How to Isolate and Enhance Speech in FFmpeg

This article provides a quick guide on how to isolate, clean, and enhance spoken-word audio using FFmpeg’s powerful audio filtering engine. While FFmpeg does not feature a native filter literally named “freespoken,” you can easily achieve the goal of freeing spoken vocals from background noise by chaining together standard filters like arnndn, afftdn, speechnorm, and silenceremove. Below, you will find the exact commands and configurations needed to optimize spoken audio files.

Suppressing Background Noise with afftdn or arnndn

To isolate spoken voice, you must first eliminate background noise. FFmpeg offers two primary filters for this:

  1. afftdn (Audio FFT De-noise): This is a native, highly effective filter that uses Fast Fourier Transforms to reduce broadband noise.

    ffmpeg -i input.mp3 -af "afftdn=noise_reduction=12" output.mp3

  2. arnndn (Recurrent Neural Network Noise Reduction): This filter uses neural networks to target and remove non-speech noise, making it highly effective for spoken-word isolation. It requires a downloaded .rnnn model file.

    ffmpeg -i input.mp3 -af "arnndn=model=cb.rnnn" output.mp3

Standardizing Volume with speechnorm

Once the noise is reduced, the spoken vocals may still have uneven volume levels. The speechnorm (Speech Normalizer) filter dynamically adjusts the volume of speech to ensure that quiet words are audible and loud words do not clip.

ffmpeg -i input.mp3 -af "speechnorm=e=4:r=0.05" output.mp3

Eliminating Dead Air with silenceremove

To make spoken recordings sound more concise, you can automatically trim silence from the beginning, middle, or end of the track.

ffmpeg -i input.mp3 -af "silenceremove=start_threshold=-30dB:start_duration=0.5:stop_threshold=-30dB:stop_duration=0.5" output.mp3

Chaining the Filters for Best Results

For the cleanest spoken-word output, combine these filters into a single command chain. This process first removes noise, normalizes the voice levels, and then trims the silent gaps.

ffmpeg -i input.mp3 -af "afftdn=noise_reduction=15,speechnorm=e=4,silenceremove=stop_threshold=-35dB:stop_duration=1" output.mp3