Enhance speech: one click to a cleaner voice.

At a glance

What it does to your recording, step by step

Most rough recordings have the same four problems, and this fixes them in the order a sound engineer would. First, noise: an AI denoiser (RNNoise, trained by Xiph on thousands of hours of speech and noise) keeps the voice and pulls down everything else, including the irregular noise a classic noise gate cannot touch, such as traffic, keyboards and a fan that changes speed. Second, rumble: a high-pass filter at 80 Hz removes the low thuds of a desk, a car or a handled phone, which eat loudness without adding anything you can hear as voice. Third, tone: a small cut at 250 Hz takes out the boxy sound of a small room, and a lift at 3.5 kHz brings forward the consonants that make speech intelligible on phone speakers and earbuds. Fourth, level: a gentle compressor narrows the gap between the loud and quiet moments, then the whole file is brought to a standard loudness, with a true-peak limiter so it never clips.

Measured on a test recording with a speaker who drops about 12 dB for one sentence, noise bursts and clicks under the voice, and a 40 Hz rumble: the result landed at -16.1 LUFS with its true peak at -1.0 dBTP; the quiet sentence came up, so the gap between it and the loud ones went from 11.5 dB to 6.8 dB; the pauses came down 24 dB relative to the voice; and the rumble all but disappeared. The voice stays where it was to a third of a millisecond and the file is exactly as long as before, so a video stays in sync. A 13-second clip took about 8 seconds on a laptop.

Enhance speech vs Adobe Podcast Enhance

Adobe's Enhance Speech is a generative model: it listens to your recording and re-synthesises the voice as if it were recorded in a studio. That can sound remarkable, and it can also change the voice, invent consonants, or turn a rough take into a slightly different person; it needs an Adobe account and an upload, with a limit on the free tier. This page does the opposite: it keeps your real voice and cleans and polishes it with tools whose effect you can predict, on your own computer. A recording made in a bathroom will still sound like a bathroom, a little less so; a recording made in a reasonable room with a fan running and an uneven speaker will sound properly finished. If you need the studio regeneration, Adobe's tool is the one for that.

Getting the best result

Choose -16 LUFS for a podcast (Apple and Spotify both play speech at about that level) and -14 for YouTube. Leave AI noise removal on unless the recording is already clean, in which case untick it: on a studio voice it has nothing to remove and can only take a little air away. Enhance before you edit if the levels are all over the place, or after editing if you are assembling several takes, so they all end up at the same loudness. For a video, the picture is copied untouched and the sound stays in sync, so you can use the result straight away. To cut out long pauses as well, run the result through auto-cut silence; to transcribe it, audio to text reads a cleaned recording more accurately.

What it can't do

It is for speech. On music, the denoiser treats instruments as noise and the EQ is wrong for them; for a song, use the EQ and the loudness normalizer separately. Other people talking in the background are voices, so the denoiser keeps them. It does not remove echo or reverb from a big room, does not fix distortion from a microphone that was too close or too loud, and cannot bring back words that the noise drowned out completely. And it does not change a voice into a studio voice: for that kind of regeneration, see the comparison above.