Make muffled calls and old songs clear. Removes noise, hum and clicks, brings back faded treble, then lets you compare before and after.
Built & maintained by Pappu Venkata Subbi Reddy, founder of Clacify · Updated October 2026
Audio Enhancer cleans up recordings that are hard to listen to: a WhatsApp call recorded next to a fan, a lecture captured from the back of a hall, a voice note with traffic behind it, or an old film song transferred from a worn cassette. You drop in the file and the tool works out whether it is mostly speech or music, removes steady background noise, electrical hum and clicks, brings back the treble that phone lines and old copies lose, and sets a sensible volume. Then comes the part most online enhancers skip: you can play the original and the enhanced version together and switch between them at the same moment, optionally at matched volume, so you judge clarity rather than loudness. If it is too strong or too thin, you move a slider, hear the change within a second, and export MP3 or WAV only when you are happy. Everything runs inside your browser; the file is never uploaded.
The tool first decodes your file with the browser's own audio decoder and analyses it: the background noise floor, peak level, clipping, stereo correlation, integrated loudness (ITU-R BS.1770 / EBU R128 method), whether mains hum is present, whether the sound is band-limited, and whether it is speech or music. Speech keeps dipping between syllables and words while music is continuous, so the tool measures how often the level drops sharply against its surroundings after removing the steady noise floor. On our test set, speech (including very noisy speech) scored 0.25 to 0.49 and music 0.09 to 0.18, so the mode is picked from the content rather than assumed, which matters because old film songs are often mono and would otherwise be treated as speech. Hum detection looks for narrow spectral peaks at 50 or 60 Hz and their harmonics across many frames and only reports hum when it is consistently there, so a sung note near 100 Hz is not mistaken for it. For speech, the cleaning engine is RNNoise, a small recurrent neural network designed by Jean-Marc Valin (Xiph.org/Mozilla) that combines classic signal processing with deep learning. It works on 10 ms frames at 48 kHz, estimates how much of each frequency band is voice, and suppresses the rest. It also reports, frame by frame, how likely it is that someone is speaking. Clacify uses that voice-activity signal in two ways: during pauses the noise is removed completely, while during speech a little of the original can be blended back (the Noise removal slider), which keeps voices natural and masks artifacts. RNNoise needs a fraction of a second to adapt, which normally lets noise through at the start of a file, so the opening two seconds are also processed in reverse by a separate instance and crossfaded in. The opening is then as clean as the rest. Music is handled differently, because a speech network treats instruments as noise. The music path repairs clicks by detecting sudden spikes in the signal's slope and interpolating across them, then applies a spectral noise reducer. It estimates the noise in each frequency bin from the quietest moments across the whole track (a minimum-statistics approach) and removes it with a Wiener-style gain that recovers quickly and fades slowly, so cymbals and reverb tails are not chopped. Many recordings are not noisy so much as muffled. The tool spots this as a cliff in the spectrum: a drop steeper than 30 dB per octave that natural voices and instruments never show. What it does next depends on what is left above that point. A phone line removes everything outside about 300 to 3,400 Hz, so for calls the missing range is rebuilt from the octave below it: that band is rectified, which creates the sums of its frequencies (these stay harmonic for a single voice), filtered to the missing range and matched in level to continue the voice's own spectral slope. The same idea restores the bass a phone line removes: neighbouring voice harmonics differ by exactly the speaker's pitch, so rectifying the 300 to 900 Hz band regenerates the fundamental. Songs are different. An old cassette or low-quality copy usually still carries its treble, just faded, and rectifying a full mix of instruments produces a gritty buzz: on a real song we measured that 87% of what it added above 3 kHz was new sound rather than the recording (coherence 0.13). So for music the faded treble is lifted with equalisation instead, continuing the recording's 2 kHz level at a natural slope of about 4 dB per octave, with a cap. That adds nothing that was not already in the recording (coherence 0.999). Cleanup has two separate controls. The difference between the original and the denoised signal is split exactly into everything below 3 kHz (background noise, rumble, crackle) and everything above it (hiss), so each can be removed independently. For music, hiss removal starts switched off: some recordings sound right with their tape hiss, and removing it is a choice the tool reports and leaves to you. After cleaning, both modes go through a 10-band graphic equaliser (31 Hz to 16 kHz). It starts flat, because on test recordings with a known original, automatically steering every file toward a generic "modern" tonal curve made the result measurably less like the original. Tone styles (Modern, fitted to your file's measured balance, plus Warm, Bright and Vocal focus) are one click away, and each band can be adjusted by hand. The slider values are the change you hear at each band; the tool solves for the filter settings that produce them, since neighbouring bands overlap. Finally comes loudness normalisation with a look-ahead peak limiter that keeps peaks at −1 dBFS so the export does not distort on phones or after MP3 encoding. Volume stays at the original level when a recording is already loud, or can be set to Soft, Standard or Loud; a mono file is measured as it sounds on two speakers, so it is not pushed into the limiter. When you move a control, the next 20 seconds from the playhead are re-rendered first (in under half a second for a song) so you hear the change almost immediately, while the whole file updates in the background. Some effects are available but off by default for music because they changed the sound in ways we could measure. Volume levelling (a compressor) softened each drum hit and lifted the lead-in by 3 dB and the reverb tail by 1.6 dB relative to the hit, which is heard as blurred beats and more echo. Stereo spread for mono songs adds slightly delayed copies, which can sound hollow. Sparkle adds shine through mild distortion, which can sound gritty. Each can be switched on from the edit panel if you like the result.
