ComfyUI Voice Workbench

A local, English-language workbench for generating speech and shaping its pitch, word duration, voiced endings and existing pauses. Edit a whole sentence, compare recordings, and return to an earlier take.

The editor runs as a ComfyUI bottom panel or a standalone local application. Both use the same editor and audio services.

Quick start

Python 3.10 or newer is required. No model download is needed to open the bundled example and edit its audio.

git clone https://github.com/Enmani/ComfyUI-VoiceWorkbench.git
cd ComfyUI-VoiceWorkbench
python -m venv .venv

Activate the environment:

# Windows PowerShell
.venv\Scripts\Activate.ps1

# Linux / macOS
source .venv/bin/activate

Then install the CPU dependencies and open the editor:

python -m pip install -r requirements.txt
python run.py

Open http://127.0.0.1:8794/. The First Light example loads automatically. On Windows, Generate & perform → System voice → Generate take also works with an installed English speech voice. System speech provides a quick start; it does not clone a voice or control emotion.

ComfyUI installation

Clone this repository into ComfyUI/custom_nodes/ComfyUI-VoiceWorkbench. Install requirements.txt with the Python environment that starts ComfyUI, then restart ComfyUI.

For Windows portable ComfyUI, from its root directory:

.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI-VoiceWorkbench\requirements.txt

Add audio → Voice Workbench → Voice Workbench · Time & Pitch and click Open editor. Leave document_path empty to use the default example. You can also load the example workflow. The Voice Workbench bottom tab provides the same editor.

Try the default example

The sun is rising. Take a breath, and let the story begin.

1. Play Original. Four preset pitch points are already visible.
2. Drag a gold point up or down, or use arrow keys on a selected point. Double-click the curve to add a point.
3. Drag a blue handle to change word duration or a green handle to extend a reliable voiced ending. Purple blocks adjust existing near-silent pauses.
4. Click Preview, then compare Original and Edited.
5. Expand Generate & perform to save a version, generate a new take, compare it and click Edit version to apply it.

Edits save automatically. Preview audio updates when you click Preview. Changing text requires a new recording; generated takes remain separate until applied.

AI speech and voice editing

Optional local adapters support IndexTTS2 for reference-voice and emotion generation, and AuK Base for whole-recording performance editing. Engines and the Qwen3 forced aligner run in separate, user-configured Python environments. Weights are not bundled or downloaded automatically.

See engine configuration. IndexTTS2 uses the current source recording as its voice reference. AuK uses the audible edited take and resets those baked controls when applied. AuK output requires transcript verification before it can be adopted.

Import your recording

Provide a mono WAV, its transcript and word timing annotations:

python import_audio.py --audio recording.wav --text "Your complete sentence." --words words.json --name my-take

The command prints a prosody.json path for the ComfyUI node. words.json is a list of objects with id, text, start_time, and end_time; see the example annotations. Timings must be ordered, non-overlapping and inside the recording. Source audio is copied into a new job.

Audio behavior and limits

• Duration and pitch are applied to the complete recording in one Praat resynthesis pass. Word waveforms are not cut and joined. An unchanged preview preserves the original WAV bytes.
• Global speed: 0.75–1.25×. Global pitch plus local curve: −3 to +3 semitones. Up to 64 points, spaced at least 20 ms apart.
• Voiced-tail targets are bounded to a stable 60–500 ms region, 75–600 Hz, with blended joins. Ending extensions are bounded by measured voicing and capped at 200 ms.
• Mono recordings: 0.15–60 seconds. AI takes: at most 200 characters; engines impose additional token/duration limits.
• Word boundaries are estimates. Unvoiced or unstable regions cannot supply reliable pitch targets. Listen before accepting a take; automatic alignment does not prove that every word was spoken correctly.
• The server binds to loopback and accepts same-origin local requests. It is intended for one local workspace, not shared public hosting.

Workspace files default to ~/.voice-workbench. Set VOICE_WORKBENCH_HOME or use python run.py --workspace /path/to/workspace. Machine configuration stays in ignored config.json; VOICE_WORKBENCH_CONFIG can point to another file.

Development

python -m unittest discover -s tests -v
npm ci
npm run check

See architecture and validation and contributing.

License and credits

Code is licensed under GPL-3.0-or-later; see LICENSE. Audio processing uses Praat through Parselmouth. Optional engines and model weights retain their own licenses. The included example was generated by IndexTTS2 using a synthetic system-voice reference, with provenance recorded in example.json.