How to Remove Background Noise From Audio With AI—Without Making the Voice Sound Robotic

This guide focuses on spoken recordings. It covers music that interferes with speech, but restoring a musical performance or preserving a stereo soundscape needs different decisions. Start with the file-cleanup steps below; the live-call section is for noise you want to prevent during a call.

Identify the sound before choosing a tool

Listen to a passage where the problem occurs during speech, then to a pause. Write down what should remain: one speaker, everyone in the conversation, music, laughter, or the sound of the location. That decision determines whether you need gentle noise reduction, speech isolation, or separate control over sound layers.

A noise gate can reduce low-level sound between phrases, but it does not separate a fan from speech while the speaker is talking. If pauses are silent and every sentence still sounds noisy, evaluate a denoiser on the sentences themselves.

Choose a cleanup path that fits your recording

For one spoken recording, Adobe Podcast gives you an upload-and-download starting point. If you need free local control over a steady fan, use the Audacity steps below. If you already edit in Descript or Resolve Studio, try that editor's cleanup before moving the recording elsewhere. The commercial tools below are described from their documentation; the downloadable audio examples were processed with DeepFilterNet3 only.

Keep an untouched copy of the original. Use separate names for the test and approved output so you can return to the recording before processing. If the source contains separate speaker tracks, keep them separate through cleanup.

Adobe Podcast Enhance Speech for an uploaded file

Adobe Podcast provides a browser workflow for spoken audio. Its plan comparison lists audio-only uploads and no strength adjustment on the free plan. Premium adds video and strength adjustment. If the free output sounds robotic, try another tool from the original; there is no free strength slider to find.

For a phone recording, save or share the original file from your recorder. If you have a video and your plan accepts only audio, export its sound from a video editor first.

Suppose you have a phone interview with a constant fan in the background and a guest who speaks more quietly than the host. Work through that file as follows:

1. Keep the original and make a 20 to 30-second excerpt that includes the guest's quiet answer, the host's voice, and a pause. If you need an editor to make that excerpt, drag the file into Audacity, select the passage, and use File > Export Audio. Choose Export selected audio in the current dialog, or Current selection in older versions, and save a WAV named interview-test-original.wav. Audacity's export guide explains the selection option. Saving an Audacity project alone does not produce an audio file you can upload.
2. Upload that excerpt to Adobe Enhance Speech. Keep the full interview out of this first attempt so a poor result costs less time to inspect and replace.
3. Download the result. Play the same quiet answer in the original and cleaned files, one at a time, at similar voice volume. Listen for the complete last word of the answer and any name or number in it. Then check the host's louder sentence. A pleasant host voice does not make up for a missing guest syllable.
4. If the guest sounds watery or a word ending disappears, reduce the strength on Premium and repeat that comparison. On the free plan, try a different tool from the original; the Audacity route below is an option for this steady fan. Do not run the damaged output through another enhancer.
5. If both speakers remain understandable and the fan is less distracting during their sentences, process the full original with the chosen tool and settings. If no version improves the speech, keep the source and investigate another microphone track or a new take.
6. Save the full result separately, for example as interview-cleaned.wav if the download is WAV. Listen through it before using it. Check a quiet answer later in the interview too, since one good excerpt does not cover a guest moving away from the phone.
7. For video, import the cleaned audio into the original timeline, mute the original soundtrack, and check timing near the start and end before exporting. Leave any separate music track out of the speech-cleanup upload.

Adobe's March 2026 update documents Premium controls for speech, background noise, and music, plus downloads of those separated tracks, called stems. For an interview filmed in a café, lower distracting background sound while keeping enough location sound to fit the picture. Listen to the speech stem during words as well as pauses; separation can leave residual sound or alter speech.

ElevenLabs Voice Isolator for speech extraction

The Voice Isolator documentation describes speech extraction from audio and video, with a maximum file size of 500 MB and duration of one hour. It also says the tool is not specifically optimized for isolating sung vocals from music.

Upload or record audio, select Isolate voice, preview, and download. Test the excerpt before using your account's credits on the whole file. Compare the download with the source, paying particular attention to quiet speech and competing voices.

Choose this route when a speech track is the intended output. For a music lesson, performance, or family video, first decide which non-speech sounds belong in the result. If the isolated version removes something you need and offers no suitable adjustment, return to the original and use a workflow with source controls. Keep residual music in mind when checking the exported file; an isolation pass is not a guarantee that every trace is gone.