| Recording | Mode | Starting preset | What to adjust if needed |
|---|---|---|---|
| Phone or WhatsApp call recording | Voice | Phone call | Lower Background noise if words sound clipped |
| Voice note recorded outdoors | Voice | Voice note | Raise Silence between words to quiet the gaps |
| Podcast or interview, two people | Voice | Interview / podcast | Raise Even out volume if one voice is quieter |
| Lecture or meeting from far away | Voice | Lecture / meeting | Try the Vocal focus tone for intelligibility |
| Old film song or cassette transfer | Music | Old film song | Raise Tape hiss if you want the hiss gone |
| Concert recorded on a phone | Music | Live / phone recording | Lower the 63 and 125 Hz EQ bands if the room boomed |
| Already decent music, needs polish | Music | Gentle polish | Keep Noise low and try the Warm or Bright tone |
The tool guesses the mode from the file name and stereo content, but you can always switch before enhancing.
We tested the engine on recordings where the clean original is known, so improvement can be measured rather than judged by ear. The score below is how far the result's tonal balance is from the original, measured third-octave by third-octave every 46 ms from 80 Hz to 10 kHz (lower is better). A speech clip squeezed through a simulated phone line (300 to 3,400 Hz, quiet) went from 25.1 dB to 13.9 dB: the missing treble above 3.5 kHz came back to within about 5 dB of the original, and the bass below 300 Hz was rebuilt. Real recorded music made mono, cut above 5 kHz, with tape hiss and 50 Hz hum added (a typical old film song transfer) went from 21.7 dB to 16.6 dB with default settings, which keep the hiss, and to 11.6 dB with hiss removal switched on. Speech buried in fan-like noise at 5 dB signal-to-noise ratio plus hum went from 11.8 dB to 6.8 dB, and the gap between speech and the noise in the pauses went from 6.8 dB to over 80 dB, meaning the background between words was removed. A one-minute song took about 6 seconds on a laptop. Results on your recording depend on how much the noise overlaps the voice or music.
Louder almost always sounds "better" to the ear, even when it is not clearer. Many enhancers exploit this by simply making the output louder. The Match volume option measures the integrated loudness of both versions and turns the original up or down to match, so when you press A and B you hear only the difference in noise and clarity. Both versions play from one clock and switching is a 12 ms crossfade, so you compare exactly the same word or note rather than restarting the file.
Noise removal works best on steady noise such as fans, AC, hiss, engine drone and hum, because it can be told apart from the voice or music. Sudden sounds that overlap speech, like another person talking, a horn or music playing in the background of a call, are much harder, and pushing removal to 100% then makes voices sound underwater. No tool can bring back detail that was never recorded: a song recorded on a phone will not become a studio master, and a heavily clipped recording stays distorted. For songs, lifting faded treble brings back detail that is still in the recording, but it also lifts whatever noise sits there. For calls, the rebuilt bass and treble are reconstructed from the sound that survived: a close estimate, not the exact lost detail. If a result sounds artificial, lower Noise removal to 50–70% and compare again.
Decoding, analysis, noise removal, equalisation and MP3 encoding all happen in your browser tab using Web Audio, WebAssembly and a background worker, so the recording never leaves your device. That matters for call recordings and meetings, which often contain personal details. Any format your browser can decode works: MP3, M4A/AAC, WAV, OGG/Opus and FLAC, plus the audio track of MP4 and WebM videos. Files up to 15 minutes are supported; longer recordings can exhaust a browser tab's memory, especially on phones, so split them first with the Audio Cutter. MP3 exports are 192 kbps stereo or 128 kbps mono; WAV exports are 16-bit with dither.
Open Clacify's Audio Enhancer and drop in your file (MP3, M4A, WAV, OGG, FLAC, or a video). Choose Voice for calls and speech or Music for songs, then click Enhance. When it finishes, press play and switch between Original and Enhanced to hear the difference. Adjust the Noise removal slider if needed and export as MP3 or WAV.
Yes, in most cases. It removes tape hiss, hum and crackle, detects when the treble has been lost and rebuilds it from the sound that survived, and can open a mono song into gentle stereo. It cannot recreate instruments that were never captured, so results depend on the source. Start with the Old film song preset; if cymbals or vocals sound watery, lower Noise. Tape hiss is kept unless you choose to remove it.
No. Decoding, noise removal, equalisation and MP3 encoding all run inside your browser using Web Audio, WebAssembly and a background worker. Your file never leaves your device.
That happens when noise removal is pushed too hard on a recording where the noise overlaps the voice or music. Lower the Noise removal slider to around 50–70% and reduce Silence cleanup. Use the volume-matched A/B switch to check you are actually improving it rather than just making it louder.
Up to 15 minutes per file. Longer recordings can run out of memory in a browser tab, especially on phones. Trim long recordings first with the Audio Cutter.