Denoise Audio with FFmpeg arnndn Filter

This guide explains how to use the arnndn (Active Recurrent Neural Network Denoise) filter in FFmpeg to remove background noise from audio tracks. You will learn how the filter works, where to obtain the required RNN model files, and the exact command-line syntax needed to clean up your audio and video files effectively.

What is the arnndn Filter?

The arnndn filter is a powerful, neural-network-based noise reduction tool built into FFmpeg. Based on the RNNoise project, it uses recurrent neural networks to distinguish between human speech and background noise (like wind, system hums, or computer fans). Unlike traditional noise gates or spectral subtraction filters, arnndn does not require you to select a “noise profile” from a silent portion of your audio; instead, it recognizes and filters out noise in real time.

Step 1: Obtain the RNN Model Files

To use the arnndn filter, you must provide it with a trained RNN model file (typically ending in .rnn). FFmpeg does not include these models by default.

You can download pre-trained models from public repositories hosting RNNoise models (such as the official RNNoise GitHub repository or community-shared configurations). Common model files include: * bd.rnn (highly recommended for general speech and vocal clarity) * cb.rnn * mp.rnn * sh.rnn

Download your preferred model file and save it in your working directory or a known system path.

Step 2: Basic Command Syntax

Once you have the model file, you can apply the filter using the -af (audio filter) flag in FFmpeg.

Use the following basic command to denoise an audio file:

ffmpeg -i input.wav -af "arnndn=model=bd.rnn" output.wav

In this command: * -i input.wav specifies your noisy input file. * -af "arnndn=model=bd.rnn" applies the audio filter and points to the location of your downloaded RNN model. * output.wav is the resulting cleaned audio file.

Step 3: Denoising Audio in a Video File

If you want to remove background noise from a video’s audio track without re-encoding the video stream, you can combine the arnndn filter with video stream copying. This process saves time and preserves the original video quality:

ffmpeg -i input.mp4 -c:v copy -af "arnndn=model=bd.rnn" output.mp4

• -c:v copy instructs FFmpeg to copy the video stream directly without re-encoding.
• The audio stream is extracted, processed through the arnndn filter using the bd.rnn model, re-encoded, and muxed back into the output MP4 container.

Best Practices for Optimal Results

• Model Selection: Different models yield different results depending on the speaker’s voice and the type of background noise. If your output sounds muffled or robotic, try switching to a different model (such as cb.rnn).
• Sample Rate: The underlying RNNoise algorithm is designed to work at a sample rate of 48,000 Hz (48kHz). If your input audio is at a different sample rate (like 44.1kHz), FFmpeg will automatically resample it, which might slightly alter the tone. For the best quality, try to feed 48kHz audio into the filter.