Descript Studio Sound inside an editing project

Descript's Studio Sound guide documents noise and echo reduction and an Intensity slider. It requires internet access and affects every instance of the source file across the project.

Select the script layer, open Properties, and enable Studio Sound under Audio Effects. For separate speaker tracks, select the track in the Sequence Editor. Compare the original with moderate Intensity before making transcript edits.

To clean one section, duplicate it into a new composition, use Flatten tracks to create a new project file, apply Studio Sound, then replace the original section. Duplicating a composition alone does not create an independent source file.

If speech becomes silent or heavily suppressed, reduce Intensity or disable the effect. Verify a short export before exporting the full project or passing a timeline to another editor.

DaVinci Resolve Studio for dialogue in a video timeline

Blackmagic lists AI Voice Isolation in DaVinci Resolve Studio. Its Fairlight training guide demonstrates the Amount control and reducing it to retain background sound. The guide uses Resolve 18; interface placement can differ in later versions.

Duplicate the timeline for comparison. Select the intended dialogue clip or track, enable Voice Isolation in the available Inspector or Fairlight controls, and loop a noisy sentence. Bypass the effect to compare, then adjust Amount while checking a quiet sentence too.

Track processing affects the material routed through that track. If only one passage needs strong isolation, separate it for local treatment instead of applying the same strength to all dialogue. Leave music and effects on their intended paths. Render a short video test and check speech timing against the picture before exporting the full timeline.

Build a useful test excerpt

About 20 to 30 seconds is usually a manageable excerpt for comparing settings. Include normal speech, a quiet phrase, the unwanted sound, and a pause. Add a name, number, or negation if those details matter in the recording.

One excerpt may not represent a long interview. If the room or microphone changes halfway through, make another sample from that section. A café interruption and a quiet home recording can need different settings even when they belong to the same project.

In an editor, copy the passage with a little audio before and after it. Keep the copy's start time if it will return to a video timeline. Do not trim only the loudest noise: you need to hear how the processor handles the nearby speech and transitions.

If strength adjustment is available, make two or three versions. Label them with the actual values, such as test-intensity-30 and test-intensity-50. For a tool with no adjustment, compare its output with the source and, if needed, another tool's output. Percentages are not comparable across products: Descript's Intensity, Resolve's Amount, and DeepFilterNet's attenuation limit describe different controls.

Keep EQ, volume adjustment, and other effects fixed while changing denoising. Once you can hear what that adjustment does, tackle the next problem separately.

Handle the noise your recording actually contains

Steady fan noise, hiss, and air conditioning

For a voice that is already understandable, test mild reduction first. Listen to the ends of words and to breathy sounds such as "s" and "f." If those become watery while the fan gets quieter, reduce the processing and compare again.

Audacity offers a local alternative when the noise is steady. You need a pause with the same fan or hiss as the speech section, without a voice or wanted music. Its Noise Reduction instructions recommend at least half a second; a longer pause is easier to select accurately.

1. Open a working copy of the recording in Audacity. Select that noise-only pause and choose Effect > Noise removal and repair > Noise reduction > Get noise profile. Older versions may place Noise Reduction directly in the Effect menu.
2. Select a short passage containing quiet speech and the fan. Reopen Noise Reduction. For a first test, set Noise reduction to 6 dB, Sensitivity to 6, and Frequency smoothing to 6 bands, the defaults listed in the current manual. These are starting values, not settings that suit every voice.
3. Select Noise only, called Residue in the Audacity 3 dialog, and apply it to that short passage. Listen to the sound the filter proposes to discard. If you can understand words in it, undo, lower Sensitivity, and try again. If your noise profile included speech, capture a better pause before continuing.
4. Undo the Noise only test. Reopen the effect with the same selection and choose Audio with noise removed, called Reduce in some older versions. Apply it and listen to the complete sentence. Undo if the voice becomes watery; lower Noise reduction and repeat. Leaving some fan noise is preferable to losing the last consonant of an answer.
5. Once the short test sounds useful, undo it before selecting the full passage or track and applying those settings once. Applying to the full track without undoing would process the test section twice. Use a new profile for a section recorded with different background noise.
6. Use File > Export Audio to save a separate WAV or FLAC, checking that the export covers the full recording you intended to clean. Play that file outside Audacity before using it.

