The Complete Guide to Cleaning Up Spoken Audio

Clean speech is not the same as silent speech. Push noise removal too hard and the voice stops sounding like a person. This guide covers the whole job in order: what the noise actually is, what to leave to the recording stage, how to clean without damaging the voice, the loudness target each platform expects, and how to check the result before you publish.

The steps use Ampliflare Audio, which purplecandy makes: a desktop app that removes noise, evens out loudness, cuts, splits and converts speech files on your computer. Nothing is uploaded. The ideas carry over to any editor, because the checks and the targets are the same.

The kinds of noise you are actually removing

Group noise by how it behaves, because each kind responds to a different fix.

• Steady noise. A laptop fan, an air conditioner, a fridge, a hiss. It holds roughly the same level across the recording, which makes it the easiest kind to remove: a tool can learn its shape from any pause. Audacity’s Noise Reduction manual says its effect suits constant background sound and is a poor fit for a one-off bang.
• Hum and low rumble. A steady low buzz from electrical or grounding issues, traffic rolling past, footsteps through a floor. It sits under the voice in the low frequencies.
• Intermittent sounds. A chair creak, a door, a click, a cough, a phone buzzing. These last a moment and land at a specific time, so they need a different tool from steady noise. iZotope’s repair documentation separates steady noise from intermittent noise for exactly this reason.
• Room and echo. Hard walls and bare floors add a short tail to every word. Ampliflare Audio’s model is built to handle room noise alongside fans, hum and traffic, per its product page. A long echo is harder, because it is the voice itself arriving late.
• Distortion and clipping. A loud peak that was cut off while recording. This is not noise, so noise removal cannot repair it.

Keep this list in mind, because the fix follows the kind.

Fix the noise in the recording when you can

The cheapest cleanup is the one you never have to do. A minute spent on setup saves several minutes in post, and some problems cannot be fixed later at all.

The podcast recording checklist has the full list. The short version:

• Turn off anything that hums, including fans, air conditioning and notifications.
• Record in the smallest, softest room you have. Carpet, curtains and a full bookshelf cut echo.
• Put the mic a hand’s width from your mouth, slightly off-axis. A close mic raises the voice and widens the gap between voice and noise.
• Record ten seconds of room tone first. A quiet stretch with no speech is useful if you later need a noise sample.
• Keep peaks well below clipping. You can make a quiet recording louder; you cannot rebuild a clipped one.

A good take with a little noise beats a clean edit of a bad take. Cleaning cannot add detail the microphone never captured.

Clean the noise without hurting the voice

Once the recording is done, clean it before you touch anything else, and before you raise the level. Fixing loudness first boosts the noise along with the voice.

A speech-trained model or a profile-based effect then removes the steady noise. The trade-off is the same in every tool: the harder you push, the more it can take from the voice. Audacity’s manual warns that too much reduction damages the sound you want, and that an unrepresentative noise profile leaves short random tones behind. iZotope’s Spectral De-noise docs describe chirpy or watery artifacts from heavy spectral subtraction.

Aim for the lowest amount that makes the recording useful, not for total silence. In Ampliflare Audio the Strength slider controls how much of the original room stays. The cleanup guide walks through the full sequence, and why noise reduction makes a voice sound robotic covers the artifacts and how to compare against the original.

Different jobs need different tools. A folder of speech files suits a one-pass cleaner; a single click over one word suits a spectral repair tool. The offline noise removal apps list and the voice noise removal picks compare the options.

Set the loudness for where it plays

Cleaning fixes the noise. It does not fix the level. Every platform normalizes to a target loudness, so a file far above the target gets turned down, and one far below it sounds weak next to everything else. Loudness is measured in LUFS, and the LUFS explainer covers the unit and the measurement standard behind it.

The targets that matter most for speech:

Apple’s audio requirements recommend that overall loudness sit around -16 dB LKFS with a plus or minus 1 dB tolerance, and that the true-peak value not exceed -1 dBFS. Spotify normalizes tracks to -14 dB LUFS according to the ITU 1770 standard. EBU R 128 sets the Programme Loudness Level at -23.0 LUFS with a true-peak ceiling of -1 dBTP.

LUFS and LKFS are the same unit under different names, both from the ITU-R BS.1770 loudness measurement, which Apple’s page names as the basis for its numbers. Audiobooks use a different shape of specification, an average level with a peak ceiling rather than a single LUFS figure, so check the retailer’s requirements before you master for one.

Fix levels after cleaning, not before. In Ampliflare Audio the Enhance tool sets loudness for a destination: music streaming, podcasts, audiobooks or broadcast. It also drops low rumble.

Check the result before you publish

Do not trust a waveform or a green checkmark. Listen.

1. Compare with the original. Play the same sentence from the untouched file and the cleaned file. Pick a passage with quiet words and a pause, so you hear both the voice and the background.
2. Listen for damage. On a held vowel, listen for a warble, a hollow tone, or a watery edge. In the pauses, listen for new chirps where there was steady noise.
3. Check the words. If any word is harder to follow after cleanup, the setting is too strong.
4. Check the noise floor. In a pause, the level should sit quietly under the speech, not rise and fall with it.
5. Check the loudness against the target for the destination, with true peak below the ceiling.
6. Listen to the first and last minute of the finished file, on headphones and on a phone speaker.

If you kept the original, you can always go back and try a lighter setting. Ampliflare Audio writes results to new files and never changes the input, so each run keeps the original as its reference.

A repeatable cleanup workflow

The order matters. Here is the sequence the cleanup guide uses, gathered in one place:

1. Back up the raw files before you touch anything.
2. Remove noise from the originals, using the lowest Strength that clears the noise.
3. Fix loudness on the cleaned files, to the target for the destination.
4. Cut and arrange, so you trim clean, correctly leveled audio.
5. Convert to the format the host wants, usually mp3 for podcasts.
6. Check the result, then publish.

In Ampliflare Audio the first steps are in the getting-started guide: drop the files on the window, pick Remove noise, choose where results go, then click Run. Version 1.1.0 added a play button and seek bar on every file and an Activity log, so you can hear the input, run the job, and play the output without leaving the app. The other tools take the whole list; Cut and Split work on one file at a time.

For a wider choice that includes transcript editors and hosted services, see the podcast audio cleanup apps.

When the source is the problem

Some recordings cannot be rescued by cleaning. It is worth knowing that before you spend an hour on them.

• Clipping and distortion. Noise removal takes noise out. It cannot rebuild a waveform that was flattened while recording.
• Heavy echo in a hard room. A model can reduce room sound, but it cannot fully undo a long reflection.
• Noise that moves. Steady noise is easier than noise that changes level or frequency mid-sentence.
• Only the processed file. Without the untouched recording you cannot compare, so you cannot tell whether cleanup caused a problem.

For a single click or creak, a spectral repair tool is the better match. The Audacity and RX comparison for audio repair explains the difference. When you are weighing hosted finishing services, Auphonic vs Descript for podcast cleanup and Ampliflare vs Auphonic lay out the trade-offs. For the browser and free-tool routes, see the Audacity vs Adobe Podcast comparison.

Takeaways

• Group noise by behaviour: steady, low rumble, intermittent, room echo, or distortion, because each kind needs a different fix.
• Fix what you can before you record. A quiet room and a close mic beat any amount of cleanup.
• Clean before you raise the level, and use the lowest reduction that keeps the words clear.
• Match the loudness to the destination: about -16 LKFS for podcasts, -14 LUFS for music streaming, -23 LUFS for broadcast.
• Always compare the finished file with the original at the same sentence before you publish.