Audio Noise Remover API: clean speech in MP3, WAV and MP4

Audio Noise Remover takes links to voice recordings and returns a cleaned file with the background noise taken out. It runs the open-source DeepFilterNet3 speech-enhancement model, evens out the loudness, and hands back a download link plus before/after measurements (noise floor and LUFS) so you can see what changed. Price: $0.02 per audio minute, no start fee, failed files are never charged.

Use it to remove hiss, fans, air conditioning, traffic, keyboard clatter, hum and light room echo from podcasts, interviews, voice memos, lectures, calls, screen recordings and video narration, with no API key, no account on another service and nothing to install.

What you get for each file

Field	Meaning
outputUrl	Download link to the cleaned MP3, M4A, WAV, FLAC or Opus (or MP4/MKV with the cleaned audio swapped into your video)
noiseFloorBeforeDb, noiseFloorAfterDb, noiseReductionDb	Level of the quietest passages before and after cleaning (dBFS RMS trough). Lower is quieter
loudnessBeforeLufs, loudnessAfterLufs, truePeakAfterDb	Integrated loudness and true peak, so every episode can land at the same level
durationSecs, processedSecs, billedMinutes	What was cleaned and what you were charged for
status, error	ok, error or skipped, with a plain-English reason

The full table of results is in the dataset (Overview view). Cleaned files sit in the run's key-value store and the links open without a token.

Quick start

1. Click Try for free and paste one or more direct file URLs into Audio or video file URLs.
2. Leave the defaults (maximum reduction, normalise to -16 LUFS, MP3) or change them.
3. Start the run. Download the cleaned files from the outputUrl column.

Use with AI agents / MCP

The Actor has one required input, sources, and a working example, so an agent can call it zero-shot.

{
  "sources": ["https://upload.wikimedia.org/wikipedia/commons/a/a5/Live_recording_of_interview_with_Dr_Amba_Ambaiowei_Pt_1.opus"],
  "reduction": "maximum",
  "targetLufs": -16,
  "outputFormat": "mp3"
}

Sample output (shortened):

{
  "sourceUrl": "https://upload.wikimedia.org/wikipedia/commons/a/a5/Live_recording_of_interview_with_Dr_Amba_Ambaiowei_Pt_1.opus",
  "status": "ok",
  "durationSecs": 67.84,
  "billedMinutes": 1,
  "outputUrl": "https://api.apify.com/v2/key-value-stores/.../records/001-Live_recording_of_interview_with_Dr_Amba_Ambaiowei_Pt_1-clean.mp3?signature=...",
  "outputFormat": "mp3",
  "noiseFloorBeforeDb": -43.6,
  "noiseFloorAfterDb": -88.2,
  "noiseReductionDb": 44.6,
  "loudnessBeforeLufs": -13.6,
  "loudnessAfterLufs": -16.3,
  "truePeakAfterDb": -3.4
}

Settings

• Noise reduction strength: maximum removes as much as the model can. strong (up to 35 dB), medium (20 dB) and gentle (12 dB) cap the reduction so some natural room tone stays.
• Extra residual-noise filter: DeepFilterNet's post-filter for very noisy recordings.
• Rumble filter (Hz): a high-pass filter, 80 Hz suits most voices, applied before cleaning.
• Normalise loudness / Target loudness: -16 LUFS for podcasts, -14 for YouTube and streaming, -23 for EBU R128 broadcast. The result is checked after encoding and the gain is corrected if it landed more than 0.7 LU off target. A limiter keeps true peaks near -1.5 dBTP.
• Output format and bitrate: MP3, M4A (AAC), WAV (48 kHz), FLAC or Opus.
• Keep the video track: for video sources the picture is copied untouched and only the audio is replaced.
• Max minutes per file and Max file size: spending guards. You pay only for minutes actually cleaned.

How it works

Each file is downloaded, probed with ffprobe, then cut into six-minute pieces with 1.5 seconds of overlap on each side. Every piece goes through DeepFilterNet3 at 48 kHz, the overlap is trimmed, and the pieces are joined, so memory stays flat for hour-long recordings. The result is measured, gain-adjusted and encoded with ffmpeg. With 8 GB of memory two pieces run in parallel.

Speed depends on the memory you give the run: roughly 3x real time at the default 4 GB and around 7x on long files at 8 GB, so a one-hour podcast takes about 20 minutes at the default. Cost per audio minute stays about the same either way, so choose more memory when you want results sooner.

How much does it cost to clean audio?

$0.02 per audio minute at the Free and Bronze tiers, $0.017 for Silver and $0.014 for Gold. Minutes are rounded up per file. A 30-minute podcast episode costs about $0.60 on the Free tier; a one-hour interview about $1.20. That compares with roughly $1.40 per hour at Auphonic, $0.60 to $2.20 per hour at Cleanvoice and about $7 per hour for ElevenLabs Voice Isolator, with nothing to subscribe to here. Errors, unsupported links and files without audio cost nothing. Set a maximum charge on the run and the Actor stops cleanly when it is reached.

Audio noise remover API

Call it from code with the Apify API or client libraries, schedule it, or plug it into n8n, Make or Zapier. Typical pipeline: download recording, run Audio Noise Remover, then send outputUrl to a transcription Actor such as Audio & Podcast Transcriber, since cleaner audio usually improves transcription accuracy.

Best results and limits

• It is trained for speech. Music and singing are not its target and may sound dull or watery. Use gentle if you must process mixed material.
• It removes steady and non-steady background noise and some reverb. It does not remove other voices, so a second speaker in the background stays.
• Very low-bitrate or heavily clipped sources keep their compression artefacts.
• Files must be direct links to audio or video. Pages from YouTube, Spotify, SoundCloud, TikTok, Instagram and similar sites are rejected, because downloading from them breaks their terms.
• Output is 48 kHz. Stereo stays stereo; mono stays mono.

Is it legal?

Yes, by design. The Actor performs pure computation on files you point it to: it does not scrape websites, log in anywhere or bypass any protection. DeepFilterNet is published under the MIT and Apache-2.0 licences, which permit commercial use. You are responsible for having the right to process the recordings you submit, and the Actor refuses streaming and social-media page URLs for that reason. The model runs inside the Actor, so your audio is not sent to any third-party service, and results are stored only in your own run storage.

FAQ

How do I remove background noise from an MP3? Paste the MP3 URL into Audio or video file URLs and run. The cleaned MP3 is in outputUrl.

Can it clean the audio of a video? Yes. With Keep the video track on, you get the same video (picture copied, not re-encoded) with cleaned audio.

Does it work in languages other than English? The model is language-independent because it works on the sound, not the words, so it handles any spoken language.

Will it change how the voice sounds? At maximum strength you get the cleanest result; at medium or gentle more of the original room character remains.

What is LUFS? A loudness measure used by podcast and streaming platforms. Normalising to -16 LUFS makes episodes sound equally loud.

Alternatives? Adobe Podcast Enhance, Auphonic, Cleanvoice and ElevenLabs Voice Isolator are polished editors or priced per hour. This Actor is for batch and API use with open pricing and no sign-ups.

Support

Found a file that does not work or want another output option? Open an issue on the Actor page and include the run link.