Tutorials Audio

Cleaning Up Dialogue with AI: Tools, Limits and a Safe Workflow

By Infinite Creation · Intermediate · ~14 min

A 1960s mid-century comic book style editorial illustration showing a sound editor at a desk with headphones, comparing a cluttered speech waveform with a clean one on a monitor while a colleague holds a boom microphone.

Overview

Dialogue cleanup means taking a recording of speech and reducing the noise, room echo and damage around the voice. AI tools now do this in one pass, where older tools asked you to capture a noise sample and tune several controls. This guide is for anyone with a recording that sounds worse than hoped. It explains what the main tools state they do, what kinds of damage each one targets, where the approach fails, and a workflow that never overwrites your original file. We did not test any tool on this page. Each statement about a tool comes from the maker's own page as read on 8 October 2026, and the Sources section says what each page established.

What You Need

One recording of speech you hold the rights to, saved as a lossless file if you have it
Headphones and, if possible, a second listening device such as laptop speakers
A free tool to start with, and a plug-in host or editor if you move to a plug-in
A way to compare before and after at the same loudness

Steps

Copy the original and never touch it again

Make a duplicate and work on the duplicate. Neural cleanup tools rebuild the signal rather than only turning it down. Some regenerate the voice. Descript's page says Studio Sound "isolates the voice and regenerates the audio", which removes echoes and noise but means the output is no longer your original sound. If a later tool improves, you will want the untouched file.

Name the damage before you pick a tool

Listen on headphones and write down which problem you hear. Steady hiss or fan noise. Room echo. Clipping, where loud peaks flattened against the limit of the recording. A thin, telephone-like sound. Tools target different problems. Accentize states that dxRevive Pro handles noise, reverb, clipped audio and band-limited audio such as phone recordings. Antinode Clear, formerly Supertone Clear, offers three controls for voice, ambience and voice reverb.

Pick the lightest tool that fits

Start with a free option. Audacity's OpenVINO noise suppression is a Windows download and, per its page, runs fully on your computer. Move to a plug-in when you need control. Waves describes Clarity Vx Pro as real-time voice noise reduction powered by neural networks. iZotope lists Dialogue Isolate in RX 12 with neural nets for offline and real-time modes. Cloud tools mean uploading your file, which matters for unreleased work.

Process gently, then compare at matched loudness

Set the effect lower than feels necessary. A louder version always sounds better, so match the loudness of the before and after versions before you judge. Switch between them while the same sentence plays. If the cleaned version sounds cleaner only because it is louder, you are hearing level and not quality.

Listen for the damage the tool causes

Aggressive settings leave artefacts. Listen to consonants such as s, t and k, which turn watery or vanish. Listen to the ends of words, where the tool sometimes cuts the decay. Play the result on laptop speakers as well as headphones. Back the amount off until the voice sounds like the same person in the same room.

Leave a little noise on purpose

A completely silent gap between words sounds unnatural and exposes edits. Keep a low room tone, or add a recorded bed of the room underneath. Then render a new file with a new name and keep the original beside it.

Pro Tips

Cleanup repairs a recording. A closer microphone or a quieter room fixes the cause. The field recording basics page covers that side.
Two light passes with different tools usually sound better than one heavy pass.
Do not run a restoration tool twice with the same setting on an already processed file. The artefacts stack.

Knowledge Base

The Longer Answer

Two ideas explain most of the results you will hear: what the tool is rebuilding, and what it cannot separate.

Reducing noise and rebuilding speech are different jobs

Classic noise reduction measures the noise and subtracts it. The site's own free denoiser works this way, using spectral subtraction, and it is not an AI tool. Neural tools learn what speech looks like and keep what matches. Some go further and rebuild missing parts of the voice. That is why a badly damaged phone recording improves more with a restoration tool, and why a rebuilt voice should be checked against the original before it ships.

What these tools do not separate well

A tool trained on speech struggles when the unwanted sound looks like speech. Overlapping voices, a television playing in the room, and music under a voice are the usual failures. The tool either removes some of the wanted voice or keeps some of the unwanted one. For those recordings the practical fix is to re-record, or to separate the sources by editing before any cleanup runs.

Where This Fits

Cleanup repairs a recording after the fact, and Audio Recording Fundamentals covers the microphone distance, room and gain choices that decide how much repair is needed.

FAQ

Q: Which tool is best?
A: We did not test them, so we cannot rank them. The makers' pages state different targets, listed above. One independent shootout we found rated newer plug-ins close to or ahead of RX for plain noise removal, and we have not been able to read it in full.

Q: Is a free tool enough?
A: For steady noise on a clear recording, often yes. Free tools give you less control and fewer repairs for clipping or heavy reverb.

Q: Do I need to tell my audience the audio was processed?
A: Cleanup that leaves the words and the voice unchanged is ordinary post-production. A tool that regenerates or rebuilds a voice is a different case. See the labelling page for current practice.

Vocabulary

Artefact A sound produced by the processing and not present in the recording. In speech cleanup it is often a watery or metallic quality on consonants. A tool set too hard creates artefacts faster than it removes noise.

Clipping Distortion caused when a signal is louder than a recording system handles, so the tops of the waveform are flattened. It sounds harsh and fuzzy. Some restoration tools rebuild the missing peaks, and none restore them exactly.

Room tone The quiet sound of a space with nobody speaking, recorded for a minute or so. Editors lay it under dialogue to fill gaps left by cleanup so the silence does not sound artificial.

Spectral subtraction A noise reduction method that measures the frequencies in a noise sample and subtracts them from the whole recording. It works on steady noise and leaves a swirling sound when set too high. It is a classic method, not an AI method.

Sources

Each maker's own page was read on 8 October 2026. Nothing was tested. Where a page was unreachable, the caveat below says so. Checked 8 October 2026.

Descript, Studio Sound. States Studio Sound removes background noise and enhances the voice, isolating the voice and regenerating the audio. It does not state where processing runs.
Accentize, dxRevive Pro. States it restores noisy, clipped, reverberant and band-limited speech with neural networks, processes locally without sending data to the cloud, and ships as VST3, AU and AAX for Mac and Windows.
Antinode Audio, Clear (formerly Supertone Clear). States three controls for voice, ambience and voice reverb, real-time processing, and Windows and macOS support in AU, VST3, VST and AAX formats.
Waves, Clarity Vx Pro. States real-time noise reduction for voice, powered by Waves Neural Networks.
iZotope, RX. Lists RX 12 with Dialogue Isolate using neural nets in offline and real-time modes, and De-reverb.
Audacity, OpenVINO AI effects. Lists noise suppression among effects that run locally. The download is Windows only.

Adobe's Enhance Speech page returned only a title and no product text, so no claim about it is made here. CEDAR's DNS page did not say whether its products use machine learning. The shootout mentioned in the FAQ was not read in full. iZotope's own page does not list platforms, and MusicTech reported on 2 July 2026 that Boris FX acquired iZotope with licences remaining active.