In this guide
Noise-profile reduction learns a steady background from a noise-only section. Source isolation estimates a named source from a mix. A small edit may be enough when a distraction does not overlap the audio you need.
Choose an audio cleanup method
This chooser uses your description. It does not listen to a file or predict the quality of a result.
Choose the closest description above, then test the suggested method on a short, representative excerpt. The comparison below explains the tradeoffs.
Audio isolation and noise reduction can both make a recording easier to use, but they answer different questions. Are you trying to make a steady background quieter, or do you need one recognizable sound out of a mix? The answer changes what you should test and what might be lost.
“Noise reduction” covers several methods, including learned speech denoisers. This guide compares noise-profile reduction with text-guided source isolation, gates, and small repairs. It is a method-selection guide, not a benchmark claiming that one product wins on every recording.
The difference between audio isolation and noise reduction
A noise-profile tool uses a sample of the unwanted background to guide reduction across the recording. Audacity's Noise Reduction documentation describes this workflow for relatively constant noise and requires a section containing only that noise. If your recording is a voice plus a steady fan, that can be a useful starting point.
Source isolation asks for a source: “spoken voice,” “vocals,” or “acoustic guitar.” A separator estimates that source from the complete mix. Meta's SAM Audio research describes general separation controlled by prompts. Here, we focus on the text-guided workflow available in Saturalabs.
The word “AI” does not settle the choice. RNNoise, for example, combines machine learning with signal processing for noise suppression. Choose by what the method does to your recording, the outputs it provides, and whether those outputs meet your editing goal.
Choose a method for the problem you hear
These are starting points to test on a representative excerpt. They are not guarantees that a particular sound can be removed cleanly.
| Problem | First test | The deciding check |
|---|---|---|
| Steady fan or hiss | Noise-profile reduction, if a matching noise-only section exists | Quiet words and sustained notes remain natural |
| Music over dialogue | Voice isolation when you need only speech; music separation when other effects matter | The full sentence and wanted effects survive the overlap |
| Typing or traffic over speech | A targeted separation or voice-focused separation | Consonants remain intact during the loudest interruption |
| Noise in pauses | A small edit or gentle gate | Breaths, word endings, and room tone do not jump |
| One tiny damaged click | A local repair in an editor | The edit is unobtrusive when heard with surrounding audio |
| Distortion inside the wanted sound | A cleaner original or a dedicated repair test | The wanted sound itself improves, not only the background |
Audacity's repair overview distinguishes profile reduction, threshold-based gates, click removal, and short-fragment repair. They solve different problems; treating each as a general “clean audio” button makes the tradeoffs harder to notice.
When a noise profile is a useful first test
Find a section where the unwanted background is audible without the wanted voice or music. Use a background that actually matches the difficult passage. A quiet room at the beginning is a poor reference for a loud vehicle arriving later.
Apply a restrained test to a section containing both wanted sound and noise. Listen to the reduced version and, if the editor provides it, the removed residue. Recognizable pieces of the wanted sound in that residue are a reason to reconsider the settings. Audacity explains this check in its noise-profile workflow.
Do not judge success by the pause alone. In an interview, a quiet pause is useful only if the next sentence still sounds complete. In music, a lower noise floor is less helpful if it makes a sustained note flutter. Save the unprocessed excerpt so the comparison remains repeatable.
When source isolation matches your goal
Use a source-focused test when you can name the sound you need, or a distinct unwanted source that overlaps it. Before running a separation, decide whether the deliverable is a voice alone, a backing track, or a more complete soundtrack with one distraction reduced.
Saturalabs produces two audio tracks for a target description. Isolated sound contains the requested source; Background audio contains the remainder. The latter label does not mean “remove background noise.” It describes the mix remaining after your requested source is extracted.
| Editing goal | Example target | Output to inspect |
|---|---|---|
| Keep interview speech | Spoken voice | Isolated sound |
| Reduce music while retaining other soundtrack elements | The background music | Background audio |
| Reduce typing while keeping the remaining recording | Keyboard typing | Background audio |
| Create a karaoke backing track | Vocals | Background audio |
For example, a cooking video might need speech, chopping sounds, and the clink of a pan. Isolating speech alone can be useful for intelligibility but unsuitable for that full soundtrack. Test separating the music and inspect the remainder if those other sounds matter. Check each wanted effect where it overlaps the music.
