Enhance voice

Ten steps in a row: room echo reduction, neural noise removal (DeepFilterNet3), mouth clicks, “um”s and stutters cut (Studio mode), long pauses shortened, breaths softened (not removed: a voice with no breaths sounds fake), voice EQ, levelling and gentle compression, then loudness set to the chosen standard. Each step can be switched off in the settings.

Fast does everything on your device: nothing is uploaded. Studio also cuts “um”s and stutters: to find them, a compressed copy of the (already cleaned) voice is sent to our server, which only returns the list of words and their timings. The cuts are made in your browser. The copy is not kept: it is deleted as soon as the transcription is done.

It rebuilds the voice instead of only cleaning it: an AI model (Sidon, by Sarulab Speech, MIT license: github.com/sarulab-speech/Sidon) listens to the recording and resynthesizes it in full-band studio quality. It works very well on a phone call, an echoey room or a poor microphone. A copy of the voice is sent to our server for this, then deleted. The timbre may change slightly: compare with the before / after player. Filler words, silences and loudness are then handled as in Studio mode.

−16 LUFS by default (EBU R128), the podcast reference (Apple Podcasts, Spotify). In the settings: −14 LUFS for social media and YouTube, −23 LUFS for broadcast (radio, TV). True peak is always limited to −1.5 dBTP to avoid clipping.

It is clearly reduced: the tail left between words drops by more than 20 dB in our tests in a large hall. A voice recorded very far from the mic keeps some room colour: nothing beats a close mic.

Each cut is made at the edge of the hesitation vowel, with a 15 ms crossfade, and the surrounding pause is brought back to a natural length. Only “um”, “uh” and involuntary repeats (“I I”, “that the, that the”) are cut; “very very” or “no no” are kept. Listen before and after: if a cut bothers you, switch that step off.

Yes. The cleaned voice is put back into the video, the picture copied as is, with no loss. By default nothing is cut in a video; if you turn on “um” or pause removal, the picture is cut at the same points (re-encoded to MP4, slower) to stay in sync.

The tool detects a phone voice (narrow band) on its own and adapts: gentler noise removal, EQ that clears the nasal tone. Highs lost by the phone line cannot be recreated.

20 minutes per file, free, no account. Expect one to two minutes of processing for ten minutes of audio on a recent computer, more on a phone. The output is mono, the standard for voice.