---
library_name: audio.cpp
license: apache-2.0
pipeline_tag: text-to-audio
language:
- en
- zh
base_model:
- OpenMOSS-Team/MOSS-SoundEffect-v2.0
tags:
- audio.cpp
- gguf
- sound-effects
---

# MOSS-SoundEffect v2 GGUF

Follow or star [audio.cpp](https://github.com/0xShug0/audio.cpp) for the native
inference code and model updates.

Generate sound effects from English or Chinese descriptions, with 48 kHz mono
output. Each package contains the Qwen3 text encoder, audio diffusion transformer,
continuous DAC decoder, tokenizer, and embedded model specification.

The upstream checkpoint is
[OpenMOSS-Team/MOSS-SoundEffect-v2.0](https://huggingface.co/OpenMOSS-Team/MOSS-SoundEffect-v2.0/tree/e35df4d82fbe87fcd5d14e5d100e349c0c3c076d),
pinned at `e35df4d82fbe87fcd5d14e5d100e349c0c3c076d`.
The implementation reference is
[OpenMOSS/MOSS-TTS](https://github.com/OpenMOSS/MOSS-TTS/tree/4cf2dab88fc8711aef93089b3b75c0ae07373e47/moss_soundeffect_v2).

| File | Transformer weights | DAC decoder |
|---|---|---|
| `moss-soundeffect-v2-q8_0.gguf` | Q8_0 matrices, F16 embeddings | F32 |
| `moss-soundeffect-v2-bf16.gguf` | BF16 matrices | F32 |

Timestep embedding, modulation and normalization retain F32, matching upstream.
The runtime family is `moss_soundeffect`; these packages implement v2.

## Performance

RTX 5090, CUDA, eight CPU threads, offline generation after a short warmup.
Mean of two different 10-second outputs, using 100 Euler steps, guidance 4,
shift 5, and the trained 30-second latent window. Python uses the official
BF16 eager pipeline. Peak VRAM is observed process memory.

| Runtime | RTF | Peak VRAM (MiB) |
|---|---:|---:|
| Official Python BF16 | 0.692 | 18,586 |
| audio.cpp BF16 | 0.999 | 11,347 |
| audio.cpp Q8_0 | 0.849 | 8,853 |

Q8_0 reduces peak VRAM by about 52% versus Python in this comparison;
native generation is currently slower. Lower RTF is faster.

## Usage

```bash
audiocpp_cli --model moss-soundeffect-v2-q8_0.gguf \
  --family moss_soundeffect --task gen --backend cuda \
  --text "Steady rain pattering on a wooden roof, with distant thunder." \
  --request-option duration_sec=10 --out rain.wav --log
```

Output duration is 0.1–30 seconds, rounded to one decimal. The model always
denoises its trained 30-second window and crops the resulting waveform.
The default is 100 Euler steps, guidance scale 4, and seed 0.
It supports offline generation, not streaming.

See the [model documentation](https://github.com/0xShug0/audio.cpp/blob/main/docs/models/moss_soundeffect.md)
for request options and conversion instructions. Model weights and upstream code
are Apache-2.0 licensed.