ComfyUI-Omnivoice Introduction
ComfyUI-Omnivoice adds speech generation to ComfyUI using OmniVoice, a multilingual text-to-speech model that supports more than 600 languages. You can turn written text into spoken audio, create speech in a voice you describe, or use a short recording as a voice reference.
The extension is designed for projects such as character dialogue, narrated artwork, and audiobooks. It includes ready-to-use voice references, a way to blend reference voices, and a tool for selecting chapters from an EPUB ebook. Its built-in text chunking lets you work with long passages without needing to split every page by hand.
How ComfyUI-Omnivoice Works
A ComfyUI workflow connects a few nodes, each with a specific job:
- OmniVoice Model Loader prepares the speech model.
- A text source supplies the words to be spoken. You can type them into OmniVoice Generate or use the OmniVoice EPUB Loader to extract ebook chapters.
- A voice source tells the model how the speech should sound. You can provide reference audio, describe a voice in words, or let the model choose one.
- OmniVoice Generate produces audio that you can connect to ComfyUI’s Save Audio node.
Think of the text as a script and the voice source as casting direction. In voice cloning mode, a short recording is the example performer. In voice design mode, a description such as “female, low pitch, British accent” provides the direction. In auto voice mode, you leave the voice choice to the model.
For a longer project, you can reuse the same voice reference and seed across Generate nodes. This is especially useful when you want separate paragraphs or chapters to sound like the same narrator.
ComfyUI-Omnivoice Features
Generate speech
OmniVoice Generate is the main node. Enter text, connect a model, and choose one of three modes:
| Mode | What you provide | When to use it |
|---|---|---|
voice_cloning | Reference audio and, optionally, its transcript | You want speech based on an existing voice recording. |
voice_design | A written description in instruct | You have a voice in mind but no reference recording. |
auto_voice | No specific voice reference or description | You want the model to choose a voice. |
The node outputs AUDIO at 24 kHz, which connects directly to ComfyUI’s Save Audio node.
You can also adjust these settings:
speedcontrols speaking speed. The default is1.0; changing the multiplier changes the pace of the resulting speech.num_stepcontrols the number of diffusion steps. The default is32; the author suggests16for faster generation.seedcontrols a source of variation in generation.0uses a random seed. For an audiobook workflow, set the same nonzero seed on all Generate nodes to help keep the voice consistent.
For voice cloning, ref_text is the transcription of the reference recording. It is optional, but the author recommends connecting a Whisper transcription node for best results.
Use a ready-made voice
OmniVoice Voice Preset supplies both reference audio and its transcript, so you do not need to prepare your own clip. Choose a preset, then connect its ref_audio and ref_text outputs to OmniVoice Generate. Preset audio is downloaded once and cached for later use.
| Preset | Gender | Listed style |
|---|---|---|
| Shadowheart | Female | Expressive |
| American actress | Female | Theatrical |
| Podcast host | Female | Casual |
| Nature | Male | Warm |
| Old Hollywood | Male | Classic |
| Rick Sanchez | Male | Casual |
| Stewie Griffin | Male | Precise |
| Harvey Keitel | Male | Intense |
| Conan O'Brien | Male | Comedy |
Preset names describe the available references; listen to generated samples to decide which one suits your project.
Blend reference voices
OmniVoice Mix Voices combines two required reference clips and an optional third clip. It joins portions of the clips into one reference, from which the model derives a hybrid speaker voice. Connect the resulting ref_audio and ref_text to OmniVoice Generate in voice cloning mode.
weight_1, weight_2, and optional weight_3 control how much of each clip’s duration is included. They are not precise sliders for traits such as pitch or accent. Equal weights of 1.0 and 1.0 are a sensible starting point; change them and listen to the result. If you supply transcripts in text_1, text_2, and text_3, the node merges them into its ref_text output.
Turn ebook chapters into narration
OmniVoice EPUB Loader reads an .epub file and sends selected chapters as plain text to OmniVoice Generate. Enter the file’s absolute path, then choose chapter_start and chapter_end. Chapter numbers start at 1, the end chapter is included, and an end value beyond the available chapters is automatically limited to the last chapter.
The node also outputs chapter_list, a numbered list of chapter titles that you can use to identify the chapters you want. Selected chapter text is joined with --- separators. For long text, OmniVoice Generate uses built-in chunking to keep VRAM use approximately constant rather than requiring the whole passage to be generated at once.
Example workflows
The included workflows/voice_cloning.json is a starting workflow. Load it with ComfyUI → Load Workflow. It connects a Model Loader, Voice Preset, Generate node, and Save Audio node.
For an ebook, replace the manually entered script with text from OmniVoice EPUB Loader. Keep the same voice reference and a fixed seed across Generate nodes when creating separate sections of an audiobook.
ComfyUI-Omnivoice Models
The provided documentation describes one speech model: OmniVoice. It does not list alternative model sizes or versions to choose between. OmniVoice Model Loader downloads the model from Hugging Face on first use and caches it in ComfyUI/models/omnivoice/.
The loader offers two settings:
deviceselectscuda:0,cuda:1, orcpu. Choose the device on which you want the model to run.dtypeselectsfloat16,bfloat16, orfloat32.
The loader’s OMNIVOICE_MODEL output connects to OmniVoice Generate. Voice presets are reference recordings, not separate speech models.
Troubleshooting ComfyUI-Omnivoice
No audio is saved
- Confirm that OmniVoice Model Loader is connected to Generate’s
modelinput. - Check that Generate has text to speak, either entered directly or connected from EPUB Loader.
- Connect Generate’s
AUDIOoutput to a Save Audio node and run the workflow.
The generated voice is not the one you intended
- Check the selected
mode. A reference clip belongs withvoice_cloning; a written voice description belongs withvoice_design. - In cloning mode, confirm that your audio or Voice Preset is connected to
ref_audio. - If you have a transcript of the recording, connect it to
ref_text. The author recommends a Whisper node to obtain one. - For a blended voice, start with equal mix weights and adjust them after listening.
A narrator sounds different between sections
- Use the same voice reference for each Generate node.
- Set the same nonzero
seedon those nodes rather than leaving it at0, which means random.
The wrong ebook chapters are selected
- Check the EPUB Loader’s
chapter_listoutput for its numbered titles. - Enter those numbers in
chapter_startandchapter_end. Remember that numbering begins at 1 and both endpoints are included. - Check that
epub_pathis the absolute path to the intended.epubfile.
Generation takes longer than you want
Try changing num_step from its default of 32 to 16, as suggested by the author for faster generation. Listen to the output to decide which setting you prefer.
Frequently asked questions
Do I need a recording to create a voice?
No. You can use voice design to describe one, auto voice to let the model choose, or a built-in Voice Preset.
Can I use my own recording?
Yes. Connect it to ref_audio in voice cloning mode. You can also supply its transcript through ref_text.
Can I narrate more than one chapter?
Yes. EPUB Loader lets you select an inclusive range of chapters, and Generate supports long text through built-in chunking.
Can I combine three voices?
Yes. Mix Voices requires two clips and accepts a third optional clip.
Learn More about ComfyUI-Omnivoice
- OmniVoice project — the underlying speech model and its project documentation.
- OmniVoice paper — a more technical explanation of the research behind the model.
- Included workflow:
workflows/voice_cloning.json— a practical example you can open in ComfyUI to see how the nodes connect.
