Custom nodes for the MOSS-TTS model family from OpenMOSS. Brings text-to-speech, zero-shot voice cloning, voice design, sound effect generation, and multi-speaker dialogue synthesis into ComfyUI workflows.

• Text-to-speech: Generate natural speech from text using the base MOSS-TTS models (8B Delay or 1.7B Local)
• Zero-shot voice cloning: Clone any voice from a 3-10 second reference clip, no transcript required
• Voice design: Describe a voice in plain language and generate speech in that style (requires MOSS-VoiceGenerator)
• Sound effect generation: Generate ambient sounds and audio effects from text descriptions with controllable duration (requires MOSS-SoundEffect)
• Multi-speaker dialogue: Synthesize conversations between two speakers with optional per-speaker voice cloning (requires MOSS-TTSD)
• Duration control: Target a specific output length using token-based duration hints
• All models output 24 kHz audio compatible with ComfyUI's native AUDIO type

1. Clone into your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/richservo/comfyui-moss-tts

2. Install dependencies:

pip install -r ComfyUI/custom_nodes/comfyui-moss-tts/requirements.txt

Note: transformers>=5.0.0 is required. Earlier versions may fail to load the MOSS model architecture. torch and torchaudio are not listed in requirements because ComfyUI provides them.

3. Restart ComfyUI.

All five nodes will appear under the audio/MOSS-TTS category in the node search.

Models download automatically from HuggingFace on first use and are cached to ComfyUI/models/moss-tts/.

MOSS-TTS (Local 1.7B) is the only model that is fast enough for practical iterative use on a single consumer GPU. All other models use the Delay 8B architecture and are significantly slower.

The Audio Tokenizer (OpenMOSS-Team/MOSS-Audio-Tokenizer) is a shared codec required by all models. It downloads automatically alongside whichever model you load first.

Loads a MOSS-TTS model and processor into a pipeline object (MOSS_TTS_PIPE) that is passed to any generation node.

Inputs:

Output: MOSS_TTS_PIPE

Notes:

• Calls mm.unload_all_models() before loading to free VRAM.
• Models are cached to ComfyUI/models/moss-tts/ on first download.
• The node uses an MD5 hash of its inputs to detect changes and avoid unnecessary reloads.

Generates speech from text. Optionally clones a voice from a reference audio clip.

Inputs:

Output: AUDIO

Notes:

• Without reference_audio: performs standard TTS with the model's default voice.
• With reference_audio: performs zero-shot voice cloning. No reference transcript is needed (unlike Qwen-TTS).
• The node re-runs whenever seed changes.

Generates speech in a voice described by a natural language instruction. Requires the MOSS-VoiceGenerator model.

Inputs:

Output: AUDIO

Notes:

• The node will print a warning if the loaded model is not MOSS-VoiceGenerator, but will still attempt generation.

Generates ambient sounds and audio effects from a text description. Requires the MOSS-SoundEffect model.

Inputs:

Output: AUDIO

Notes:

• Duration is converted to tokens internally: tokens = int(duration_seconds * 12.5).
• The node will print a warning if the loaded model is not MOSS-SoundEffect.

Synthesizes multi-speaker dialogue with two distinct voices. Supports optional per-speaker voice cloning from reference audio. Requires the MOSS-TTSD v1.0 model.

Inputs:

Output: AUDIO

Dialogue text format:

[S1] Hello, how are you today?
[S2] I'm doing great, thanks for asking!
[S1] That's wonderful to hear.

Accepted tag formats: [S1]/[S2], [s1]/[s2], or [1]/[2]. When normalize_text is enabled, all formats are normalized to [S1]/[S2] automatically.

Notes:

• Without reference audio: the model assigns random but distinct voices to each speaker.
• With reference audio for one or both speakers: the node uses continuation mode, conditioning generation on the reference clips.
• s1_prompt_text / s2_prompt_text are used as the continuation prefix transcript. If left empty, the reference audio is still used for voice conditioning.
• The node will print a warning if the loaded model is not MOSS-TTSD-v1.0.

MOSS-TTS Model Loader  →  MOSS-TTS Generate  →  PreviewAudio

Load the MOSS-TTS (Local 1.7B) model for fast iteration. Enter text in the Generate node and queue the prompt.

LoadAudio (reference clip)  ──┐
                               ├─→  MOSS-TTS Generate  →  PreviewAudio
MOSS-TTS Model Loader  ───────┘

Connect a short audio clip (3-10 seconds) to the reference_audio input of the Generate node. The model will clone the voice without requiring a transcript.

MOSS-TTS Model Loader (VoiceGenerator)  →  MOSS-TTS Voice Design  →  PreviewAudio

Load the MOSS-VoiceGenerator model. Describe the desired voice in the instruction field and provide the text to speak.

MOSS-TTS Model Loader (SoundEffect)  →  MOSS-TTS Sound Effect  →  PreviewAudio

Load the MOSS-SoundEffect model. Describe the desired sound in ambient_sound and set duration_seconds.

MOSS-TTS Model Loader (TTSD)  →  MOSS-TTS Dialogue  →  PreviewAudio

Load the MOSS-TTSD v1.0 model. Write dialogue in the dialogue_text field using [S1]/[S2] tags. Optionally connect reference audio clips for each speaker to clone their voices.

Reference audio quality:

• The optimal reference clip length is 3-10 seconds.
• Use clean recordings with a single speaker and no background noise.
• Clips longer than ~15 seconds may introduce noise artifacts or degrade quality.

Model selection:

• Use MOSS-TTS (Local 1.7B) for all day-to-day work — it is the only model fast enough for iterative use on a single GPU.
• Use the Delay 8B models (base TTS, TTSD, VoiceGenerator, SoundEffect) when maximum quality is the priority and generation time is not a concern.

Sampling parameters:

• Each node ships with defaults tuned for its intended model. If you switch models, refer to the recommended hyperparameters table below.
• Lowering temperature produces more consistent but less expressive output. Raising it increases variety but may introduce instability.

Recommended parameters per model:

Duration control:

• All models use a codec frame rate of 12.5 tokens per second.
• Formula: tokens = int(seconds * 12.5) — so 325 tokens ≈ 26 seconds, 125 tokens ≈ 10 seconds.
• Duration control is a hint, not a hard limit. Actual output length may vary.

Audio output:

• All models output 24 kHz mono audio.
• Output is a standard ComfyUI AUDIO type and is compatible with all built-in ComfyUI audio nodes (PreviewAudio, SaveAudio, etc.).

Voice cloning vs. Qwen-TTS:

• MOSS-TTS does not require a reference transcript for voice cloning. Connecting a reference clip to reference_audio is sufficient.

Dialogue tags:

• The Dialogue node accepts [S1]/[S2], [s1]/[s2], or [1]/[2] — all are normalized automatically when normalize_text is enabled.
• Keep each speaker turn on its own line for best results.

MOSS-TTS is developed by the OpenMOSS team.

• MOSS-TTS GitHub repository: https://github.com/OpenMOSS/MOSS-TTS
• HuggingFace organization: https://huggingface.co/OpenMOSS-Team

This ComfyUI node package is an independent wrapper and is not officially affiliated with OpenMOSS.