⚙️ MOSS SoundEffect v2 Engine

This one isn't TTS at all. MOSS SoundEffect v2 makes sound effects from a text prompt - "thunder rolling over a distant city," "footsteps on gravel," "a sword being drawn" - as 48 kHz mono audio. It's the audio equivalent of text-to-image, dropped into the same suite as all the voice engines because it shares the infrastructure. If you're building a video or an audiobook and need foley or ambience without hunting through a stock library, this is a genuinely handy tool. It has the highest click-through of any node in this pack for a reason: people want it. This node outputs a TTS_ENGINE that runs through the suite's Sound Effects workflow.

How it works

Despite the shared TTS_ENGINE plumbing, under the hood v2 is a diffusion pipeline, not an autoregressive speech model - which is why its controls (inference_steps, cfg_scale, sigma_shift, negative_prompt) look exactly like image-diffusion controls rather than TTS ones. You prompt what you want to hear, and it denoises a 48 kHz mono clip toward that description. The text prompt itself comes from the generation node it drives (via the Unified Sound Effects flow), while this node holds the model and the diffusion settings.

The inputs and outputs that matter

• inference_steps (10–150, default 100) - diffusion steps. More steps generally means cleaner, more detailed sound and slower generation. 100 is a solid default; drop to ~40–60 when iterating on a prompt to speed things up, then push back up for the final.
• cfg_scale (1–8, default 4) - prompt adherence, same idea as CFG in image models. Higher sticks to your description harder; too high can get harsh or artifact-y. 4 is a good middle.
• sigma_shift (0–10, default 5) - shifts the noise schedule, affecting the balance of coarse structure vs. fine detail in the result. Leave at default unless you're chasing a specific texture.
• negative_prompt - describe what you don't want (e.g. "music, speech, hiss") to steer the effect away from it. Optional but useful for keeping ambience clean.

model is fixed to MOSS-SoundEffect-v2.0, device stays auto, and dtype on auto is fine. The single output is TTS_engine, which you run through the Sound Effects workflow rather than TTS Text.

Installing it

Ships with TTS Audio Suite. Easiest: ComfyUI Manager → search TTS Audio Suite → install → restart, which runs the pack's install.py and handles the dependency conflicts (NumPy, librosa, s3tokenizer) and Python 3.13. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/diodiogod/TTS-Audio-Suite.git, then python install.py from inside the folder with your ComfyUI venv active (run the script, not just a requirements install). Linux: portaudio19-dev libsamplerate0-dev first. The v2 model (~11.2GB) auto-downloads into ComfyUI/models/TTS/moss_soundeffect_v2/ on first use - it's a chunky download, so the first generation is slow. Use the "Unified 🌩️ Sound Effects" example workflow, which wires v1 and v2 with per-segment parameters and long-duration chunking.

Common issues

• Big download, big first run. ~11GB before it makes a peep. That's the model, not a hang - watch the console.
• Result doesn't match the prompt. Sound-effect prompting is a skill like image prompting. Be concrete and sensory ("heavy rain on a tin roof, no thunder") and use the negative_prompt to exclude what keeps sneaking in. Raise cfg_scale a little if it's ignoring you.
• Harsh or clipped output. cfg_scale too high or inference_steps too low. Back cfg toward 4 and give it more steps.
• Expecting speech or music. It's a sound-effects model - 48 kHz mono foley and ambience, not dialogue (that's the TTS engines) and not full musical arrangements.
• It won't behave like a normal TTS engine. Right - don't route it into TTS Text expecting spoken words. Use the Sound Effects workflow.