If no pause contains noise alone, do not capture a profile from a spoken sentence. Try an AI speech cleaner on an excerpt instead. A noise profile suits continuous sound; a door slam needs separate treatment.

Electrical hum or a high-pitched whistle

A pitched buzz calls for a more specific diagnosis than broad enhancement. In Audacity, select a representative passage and open Analyze > Plot Spectrum. Look for a prominent narrow peak rather than assuming that every buzz is mains hum.

Audacity's Notch Filter guide describes targeting hum around 50 or 60 Hz and checking its harmonics. Apply a notch at the measured frequency, compare the voice, and inspect remaining peaks before adding more cuts. Higher Q makes the notch narrower; the guide gives Q values of 2 to 10 as a useful range for mains hum, not a preset for every recording.

If the voice becomes thin, undo and narrow or reduce the treatment. Avoid cutting a large portion of the low end merely to make a buzz disappear. After addressing the tone, evaluate any remaining hiss separately.

Room echo and reverberation

Listen after a word stops. A trailing copy or wash of that word points to reflected speech, which can remain even when pauses become quiet. Raising an ordinary denoiser until the pauses are silent may leave the room tail while harming the voice.

Use a tool that addresses echo or reverberation. Adobe Enhance Speech and Descript Studio Sound document this capability. For more specific control, iZotope's De-reverb guide explains learning the reverb profile, adjusting its frequency emphasis and tail length, then setting reduction. Excessive reduction can introduce artifacts.

Preview a phrase with a noticeable tail and a quiet phrase. Reduce enough that the words are easier to follow, then listen for hollow vowels, abrupt endings, or an unstable room sound. For a video shot in a large room, some reverberation can fit the scene. A completely dry voice may sound disconnected from the picture.

If every tested setting makes the words less clear, keep a gentler version. For a new recording, move the microphone closer and address the reflective space before relying on another cleanup pass.

Wind, plosive pops, and clothing rub

These sounds often occupy short sections. Mark each burst, include the surrounding words in the selection, and process that section independently. Compare the repaired section in context so you hear any change in tone at its boundaries.

iZotope's RX feature documentation distinguishes De-wind for wind rumble, De-plosive for plosive pops, and De-rustle for microphone contact with clothing. These are different repair tasks, and module availability depends on the edition.

For mild low-frequency rumble, try a high-pass filter, which reduces sound below its cutoff frequency. Apply it to the affected passage, raise the cutoff gradually, and stop if the voice loses body. This can reduce rumble; it will not reliably repair speech overwhelmed by turbulence or microphone overload.

Clothing rub can cover consonants as well as low frequencies. Check the exact affected word after processing. When a key phrase remains unintelligible, seek another microphone track or another take. For the next recording, secure the lavalier cable and capsule away from rubbing fabric; outdoors, use suitable wind protection.

Keyboard clicks, mouth clicks, and sudden impacts

If a click occurs during an otherwise empty pause, edit or attenuate that short region and smooth the boundary. If it overlaps speech, cutting the whole region also cuts the word. Use a repair tool that can target the disturbance, then compare the complete phrase.

iZotope's AI noise-removal walkthrough distinguishes dialogue cleanup from Spectral Repair for sudden unwanted sounds. In a spectral editor, select the visible disturbance as narrowly as you can, preview the repair, and expand the selection only if the remaining defect warrants it. If you cannot distinguish the unwanted sound from speech, use a conservative selection and compare carefully.

Keep natural breaths unless they are a separate editing problem. A row of silent gaps between words can make otherwise clean narration sound choppy. Small fades and a little sound from the original room help joins sound continuous, but they cannot restore a syllable removed by an edit.

Music behind speech

Decide whether you want to remove, lower, or preserve the music. If speech and music already occupy separate tracks, reduce the music track directly. Processing the combined mix creates a separation problem you can avoid.

For a mixed file, use speech isolation when you need speech alone. Use source separation when you want independent music and background control, such as the Premium stems documented by Adobe. Listen to sustained vowels, consonants, and the ends of sentences for both music leakage and altered speech.

If the background contains singing, include that passage in the test. Both wanted dialogue and unwanted vocals contain voice-like material. A quiet pause between sentences tells you little about how well the tool separates them while they overlap.

