MOSS-SoundEffect-v2.0

MOSS-SoundEffect-v2.0

Generates sound effects from a text prompt, such as nature, city scenes, animals and human actions, up to 30 seconds at 48 kHz. For creators and developers who describe sounds in English or Chinese.

Made by OpenMOSS-Team on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MOSS-SoundEffect v2.0 turns a written description into audio. You describe a sound in words, and it generates the clip. It uses a Diffusion Transformer trained with flow matching, together with an audio VAE and a Qwen3 text encoder.

It belongs to the MOSS-TTS family, where it is the model for non-speech sound. Version 2.0 replaces the earlier token-by-token design with a continuous-latent one.

Who it suits: It suits creators and developers who need sound effects without recording them, and who can write a short description of the sound in English or Chinese.

Industry | Who uses it | How they use it | Example
Games | Sound designers | Draft ambient and creature sounds from text | Make a forest background loop
Video | Editors | Add placeholder effects to a cut | Generate a city street ambience
Software | Developers | Build sound generation into an app | Create effects on demand
Research | Audio researchers | Study text-to-audio generation | Compare prompts and settings

If you need speech, a voice or full songs, you can skip it, since this model makes sound effects and short music-like clips.

Where to use it: The card shows a Python pipeline that runs on a CUDA GPU, on a machine where you can set up a Python 3.12 environment.

• A Python script that loads the pipeline
• A Gradio demo that comes with the project
• Windows, macOS or Linux for fetching the files

Use case | Who | What they get back
Ambience | Sound designers | Natural and urban environment sounds
Creature and action sounds | Game teams | Animal, creature and human action effects
Short clips | Creators | Short music-like or percussive clips

How to set it up: You need conda or another way to make a Python 3.12 environment, a CUDA GPU, and the project's code from the MOSS-TTS repository on GitHub.

• Python 3.12 in a clean environment
• A CUDA GPU
• The moss_soundeffect_v2 folder of the MOSS-TTS repository

1. Create and activate a clean Python 3.12 environment.
2. Get the MOSS-TTS repository from GitHub and open the moss_soundeffect_v2 folder.
3. Install it with pip, using the inference-only option if you do not plan to fine-tune.
4. Download the model files.

conda create -n moss-soundeffect-v2 python=3.12 -y
conda activate moss-soundeffect-v2

pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-SoundEffect-v2.0

How to use it: You load the pipeline once, then call it with a text prompt and a length in seconds. It returns a waveform that you save as a WAV file.

import torch
from moss_soundeffect_v2 import MossSoundEffectPipeline

pipe = MossSoundEffectPipeline.from_pretrained(
    "OpenMOSS-Team/MOSS-SoundEffect-v2.0",   # this repo, or a local dir
    torch_dtype=torch.bfloat16,
    device="cuda",
)

audio = pipe(
    prompt="A dog barking loudly in a park.",
    seconds=10,
    num_inference_steps=100,
    cfg_scale=4.0,
)                                            # (B, C, T) waveform tensor
pipe.save_audio(audio, "out.wav")

Setting | Default | What it controls
num_inference_steps | 100 | How many solver steps the sampler takes
cfg_scale | 4.0 | How strongly the prompt guides the result
sigma_shift | 5.0 | A scheduler shift applied on each call
seconds | 10.0 | The length of the clip, up to 30

Example: You write a prompt asking for rain on a metal roof and set the length to 10 seconds. The pipeline returns a waveform, and you save it as out.wav to listen to it.

For newcomers and experts: If you are new, run the sample script once with the default settings and a short prompt, then change the prompt and length. The first call can take a few minutes, because the model is compiled then.

If you know the field, the Diffusion Transformer is trained with a flow matching objective and sampled with a flow-match solver. The card notes it is wrapped with torch.compile and Triton CUDA Graph, and suggests setting TORCHDYNAMO_DISABLE=1 if compiler errors appear.

A text prompt is encoded by Qwen3
The Diffusion Transformer samples a latent
The audio VAE decodes it
A waveform is saved as WAV

Good to know: Clips run up to 30 seconds at 48 kHz, and the model was trained on both English and Chinese captions. The card points to a project README for the Gradio demo and fine-tuning recipes.

• License: apache-2.0
• Cost to run: none named, but it needs a CUDA GPU
• Data it reaches: files you download, and the prompts you write
• Not for: speech or voice generation, since it makes non-speech sound

On Hugging Face, 13 November 2026

• Task: text to audio
• Library: diffusers
• Parameters: 1.4B
• Weights: about 5.7 GB in F32; about 708 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 6,096
• Likes: 158
• License: apache-2.0
• Languages: en, zh
• Based on: Qwen/Qwen3-1.7B
• Quantized versions: 1
• Fine-tunes: 3
• Tags: diffusers, safetensors, text-to-audio, diffusion, flow-matching, sound-effects, audio-generation

Links

• Model: https://huggingface.co/OpenMOSS-Team/MOSS-SoundEffect-v2.0
• Files: https://huggingface.co/OpenMOSS-Team/MOSS-SoundEffect-v2.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-SoundEffect-v2.0

Then follow the usage in its model card. Nothing in this kit runs by itself.