Free Audio Enhancer
Audio Enhancer cleans podcasts, interviews and voice recordings in one pass: noise reduction, rumble and pop filters, a de-esser, a leveler that brings quiet and loud voices together, and loudness normalization to −16, −14 or −23 LUFS with a true-peak limiter. It can also shorten long pauses and cut filler words before you save WAV, MP3 or the video.
- Loudness:
- Podcast −16, streaming −14, EBU −23 LUFS
- Cuts:
- Long pauses and filler words
- Output:
- WAV, MP3, video, Audacity labels
- Price:
- Free
- Runs:
- Locally, in your browser
- Uploads:
- None: files never leave your device
- Account:
- Not needed
Quick answers
Is my recording uploaded?
No. Noise reduction, leveling, loudness and cuts all run in your browser. Only when Cut filler words is on, the speech recognition model (Whisper Base: 140 MB on the CPU, 210 MB with WebGPU) is downloaded once from Hugging Face; it also runs on your device, and the recording never leaves it.
What does the leveler do?
It measures the speech level over a 2-second window, ignoring the pauses, and turns quiet passages up and loud ones down by up to 12 dB, so every speaker ends up at about the same level. In pauses the gain never rises above the speech around them, so background noise is not pumped up. A gentle compressor then evens out what is left.
Processed on your device: your files are never uploaded. How it works
Network activity in this tab
What this page has requested since you opened the site in this tab. An upload is any request that sends data: POST, PUT, PATCH or a request with a body.
No uploads: nothing has been sent from this page.
No network requests yet.
To see every request, open your browser's developer tools (Network tab).
How to use the Audio Enhancer
- Drop a podcast, interview, lecture or video recording.
- Choose the clean-up (noise reduction, rumble, de-esser, pop filter), the leveler and the loudness target, and whether to shorten long pauses or cut filler words.
- Click Enhance, then compare A (original) and B (enhanced) while listening; untick any filler word you want to keep and click Update.
- Download WAV or MP3, the video with its new soundtrack, or the cut list for Audacity.
Credits & open source
Our own code plus open-source libraries. Thank you to the people who make these projects.
- web-noise-suppressorAudioWorklet noise suppression (RNNoise, GTCRN, Speex) compiled to WebAssembly. © 翠 / green (sapphi-red); includes RNNoise (Xiph.Org, Jean-Marc Valin; BSD-3-Clause), SpeexDSP (Xiph.Org; BSD-3-Clause), GTCRN (Rong Xiaobin; MIT) and the PFFFT FFT library (Julien Pommier; FFTPACK-derived BSD-like license) · MIT
- wavesurfer.jsAudio waveform display with selectable regions. © katspaugh and contributors · BSD-3-Clause
- MediabunnyReads, writes and converts media files with WebCodecs (MP4, WebM, MKV, MOV, MP3, WAV, Ogg...). © Vanilagy and contributors · MPL-2.0
- lamejsMP3 encoder (port of LAME), loaded as a separate unmodified chunk. © LAME project; JavaScript port by Alex Zhukov; ESM fork by Breezystack · LGPL-3.0
- Transformers.jsRuns Hugging Face AI models in the browser with ONNX Runtime. © Hugging Face, Inc. and contributors · Apache-2.0
- ONNX Runtime WebExecutes ONNX models with WebGPU or WebAssembly (self-hosted from /vendor/ort). © Microsoft Corporation · MIT
Web platform APIs used: Web Workers, WebAssembly, Web Audio API, WebCodecs, File API, localStorage.
- RNNoise model weights by Jean-Marc Valin / Xiph.Org Foundation (rnnoise-wasm build by Shiguredo Inc.) (BSD-3-Clause (rnnoise-wasm Apache-2.0)) — Standard noise reduction and the high band of the Strong one
- GTCRN speech enhancement model by Rong Xiaobin (GTCRN); WebAssembly build by sapphi-red (MIT) — Strong noise reduction (band below 7 kHz)
- SpeexDSP preprocessor by Xiph.Org Foundation, Jean-Marc Valin and contributors (BSD-3-Clause) — Light noise reduction
- Whisper tiny / base / small (ONNX, timestamped exports) by OpenAI (model); onnx-community (ONNX conversion) (MIT (openai/whisper; the Hugging Face model cards list Apache-2.0)) — finding filler words and their timings (optional, Base by default)
- Whisper large-v3-turbo (ONNX, timestamped export) by OpenAI (model); onnx-community (ONNX conversion) (MIT) — most accurate word timings, WebGPU only (optional)
Full license texts: third-party notices · all credits
Frequently asked questions
Which loudness target should I choose?
Podcast (−16 LUFS, −1 dBTP) is the common spec of podcast hosts and Apple Podcasts; YouTube and streaming use −14 LUFS; EBU R128 (−23 LUFS) is for European TV and radio and ATSC A/85 (−24 LUFS, −2 dBTP) for US TV. The result card shows the measured loudness and true peak of your file.
How are filler words found?
Speech recognition (Whisper, Base by default) gets a sample transcript that keeps its “um”s and “uh”s, so it writes most fillers instead of leaving them out; those are ticked for cutting. Held sounds between words (a voice held for 0.2 s or more on an almost flat pitch) are listed unticked. On a 14-minute English meeting from the AMI corpus, 78% of the ticked cuts were fillers or hums, and 33 of its 70 fillers were found. Check the list with Listen before downloading.
Can I enhance the sound of a video?
Yes. Drop the video and use Video with enhanced audio: the picture is copied untouched and only the soundtrack is replaced. Pause and filler cuts are turned off for videos, because removing audio would move the sound away from the picture.
Does it remove echo or rebuild a muffled voice?
No. It reduces background noise behind speech, not the reverberation of the room, and it does not recreate missing high frequencies of a phone or low-bitrate recording.