For a recording whose purpose is music, start with a music-oriented repair or separation workflow. A speech cleaner may treat instruments as unwanted sound. The ElevenLabs documentation's qualification about sung vocals is relevant here; do not turn a spoken-audio recommendation into a promise about music restoration.

Several speakers, audience questions, and overlapping speech

List the speakers who need to remain audible, including anyone farther from the microphone. Test the quietest speaker independently. A lecture can sound cleaner after isolation while losing the question that makes the answer understandable.

When separate microphone tracks exist, clean each track with settings appropriate to that speaker, then listen to the combined conversation. Soloing tracks helps locate noise, but the final mix reveals whether transitions or overlapping passages still sound natural.

With one mixed track, lower processing if a wanted voice is suppressed. Ordinary speech denoising does not guarantee that two simultaneous voices become independent tracks, or that a tool knows which person you intend to retain. If overlap remains unintelligible, preserve the source and mark the uncertain passage in the transcript. Do not replace an unclear word with a confident guess.

Clipping, crackle, and missing audio

Harsh distortion concentrated on loud syllables needs inspection before more denoising. Flat-topped waveform peaks can indicate clipping, although a later volume reduction can conceal the original peak level. Turning the file down does not undo distortion already recorded.

RX includes De-clip for clipping artifacts and De-click for impulse noises, as described in its feature documentation. A specialized repair may improve a damaged passage, but judge the words after repair. Avoid claiming complete recovery from a smoother waveform alone.

For continuous fine crackling, RX provides De-crackle. Try it on a short passage and compare the texture of the voice before extending the treatment. A continuous crackle, broad hiss, and an isolated click are different sounds; using the same aggressive filter for all three can damage speech without fixing the defect.

Recording dropouts leave a different problem: there may be no speech in the missing interval. Check other tracks or backups, and rerecord a critical line when possible. A generated replacement is a new recording and should not be treated as recovered source audio.

Compare the voice as well as the noise

Bring the original and cleaned versions into one editor and align the same phrase. Adjust playback gain until their voices sound similarly loud, keeping peaks below clipping. Leave tone controls fixed. A louder version can make a comparison misleading, so do this before judging clarity.

Listen once for remaining noise, then for the voice. Does the speaker retain their familiar tone? Are quiet consonants complete? Does noise surge between syllables? Replay any suspicious phrase directly against the original rather than relying on memory of a whole minute.

For a concrete word check, use a phrase such as "not approved" or "fifteen" only if it occurs in your recording. Listen for the quiet "not" and the end of "fifteen" in both versions. If that detail is clear in the source and blurred after cleanup, reject that setting even if the pause is almost silent. If it is unclear in both, the comparison cannot tell you what was said. Stop adjusting once the distraction is lower and the words still survive; the remaining faint noise does not need to reach zero.

If the recording will be transcribed, compare transcripts from the same recognizer and settings. Inspect names, amounts, dates, codes, and negation. A changed transcript is a reason to listen again, not proof that the denoiser changed the spoken word. Recognition itself can fail on an unchanged name.

Use the blank listening sheet if you need to record the decision. Keep noise reduction, naturalness, and important-word checks separate. Detailed transcript scoring belongs in the site's transcript accuracy guide; cleanup does not require you to calculate a score for every casual voice memo.

If the voice sounds robotic, compare the untouched source first. When only the cleaned version has that sound, lower the processing or change tools. If the source already sounds robotic, seek an unprocessed local recording or another take before applying more enhancement.

Use the tool's own original/processed mix control when available. If you manually play both files together, check their timing first. A slight delay between similar copies can create comb filtering, heard as a hollow or phasey voice. Compare each track alone, then together; if the problem appears only in the combined playback, mute one track and investigate alignment instead of increasing denoising.

Avoid feeding a heavily processed result into several competing speech enhancers while searching for a cleaner version. Compare each tool from the original. Targeted repair of a remaining click or hum can be useful after denoising, but evaluate that extra operation separately so you can undo it if it harms the speech.

Handle long recordings, upload failures, and export problems

The speaker is quiet and the noise is loud

Raise the playback volume enough to hear the original phrase before judging the cleanup. If you change clip gain, watch the peaks so loud moments do not clip. Turning up a mixed recording raises both speech and noise; it does not separate them.

