ComfyUI-MOSS-TTS Introduction
ComfyUI-MOSS-TTS adds custom nodes that let you use MOSS-TTS inside a ComfyUI workflow. MOSS-TTS is a 1.7-billion-parameter text-to-speech model: it turns written text into spoken audio. It also supports zero-shot voice cloning, which means you can provide a short recording as a voice reference without training a new model.
For an AI art project, you can use the extension to make spoken lines for a character, narration for a video, or audio to accompany a scene. You can generate speech from text alone, or add a reference clip when you want the result to follow a particular voice. The nodes appear in ComfyUI under the MOSS-TTS/ category.
MOSS-TTS was trained primarily on Chinese. Its default voice may pronounce English and other non-Chinese text with Chinese-leaning phonetics. For speech in another language, the author recommends supplying a clean 5–15 second reference clip in the language you want to hear.
How ComfyUI-MOSS-TTS Works
Think of the workflow as giving a performer a script and, optionally, an example voice:
- Load the performer.
MOSS-TTS Loadermakes the model available to your workflow. - Provide the script. Enter the words you want spoken in
MOSS-TTS Generate. - Optionally provide a voice example.
MOSS-TTS Voice Referenceprepares an audio clip for the generation node. The clip helps guide the voice used for the new speech. - Generate and save the audio. Connect the generation node’s
audiooutput to ComfyUI’s built-inSave Audionode.
You do not need a reference clip to generate speech, but it is especially useful when you want to guide the voice or produce speech in a language other than Chinese. A short, clean clip gives the model a clearer example to follow than one with background noise or competing voices.
The first generation downloads approximately 13 GB of model weights from Hugging Face. This initial run takes longer; later runs use the cached weights.
ComfyUI-MOSS-TTS Features
| Node | What it does | When you might use it |
|---|---|---|
MOSS-TTS Loader | Loads the model and its processor, then keeps them cached while they are in use. | Add it at the start of a speech-generation workflow. |
MOSS-TTS Voice Reference | Accepts an AUDIO input or a filesystem path and prepares it as a voice reference. | Add it when you want to guide the generated voice or pronunciation. |
MOSS-TTS Generate | Turns your text into speech, with an optional voice reference. | Use it for dialogue, narration, or other spoken audio. |
MOSS-TTS Unload | Frees the loaded model from GPU video memory (VRAM). | Place it at the end of a workflow when you need that memory for other tasks. |
To try text-to-speech without a voice reference, connect MOSS-TTS Loader to MOSS-TTS Generate, type your text, and connect the audio output to Save Audio. To use a reference recording, connect Load Audio → MOSS-TTS Voice Reference → the generation node’s voice input.
Quality presets
The preset dropdown on MOSS-TTS Generate offers a trade-off between audio quality and how much audio can be generated within the stated T4 GPU limits. RVQ layers are part of the model’s audio-generation process; for everyday use, you can treat the presets as a simple quality control.
| Preset | RVQ layers | What to expect |
|---|---|---|
Custom | Set by individual widgets | Use your own sampling settings. |
Fast (8 RVQ) | 8 | Lower quality; approximately 12 minutes of audio on a T4. |
Balanced (16 RVQ) | 16 | Default; approximately 8 minutes on a T4. |
High Quality (24 RVQ) | 24 | Higher quality; approximately 5 minutes on a T4. |
Maximum (32 RVQ) | 32 | Best quality of these presets; approximately 4 minutes on a T4. |
These durations are the author’s approximate T4 figures, not guaranteed lengths for every machine or workflow. Any preset other than Custom overrides the individual sampling widgets when the workflow runs. Choose Custom if you want your adjustments to those widgets to take effect.
The defaults target a T4-class GPU with roughly 16 GB of VRAM, using FP16 and sdpa attention. The author notes that users with at least 24 GB of VRAM may want to try dtype=bf16 and a higher max_tokens value.
ComfyUI-MOSS-TTS Models
This extension integrates MOSS-TTS, the 1.7-billion-parameter model linked above. The context does not identify separate model variants to choose from. The quality presets change generation settings; they are not different models.
Troubleshooting ComfyUI-MOSS-TTS
The audio sounds like noise instead of speech
The author identifies an unsupported transformers version as the most common cause.
- Check the installed version with
pip show transformersin the Python environment used by ComfyUI. - Confirm that it is at least 4.50 and below 4.58. Earlier versions lack APIs the model expects, while later versions, particularly 5.x, have removed APIs it still uses.
- If the version is outside that range, use a compatible 4.5x release and try the generation again.
The extension includes compatibility shims for some specific gaps, but the supported version range is the recommended choice.
English speech has unexpected pronunciation
- Select a clean 5–15 second recording spoken in English.
- Connect it through
MOSS-TTS Voice ReferencetoMOSS-TTS Generate.voice. - Generate the speech again.
Use the same approach for another non-Chinese language: choose a reference clip spoken in the language you want for the output.
The first generation takes much longer than expected
The first run downloads approximately 13 GB of model weights. The author’s smoke-test estimate is about 5–8 minutes for that download, though the time you experience can vary. Subsequent runs use cached weights.
You need GPU memory for another workflow
Add MOSS-TTS Unload at the end of your MOSS-TTS workflow. It frees the model from VRAM after you have finished generating audio.
Frequently asked questions
Do I need a voice recording?
No. You can generate speech from text alone. A reference clip is useful for guiding the voice and is recommended for non-Chinese output.
Why do my changes to the sampling widgets have no effect?
A selected quality preset overrides those widgets at execution time. Set the preset to Custom to use your individual settings.
Which preset should I try first?
Balanced (16 RVQ) is the default. You can then try Fast when you need a lower-quality, longer output, or a higher-quality preset when audio quality matters more.
Learn More about ComfyUI-MOSS-TTS
- MOSS-TTS model project: Learn about the underlying speech model from the OpenMOSS team.
- ComfyUI-MOSS-TTS extension repository: Find the extension’s source, license, and project information.
The extension is licensed under the MIT license; see the repository’s LICENSE file for its terms.
