Generates MOSS Sound Effect

The dialogue and TTS nodes get the attention, but MossSoundEffectGenerate is the sleeper. It's a text-to-sound-effects model: describe an ambient noise or effect in plain English and it generates the audio, with a duration knob. No foley library, no hunting through ten-year-old royalty-free packs for "crackling fireplace." For video work - a rain scene, a tavern ambience, a crowd murmur behind a voiceover - this is genuinely useful, and it's the part of the MOSS family nobody else in the ComfyUI TTS space is really competing with yet.

How it works

Same architecture as the rest of the pack: the MOSS-SoundEffect model (from OpenMOSS) takes your text description, encodes it into audio tokens alongside the shared MOSS-Audio-Tokenizer codec, samples tokens, and decodes them into a waveform. The one thing this node adds is duration control. The tooltip is your math: roughly 12.5 audio tokens per second. The node converts your requested seconds into a token count, asks the model for that many tokens, then caps generation at max_new_tokens as a safety limit.

Inputs that matter

• text - the sound description. "birds chirping in the forest" and "rain on a tin roof" are the defaults and both work well. Be specific about the scene, not the mood - the model renders environments, not abstract feelings.
• duration_seconds - 0.5 to 30 seconds, default 5. This is your main creative control. It's approximate (the token-to-seconds ratio is a heuristic), so a "5 second" clip may come out a bit short or long.
• audio_temperature - default 0.7, notably lower than the speech nodes' defaults, and for good reason: effects benefit from consistency. Crank it to 1.5 and your "rain" might turn into a downpour with a thunderclap you didn't ask for.
• max_new_tokens - default 1024, capped at 4096. It's a safety limit, not a length control - leave it and use duration_seconds instead.

Output is a standard ComfyUI AUDIO tensor. Wire it to SaveAudio and drop the file into your edit.

Install

ComfyUI Manager → search ComfyUI Kaola MOSS-TTS, or:

cd ComfyUI/custom_nodes
git clone https://github.com/kana112233/ComfyUI-kaola-moss-tts.git
cd ComfyUI-kaola-moss-tts
pip install -r requirements.txt

The SoundEffect model (~3GB) plus the shared codec (~1GB) auto-download to ComfyUI/models/moss_ttsd/ on first run. Remember transformers>=5.0.0 is required and a fresh Python 3.12 environment is the recommended way to avoid dependency clashes with your other packs.

Where people get burned: they treat duration_seconds as a hard guarantee and are surprised by the fuzz around the edges, and they push temperature too high expecting "more creativity" - with effects, higher temperature mostly means more artifacts. Keep the temperature near default, generate a couple of takes, and pick the keeper. It's also brand-new tech with a thin community footprint, so expect some prompts to just miss; that's the model, not your workflow.