Include the quietest important sentence in your test. Try moderate cleanup, then compare a milder version if syllables vanish. If the words are already masked in the source, inspect another microphone track or take. A clearer-sounding output is not evidence that the missing words were recovered.

A long recording changes conditions

Mark changes in room, microphone, speaker distance, and noise level. Test those sections separately and keep clean passages on a lighter setting. In an editor that supports local effects, give the treated section some surrounding audio and use fades or crossfades where the noise floor changes.

For file-based processing, keep each segment's original start time and sufficient surrounding audio to rebuild the join. If clips return to a video timeline, place them at those recorded positions. Listen across every join; a sudden change from room tone to silence is easy to miss when auditioning segments separately.

A cloud tool rejects or stalls on a file

Check the accepted format, file-size limit, duration limit, and remaining credits or daily allowance for your account. Adobe's plan documentation distinguishes Premium features such as video and bulk upload. ElevenLabs documents both size and duration limits; satisfying only one may still leave a file ineligible.

Try a short supported-format excerpt. If that works, investigate the full file's limits or split it with preserved timing. If it fails too, check the service status or error message before converting repeatedly. Renaming .m4a to .wav does not convert its audio encoding; export a real WAV from an editor when conversion is necessary.

Cleanup leaves the video out of sync

Check the downloaded audio's start position and duration against the original. A similar offset throughout suggests an alignment or processing-delay issue. An offset that grows toward the end calls for investigation of speed, timestamps, sample-rate interpretation, or the export path. These are diagnostic clues, not a diagnosis from the sound alone.

Place the cleaned audio at the source's intended start, compare a visible consonant or clap near the beginning and end, and check a middle point on a long recording. Do not stretch a track blindly to hide a mismatch. Descript's Repair audio drift guide documents a Chrome encoding issue that can cause echo or drift. In the Project panel's Files section, open the file's ellipsis menu and select Repair audio drift, then recheck the replacement file.

The export sounds worse than the preview

Keep a WAV or FLAC working copy where the tool allows it. These formats avoid another lossy encoding stage, but converting an MP3 to WAV does not restore information already discarded. Encode to your delivery format after editing, using the destination's specifications.

Inspect channel layout before export. One speaker can be appropriate in mono; separate microphones and intentional stereo ambience may need a different arrangement. Changing the sample-rate setting or peak level alone is not an improvement in noise removal. Open the exported file outside the processing tool and listen for new distortion, missing channels, altered duration, and problems at edit boundaries.

Reduce noise during a live call

Krisp describes its noise cancellation as real-time processing and says it is not intended for removing noise from an existing recorded file. Use the file workflows above for a recording you already have.

In Krisp, choose your physical microphone and output device. In the call app, choose the Krisp virtual microphone, and the virtual speaker if you want incoming-audio processing. Test normal speech, a quiet sentence, and the keyboard or fan you actually use. Confirm the call app is receiving the processed input.

For Zoom, Krisp's setup instructions recommend Low background-noise removal because combining Krisp with Medium or High can suppress the voice. If words cut out, compare with other suppression disabled before increasing Krisp processing. When recording music or several wanted people on one microphone, also check whether a speech-focused call setting is suitable for what you need to capture.

Use DeepFilterNet when speech processing must stay local

DeepFilterNet is an open-source speech-enhancement project with local processing. This section is for readers comfortable installing Python packages. For a steady fan and a simpler local interface, use the Audacity section above. Download packages and model weights before expecting to work offline.

The downloadable examples used Python 3.11, DeepFilterNet 0.5.6, and Torch/Torchaudio 2.2.2 on an Apple-silicon Mac. With Python 3.11 already installed, the package setup recorded for that run was:

python3.11 -m venv .venv
.venv/bin/python -m pip install \
  torch==2.2.2 torchaudio==2.2.2 \
  deepfilternet==0.5.6 soundfile
.venv/bin/deepFilter --help

These paths are for macOS or Linux. This is the archived setup, not a claim that those versions install on every current computer. For another platform or a new package release, use the project's installation instructions. Keep this environment separate from other Python projects.

