Speech Stage - ComfyTV

What this node does

Speech Stage is ComfyTV’s text-to-speech (TTS) node. Put the full script in main_prompt, pick a TTS workflow, click ▶ Run, and get a COMFYTV_AUDIO speech snapshot. The stage is model-agnostic: the UI exposes parameters common to most TTS backends (text, voice, language, speed, optional reference audio). Engine-specific options live in sidebar Custom Params via preset JSON. Do not use Speech Stage like Music Stage: there are no genre/BPM tags—main_prompt is the script to read aloud.

When to use it

Short-form narration, ad voice-over, character dialogue.
Copy dialogue from Storyboard shots into Speech Stage.
Audio for Video Stage IA2V (Speech audio → Video audio).
Voice cloning with reference_audio + reference_text (F5-TTS, GPT-SoVITS, etc.).

How ComfyTV stages work

▶ Run executes only the subgraph in workflows/speech/.
Snapshots: MP3/FLAC URL stored in the project for IA2V and Timeline.
Native ComfyUI has no TTS (ACE-Step is music-only); Speech relies on custom nodes (e.g. Kokoro).

Division of labor: Music = melody, arrangement, optional sung lyrics; Speech = accurate reading of given text.

Types (COMFYTV_AUDIO vs native ComfyUI)

ComfyTV type	What it is	vs native ComfyUI
`COMFYTV_AUDIO`	Speech clip project snapshot	Not an in-memory AUDIO tensor
Native AUDIO	Sample tensor	Needs Bridge for ComfyTV

Conversion:

Native → ComfyTV: → ComfyTV Audio
ComfyTV → native: ← ComfyTV Audio

See bridges.md.

Speech Stage vs Music Stage (important)

	Speech Stage (this node)	Music Stage
Purpose	Read scripts, narration, dubbing	Songs, instrumentals, SFX
main_prompt	Full script to speak	Music style tags
voice / language	TTS voice and language	N/A (lyrics for singing)
Typical workflow	Kokoro TTS	ACE-Step v1 Song
Sound	Spoken voice, no full arrangement	Musical structure, possible vocals

Parameters

workflow

TTS backend from workflows/speech/. Built-in: Kokoro TTS (kokoro-tts.json)

Kokoro-82M, ONNX on CPU, lightweight for end-to-end testing.
Install custom node:

cd ComfyUI/custom_nodes
git clone https://github.com/stavsap/comfyui-kokoro.git
pip install kokoro-onnx onnxruntime

Model (~300MB) downloads on first run. Details: workflows/speech/README.md.

main_prompt (script)

The full text to speak—not music tags. Example: Welcome to today's show. Next, we visit a cyberpunk city of the future.

voice (preset speaker)

Named speaker id for preset-voice models (Kokoro, Bark, ElevenLabs). Kokoro uses preset defaults like af_sarah. Leave empty when cloning from reference_audio (if the workflow supports it).

language

Language for multilingual models. Auto = workflow/model default (Kokoro falls back to English). Kokoro requires exact names: English, Mandarin Chinese, Japanese, French, etc. Mistake: Expecting Chinese with Auto on Kokoro—pick Mandarin Chinese explicitly.

speed

1.0 = natural rate, range 0.5–2.0.

reference_text

Transcript of the reference clip; required for F5-TTS / GPT-SoVITS cloning. Auto-transcribing engines may leave it empty.

reference_audio

Optional COMFYTV_AUDIO for voice cloning. Wire upstream audio or Load Audio from Asset; bind via upstream_audio:value in preset. Kokoro does not support cloning—these fields are ignored for Kokoro.

custom_params

Sidebar bindings for engine-specific params.

Outputs

Output	Type	Meaning	Downstream
audio	`COMFYTV_AUDIO`	Synthesized speech snapshot	Video IA2V audio; Timeline; Bridge

Step by step for beginners

Install Kokoro (above), add Project + Speech Stage.
workflow → Kokoro TTS.
main_prompt: a short narration (2–3 sentences).
language → Mandarin Chinese for Chinese (avoid Auto).
voice empty (preset default), speed 1.0.
▶ Run and preview on the node.
For lip-sync video: Video Stage → IA2V, Speech audio → Video audio, portrait on images.

Full guides (recommended reading)

This page covers one node only. For end-to-end workflows, multi-stage pipelines, type conversion, and design rationale, see the ComfyTV user guides on GitHub:

Guide	Contents
Getting started	Install, canvas basics, per-stage Run, snapshots, Project, Image Picker
Generate	Text, Image, Video, Music, and Speech stages and workflow selection
Model files	Checkpoints, LoRAs, and folder paths per workflow
Custom workflows	Import your ComfyUI workflow JSON without Python changes

Repository and workflows

Resource	Link
GitHub repository	https://github.com/jtydhr88/ComfyTV
User guides index	https://github.com/jtydhr88/ComfyTV/tree/main/docs
Built-in workflows	https://github.com/jtydhr88/ComfyTV/tree/main/workflows
Model checklist	https://github.com/jtydhr88/ComfyTV/blob/main/docs/models.md
This node’s workflow folder	https://github.com/jtydhr88/ComfyTV/tree/main/workflows/speech
Workflow README	https://github.com/jtydhr88/ComfyTV/blob/main/workflows/speech/README.md
Custom workflows	https://github.com/jtydhr88/ComfyTV/blob/main/docs/custom-workflows.md

FAQ

Q: Reads English but I wrote Chinese?
A: On Kokoro, set language to Mandarin Chinese, not Auto. Q: Speech or Music Stage?
A: Read script → Speech; BGM or song → Music Stage. Q: Kokoro not found?
A: Install comfyui-kokoro + kokoro-onnx, restart ComfyUI. Q: reference_audio has no effect?
A: Kokoro has no cloning—use F5-TTS etc. with proper preset bindings.

Music Stage (Audio Stage)—songs/instrumentals, not TTS.
Video Stage (IA2V)—audio from Speech drives visuals.
Storyboard Stage—copy shot dialogue to Speech main_prompt.
Load Audio from Asset—pick existing voice tracks.
Bridge → / ← ComfyTV Audio—native interop.
Project—project context.

Music Stage (Audio Stage)

3D Model Stage

⌘I

​What this node does

​When to use it

​How ComfyTV stages work

​Types (COMFYTV_AUDIO vs native ComfyUI)

​Speech Stage vs Music Stage (important)

​Parameters

​workflow

​main_prompt (script)

​voice (preset speaker)

​language

​speed

​reference_text

​reference_audio

​custom_params

​Outputs

​Step by step for beginners

​Full guides (recommended reading)

​Repository and workflows

​FAQ

​Related nodes