Voice Isolator: removing noise from audio with ElevenLabs

👁 129 reads

Voice Isolator from ElevenLabs removes background noise from recordings: it cuts out ambient sound, reverb and music, leaving only the voice. We tested it in English on a recording we deliberately corrupted: a clean monologue mixed in ffmpeg with fan noise, street rain and cafe chatter. We measured how much noise remains, whether the voice sounds like the original and whether the cleaned file transcribes better. The last result surprised us.

01ElevenLabs Scribe: English transcription test with hard namesElevenLabs Scribe v2 transcribed a 59-second English interview with tricky names and 12 numbers in 3.6 s with 2 automatic differences, correctly attributed 6 of 6 turns.Read →02ElevenLabs Voice Changer: AI Voice Swap Test & ResultsVoice Changer keeps your timing and melody while swapping the voice. We tested it on clean and noisy recordings - here's what survived and what didn't.Read →03ElevenLabs Voice Cloning: clones from 10 s, 30 s and 2 minWe cloned a synthetic voice from 10 s, 30 s and 2 minutes of audio. The 10-second clone reached 0.992 similarity, the 2-minute clone 0.996.Read →04ElevenLabs Dubbing v1 vs v2: Test on a 24-Second ClipElevenLabs Dubbing v1 vs v2 tested on a 24-second clip: time, credits, voice similarity, and one failure that still cost credits.Read →05ElevenLabs Sound Effects: Prompt Influence 0.3 vs 1.0 TestWe tested ElevenLabs Sound Effects with five English prompts at prompt influence 0.3 and 1.0. Both settings hit the target 5 of 5 times.Read →

What Voice Isolator does

The tool accepts an audio file (AAC, AIFF, OGG, MP3, OPUS, WAV, FLAC, M4A) or video (MP4, MOV, MKV, WEBM among others) up to 500 MB and 1 hour. It returns speech only. The documentation notes that the tool is not optimized for extracting vocals from music, though it sometimes works. The same model also powers the “remove background noise” switch in voice cloning and Voice Changer.

Our test: clean recording, four types of noise

We started with a 28.3-second monologue about apartment hunting (87 words), generated with a synthetic voice in the Eleven v4 model. This gave us a reference: we know exactly how the recording sounded without noise.

We added noise in ffmpeg. “Fan” is pink noise with 50 and 150 Hz hum, generated in ffmpeg itself. “Street” (rain and tram) and “cafe” (chatter, cups, espresso machine) are backgrounds from the sound effects generator. We set the level so the noise was as loud as the speech (0 dB SNR), and in the hardest version the cafe was 5 dB louder than the voice. Example for the fan, simplified:

ffmpeg -f lavfi -i "anoisesrc=color=pink:amplitude=0.5:duration=28.3" 
  -f lavfi -i "sine=frequency=50:duration=28.3" 
  -f lavfi -i "sine=frequency=150:duration=28.3" 
  -filter_complex "[1:a]volume=0.6[h1];[2:a]volume=0.25[h2];[0:a][h1][h2]amix=inputs=3:normalize=0" 
  -ac 1 noise.wav
ffmpeg -i clean.wav -i noise.wav 
  -filter_complex "[1:a]volume=0.675[n];[0:a][n]amix=inputs=2:normalize=0:duration=first" 
  noisy.wav

We calculated the noise gain (here 0.675) separately for each background, from the RMS level of speech and noise. We sent each noisy file on September 29, 2026 to Voice Isolator via the API.

Measurement results

We did not judge quality by ear. We measured three things against the clean reference: the ratio of speech level to what you hear in pauses between words (more dB means quieter pauses), the distance of the spectrum from the reference (less is closer to the original), and voice similarity using the WavLM speaker verification model.

In terms of sound this is a very good result: noise in pauses dropped by more than 50 dB, and the spectrum distance from the reference fell roughly threefold. In the clean reference this ratio is about 62.6 dB, so it does not quite reach studio silence. The output file has the same length as the input, so you can sync it to video without shifting.

Surprise: clean does not mean faithful

We transcribed each recording before and after cleaning with ElevenLabs Scribe and counted wrong words out of 88.

Sample-by-sample comparison with the reference came out very poorly, even though the spectrum is close to the original. It looks like the model does not just cut noise but reconstructs speech, and where the voice was masked it guesses phonemes. The ear gets clean sound, and transcription sometimes gets different words. Practical conclusion: for transcription send the raw file straight to the transcription service, and use Voice Isolator for what someone will listen to.

How to use Voice Isolator

1. In the app: Voice Isolator, upload an audio or video file and download the result.
2. Via API (tested September 28, 2026):

curl -X POST https://api.elevenlabs.io/v1/audio-isolation 
  -H "xi-api-key: $ELEVENLABS_API_KEY" 
  -F audio=@recording.wav -o clean.mp3

Listen to the result in places where the noise was loudest, and compare with the original before you publish a quote. If the recording will be used for voice changing, you do not need to clean it separately: Voice Changer has the same filter as an option, and in our test it made a huge difference there.

How much does Voice Isolator cost

The documentation states 1000 credits per minute of recording, which matches our measurement: 472 credits for 28.31 s. API pricing from September 28, 2026: $0.12 per minute. An hour-long podcast is therefore $7.20. In the Creator plan at $22 (121,000 credits) you can fit about 2 hours of cleaning per month if you use credits for nothing else. We compare the other plans in the article how much ElevenLabs costs, and all tools in the ElevenLabs guide.

Frequently asked questions

How do you remove noise from a recording with AI?

Upload the file to Voice Isolator in ElevenLabs, in the app or via API. In our test the noise in pauses between words dropped by 50-56 dB, and a 28-second recording was ready in about 3 seconds.

How much does Voice Isolator cost?

1000 credits per minute of recording, via API $0.12 per minute before VAT. Our 28.31 s cost 472 credits.

Does Voice Isolator improve transcription?

In our test no, it made it worse. With chatter louder than speech the number of wrong words rose from 2 to 14 out of 88, because the model reconstructs different phonemes in masked sections.

Does Voice Isolator work with video files?

Yes. It accepts MP4, MOV, MKV and WEBM among others up to 500 MB and an hour long, and returns the cleaned audio track.

// read moreElevenLabs Voice Changer: AI Voice Swap Test & ResultsAI ToolsElevenLabs Voice Cloning: clones from 10 s, 30 s and 2 minAI ToolsElevenLabs Dubbing v1 vs v2: Test on a 24-Second ClipBenchmarks

The link to ElevenLabs is a partner link: if you buy through it, QUANTSLANT earns a commission and you pay the same price. It has no influence on the content, including where the coverage is critical.