For the Python command used in that kit, a short speech excerpt can be prepared with FFmpeg:

ffmpeg -i input.wav -ss 00:01:00 -t 25 \
  -ar 48000 -ac 1 -c:a pcm_s16le test.wav

This example creates a 25-second, 48 kHz mono test beginning one minute into the source. Replace the filename and start time. Use mono only for a speech source you intend to downmix; keep separate microphones and wanted stereo material separate. The excerpt's new zero point is not its original position in a video timeline.

With that environment available, create separate versions from the same input:

# Attenuation limited to 6 dB
.venv/bin/deepFilter -m DeepFilterNet3 --atten-lim 6 \
  --output-dir out-mild test.wav

# Attenuation limited to 12 dB
.venv/bin/deepFilter -m DeepFilterNet3 --atten-lim 12 \
  --output-dir out-medium test.wav

# No attenuation limit
.venv/bin/deepFilter -m DeepFilterNet3 \
  --output-dir out-strong test.wav

These are the kit's test settings, not recommended levels for every voice. The Python deepFilter command and the project's compiled deep-filter program have different options. Do not transfer this command line unchanged to the compiled program.

If installation fails with a missing Torchaudio API, inspect the installed versions and compare them with the project's requirements or the kit README. Repeatedly reinstalling the newest package pair can reproduce the same incompatibility. If processing finishes but the file sounds bad, return to the original test and compare a milder setting; a successful command only establishes that processing completed.

For video, check alignment after processing. The current Python implementation compensates for its STFT processing delay by default and has a --no-delay-compensation option to disable that behavior. Inspect your installed command's help and verify the result against picture.

Try the downloadable audio examples

The audio review kit contains 17 lossless audio files: one clean example, four versions with added noise or reflections, and three DeepFilterNet3 outputs for each noisy version. All voices and added sounds are synthetic. The kit has no human listening scores and no processed audio from Adobe, ElevenLabs, Descript, Resolve, or Krisp.

You can listen without installing DeepFilterNet. Unzip the kit and open the FLAC files in a player or editor that supports them:

1. Play fixtures/noisy-steady-fan.flac, then outputs/mild/noisy-steady-fan_DeepFilterNet3.flac at similar voice volume. Replay "Mara Chen" near the beginning and "no payment was approved" near the end.
2. Compare the same phrases with the matching file in outputs/strong/. Decide whether quieter background sound comes with any change you dislike in the voice. Do not decide from the opening pause alone.
3. Check fixtures/clean-reference.flac when you want to hear the source without added noise. Repeat with the room-echo files if reverberation is your problem. Write your own listening notes before consulting the measurements below.

Its 19.62-second clean recording follows a 47-word script containing a fictional name, dates, an amount, a code, and negation. Two measurements in the results file need explanation before reading the table. Word error rate compares automatically recognized text with the known script: a value of 0.0426 means 4.26%, and it can reflect mistakes by the recognizer. SI-SDR compares the audio signal with the clean example; a higher value does not establish that the voice sounds more natural.

Most real recordings have no clean example of the same performance, so you cannot make that SI-SDR comparison for your own file. You can still compare the original with the cleaned version by listening to the same words.

Even the clean reference produced a word error rate of 0.0426 because the recognizer rendered "Mara Chen" as "Marichand." That prevents treating the failed name check as proof of audible damage. The results identify passages to verify; they do not establish a product winner or a human preference.

The kit README explains the setup. An optional file checker verifies the archived records; it does not judge how natural the voice sounds.

Check the complete output before using it

For material you plan to publish or use as a record of a conversation, listen through the exported file. Spot checks are useful while adjusting settings; they do not establish that an unreviewed passage stayed intact. For a quick personal memo, concentrate on the parts you need to understand and keep the source available.

Check every speaker, quiet phrases, the worst noise, repaired sections, and transitions. For video, check timing near the beginning and end. Listen on headphones and the playback device your audience is likely to use. Keep the original, approved export, tool name, and settings together so another revision starts from the source.

For a recording that cannot be uploaded, choose local processing. Otherwise confirm that your use of the cloud service fits the recording's permissions and confidentiality requirements, including storage and deletion. Keep an untouched source if the recording will support an important decision, and have another person check critical phrases when their interpretation matters.

If a word remains unclear after reasonable repair, preserve that uncertainty in the transcript or replace the line with a new take when possible. Keep the version that reduces the distraction while retaining understandable speech and the sounds the recording needs.