For an interview where the words are the priority, start with the voice isolator. For typing, the keyboard noise workflow explains which output to compare. AI separation is a separate service from the free preparation tools; the product shows account and download credit details.
When a small edit is the better starting point
If a chair squeak happens entirely between sentences, you may not need to process the spoken words at all. Try a local edit with natural surrounding room tone. A gentle gate can also reduce low-level gaps, but it does not remove overlapping noise while the wanted sound is playing. Audacity's alternative techniques guide explains that limitation.
For a very short damaged fragment, a repair tool may estimate a replacement from nearby audio. Audacity's Repair documentation limits this effect to 128 samples, not a whole word or a long burst of typing. That distinction matters when deciding whether a local repair is even applicable.
Do not confuse trimming with removing an interruption from the middle. The free Saturalabs trimmer keeps one continuous excerpt. A multitrack editor gives you the control needed to join sections or replace a gap with consistent ambience.
A listening test that catches the tradeoff
- Choose the hard overlap. Include a whole sentence or musical phrase, not only the cleanest moment. Use the preparation guide to retain context and the original timeline offset.
- Write down what must survive. For dialogue, that might be every word and breath needed for the delivery. For a backing track, it might be drum cues and chord changes.
- Compare at a similar listening level. A louder result can appear more present without resolving the actual problem. Alternate with the original excerpt and avoid changing several settings at once.
- Check continuity. Listen before, during, and after the interruption. Check soft starts, long tails, and transitions rather than rewarding complete silence between phrases.
- Use the final context. Put the selected WAV into the edit or practice session. A flaw that is acceptable under a soundtrack may be distracting when the sound is exposed on its own.
Keep the first attempt so later versions have a fixed reference. A simple note—original file, excerpt times, target, chosen output, and the problem you still hear—makes the next decision more useful than repeatedly processing the latest result.
If the chosen clip needs an overall level adjustment for an edit, the free peak normalizer applies one gain locally. Read the peak and loudness normalization guide before using equal peaks as a comparison: they do not ensure equal perceived loudness, and raising gain raises remaining noise too.
When cleanup cannot solve the recording problem
If the wanted sound is already clipped or heavily obscured, separating it does not guarantee that its missing detail will be reconstructed. Look for another microphone, an original track, or a different take. A dedicated repair may be worth a restrained test for light damage: Audacity describes Clip Fix as most likely to help lightly clipped audio.
Likewise, a voice alone is not automatically a natural voice. Keep some unobtrusive room sound when a quieter version makes the delivery hollow or incomplete. The useful endpoint is an edit that communicates what your audience needs, not the largest possible amount of removed audio.
Common questions
Does a noise gate remove background music from speech?
A gate can make quiet gaps less audible. It cannot independently remove music while speech above the threshold is passing through. If music overlaps the words, evaluate a source-focused separation instead.
Should I isolate the voice or isolate the music?
Request the voice and keep Isolated sound when you need the voice alone. Request music and inspect Background audio when other non-musical soundtrack elements should remain. Compare what each method changes before choosing.
Why does a quieter result sound worse?
The result may have lost wanted sound or changed its continuity. Compare the word endings, attacks, and tails at a similar level, then choose the version that serves the edit rather than the one with the quietest pauses.
Can I use noise reduction after source isolation?
Consider it only if the isolated track has a remaining problem suited to that method. Save the first separation and test on a copy. More processing steps make it harder to trace damage, so keep each stage available for comparison.
Is the chooser analyzing my audio?
No. It maps the problem you select to a possible starting method. You do not upload a file to the chooser, and it does not score a recording or guarantee a result.