---

Audio-Denoiser-ONNX

Audio denoising powered by ONNX Runtime for enhanced clarity.

Features

1. Supported Models:

   • ZipEnhancer
   • GTCRN
   • DFSMN
   • Mel-Band-Roformer
   • MossFormerGAN-SE-16K
   • MossFormer2-Super-Resolution

2. Dynamic Quantization:

   • Dynamic quantization is not recommended for Denoiser as it significantly reduces performance due to increased computational overhead. Except, Mel-Band-Roformer.

3. End-to-End Processing:

   • The solution includes internal STFT/ISTFT processing.
   • Input: Noisy audio
   • Output: Crystal-clear denoised audio

4. Resources:

   • Download Models
   • Explore More Projects

5. Note

   • Please note that the denoiser model optimized (opt_level=99) on Windows cannot be used on Linux, and vice versa.

---

性能 Performance

OS	Device	Backend	Model	Real-Time Factor (Chunk Size: 4000 or 250ms)
Ubuntu-24.04	Desktop	CPU i3-12300	ZipEnhancer f32	0.32
Ubuntu-24.04	Desktop	OpenVINO-CPU i3-12300	ZipEnhancer f32	0.25
macOS 15	MacBook Air	CPU M3	ZipEnhancer f32	0.25
Ubuntu-24.04	Desktop	CPU i3-12300	GTCRN f32	0.0036
macOS 15	MacBook Air	CPU M3	GTCRN f32	0.0013 ~ 0.0019
Ubuntu-24.04	Laptop	CPU i5-7300HQ	DFSMN f32	0.0068 ~ 0.012
Ubuntu-24.04	Laptop	CPU i7-1165G7	MelBandRofomer q8f32	1.40 (Chunk Size: 1000ms)
Ubuntu-24.04	Desktop	CPU i3-12300	MossFormerGAN_SE_16K f32	1.085
Ubuntu-24.04	Desktop	OpenVINO-CPU i3-12300	MossFormerGAN_SE_16K f32	0.95
Ubuntu-24.04	Desktop	CPU i3-12300	MossFormer2-SR f32	1.49

---

To-Do List

• Denoiser-MossFormer2-48K
• ExNet-BF-PF
• MP-SENet

---

Audio-Denoiser-ONNX

通过 ONNX Runtime 实现音频降噪，提升音质清晰度。

功能

1. 支持的模型：

   • ZipEnhancer
   • GTCRN
   • DFSMN
   • Mel-Band-Roformer
   • MossFormerGAN-SE-16K
   • MossFormer2-Super-Resolution

2. 动态量化：

   • 除了 Mel-Band-Roformer 之外，不建议对其餘 Denoiser 应用动态量化，因为这会由于计算负载增加而显著降低性能。

3. 端到端处理：

   • 解决方案内置 STFT/ISTFT 处理。
   • 输入：带噪音的音频
   • 输出：清晰无噪音的音频

4. 资源：

   • 下载模型
   • 探索更多项目

5. Note

   • 请注意，在 Windows 系统上优化的(opt_level=99)降噪模型无法在 Linux 系统上使用，反之亦然。

---