Lightweight Noise Suppression for Audio Processing

rnnoise_light is a streamlined, minimalistic version of the Xiph RNNoise library, designed for efficient real-time noise suppression on embedded systems, desktop UIs, or constrained environments.

Unlike the full RNNoise implementation, this version reduces dependencies and computational overhead while retaining usable audio quality. For integration into Flutter or real-time streaming systems, see my other project: dgofman/audiostreamer.

---

🔊 Test Audio Samples

• 🌐 index.html
• ▶️ test_input.wav
• ▶️ test_output.wav
• ▶️ test_full_output.wav

---

🔍 What is RNN?

RNN (Recurrent Neural Network) is a type of artificial neural network designed for processing sequential data such as audio. It uses memory of previous inputs to inform future predictions, making it suitable for speech and noise modeling.

---

🎧 Test Setup

1. Input Audio

• Sample from YouTube:
  https://www.youtube.com/shorts/nRLKJq1eF0Q

• Dataset:
  Google's Noisy Speech Commands Dataset
  Includes samples like:

  speech_commands_v0.02/_background_noise_/running_tap.wav
  

2. Record Test Audio with Background Noise

Use ffmpeg to record your microphone with injected background noise:

ffmpeg -f dshow -i audio="{{MICROPHONE NAME}}" -t 13 -filter:a "volume=1.0" -ar 48000 -ac 1 test_input.wav

To list your available microphone devices:

ffmpeg -list_devices true -f dshow -i dummy

---

⚙️ Build Instructions

Step 1: Compile

make

Step 2: Run Noise Suppression

examples/rnnoise_demo.exe docs/test_input.wav docs/test_output.raw

Step 3: Convert Output to WAV

ffmpeg -f s16le -ar 48000 -ac 1 -i docs/test_output.raw docs/test_output.wav

---

💡 Notes

• The test_output.raw is 16-bit PCM (mono, 48 kHz).
• Noise suppression is performed using a lightweight RNN-based model.
• Reduced rnnoise_tables can be used for performance experiments.