). Point it at a fine-tune if the server has one.
• attn_implementation - eager by default; on newer GPUs the server may prefer a flash-attention setting for speed. Leave it unless you're chasing performance.
• compile_mode - empty by default; a server-side torch.compile option for the same "make it faster" goal.

Outputs

audio (AUDIO) into a preview/save node, wav_path (STRING) when also_save_wav is on, and metadata_json (STRING) with the request details.

Installing it

ComfyUI Manager → search "TTS WebUI API nodes for ComfyUI", or:

cd ComfyUI/custom_nodes
git clone https://github.com/rsxdalv/ComfyUI-TTS-Webui

then restart. The pack only needs requests, numpy, torch. The Parler model downloads into the TTS WebUI server on first load - the mini model is a real download, so the first request can be slow.

When it breaks

• The voice ignores your description: the description influences style, not identity. Parler won't clone a person no matter how you phrase it - it renders the described delivery. If you want a specific person's voice, use a cloning node (F5-TTS, Chatterbox) instead.
• First call times out: model download/load on the server. Raise timeout_sec (max 600).
• Connection refused: server not running on 127.0.0.1:7778, or api_base points at the wrong host.
• Slow generation: mini is already the compromise. Messing with compile_mode/attn_implementation is where you go from there.

Parler is the fun node in this pack - the one where you can get a genuinely different-sounding narrator just by writing a better sentence. Just remember the division of labor: text says what, description says how.