AI VOICE CHANGER
AI voice changer that runs in your browser
Two clips in, one out. The first supplies the words, the timing and the delivery; the second supplies the voice. The Audiofilter audio filter does the conversion inside this tab — neither clip is uploaded, and there is no account.
LOCAL AUDIO LAB
Change a voice
Give it something to convert and a sample of the voice to convert into.Whatever is said here is what comes out — only the voice changes.
A sample of the target speaker. Clean speech from one person works best.
Do not use this to impersonate anyone or for any other unlawful purpose. See the Terms.
What it actually does
This is a tone colour converter, which is a narrower and more predictable thing than “voice changer” usually implies. It does not alter what is said or how it is said. Your recording keeps its words, its rhythm, its pauses and its emphasis; what gets replaced is the identity of the speaker.
The mechanism, in the order it runs: a reference encoder listens to your voice sample and distils it into 256 numbers — that is the entire representation of a voice here. A posterior encoder turns your recording into a latent sequence. A normalising flow then runs that sequence forward into a speaker-neutral space and back out again conditioned on those 256 numbers, and a vocoder rebuilds a waveform from the result.
The flow is the only part that sees the speaker embedding, which is why the conversion is as clean as it is: the encoder and the vocoder are doing speaker-independent work, and the identity swap happens in one place.
It needs audio to work from and will do nothing with typed words. If you want speech generated from text — with its own voice cloning — that is the text-to-speech tool instead.
What it costs to run locally
Honest numbers, measured on this build rather than guessed:
| Item | Figure | Why |
|---|---|---|
| First-visit download | About 102 MB, once | The converter weights, in five pieces because a static host caps files at 25 MiB. Cached afterwards. |
| Conversion speed | About 18 s of computing per 1 s of audio | Three WaveNet stages take 82 per cent of it and the vocoder the rest. There is no single hotspot to remove. |
| Clip length | Up to 10 seconds in, 30 seconds of reference | Five seconds already means about a minute and a half of waiting, and going longer only makes it longer. The reference is cheap — its encoder costs about 6 ms. |
| Output | 22 050 Hz mono, 32-bit float WAV | 22 050 Hz is the model's internal rate. Float rather than 16-bit so quiet passages keep their detail. |
Getting a result worth keeping
- Give the reference five to fifteen seconds of clean speech. One speaker, no music, no second voice. Longer is not better — it all collapses into 256 numbers regardless.
- Record the reference somewhere quiet. The converter reproduces the room along with the voice, so a fan in the sample becomes a fan in the output. If your only sample is noisy, run it through the noise reducer first.
- Feed it speech, not singing. Sustained pitch and vibrato are outside what this model was built for.
- Keep the source clean too. Background noise in the source survives the conversion; it is not a cleanup tool.
- Try more than one reference. Some voices transfer better than others, and it is quicker to test two references than to argue with one.
Where it will not work
- Anything long. Ten seconds a pass, and the wait scales linearly with the length — ten seconds is already about three minutes. A voiceover is not a good fit.
- Real time. At eighteen seconds of work per second of audio, live conversion is not on the table in a browser tab.
- Singing and music. Speech only.
- Changing what was said. There is no editing here, only re-voicing.
- Low-memory devices. The model needs roughly 200 MB of working memory while it loads. A modest phone may not have the headroom.
The tool will convert to whatever reference you hand it, which is exactly why the choice is yours to answer for. Cloning a real person's voice without their agreement can infringe rights of publicity, and several jurisdictions now have specific voice-likeness laws — copyright is not the only question in play. Use your own voice, or one you have permission to use. Nothing here is uploaded, so this is not us policing you; it is a note that the responsibility sits with whoever presses the button.
Not for unlawful use. Passing a converted voice off as someone else, deceiving a person or a voice-authentication system, harassment, defamation — none of that is a use this tool is offered for, and none of it stops being unlawful because the conversion happened on your own machine. What the law permits differs from country to country, and staying within the law where you are is your responsibility. Your use of this site is governed by the Terms.
The rest of the Audiofilter audio filter set
- AI vocal remover — split a song into an isolated acapella and a karaoke-ready instrumental.
- AI noise reducer — take fan hum, air conditioning and mic hiss out from behind speech.
- AI text to speech — 19 voices, plus cloning from a short sample.
- Audio converter and MP3 compressor — convert between formats, or shrink an MP3.
QUESTIONS & ANSWERS
Frequently asked questions
How does this AI voice changer work?
It is a tone colour converter. Your recording supplies the words, the timing and the delivery; the reference clip supplies the voice identity. A reference encoder turns the sample into a 256-number speaker embedding, and a normalising flow re-conditions your audio on that embedding before a vocoder rebuilds the waveform.
Do I need to type anything, like a text-to-speech tool?
No, and it will not work from text. This converts audio you already have. If you want speech from typed words instead, use the text-to-speech tool, which has its own voice cloning.
Is my audio uploaded anywhere?
No. The model is downloaded to your browser once and the conversion runs there. Open the network panel before converting: every request is a download of the model from this site, and nothing carries your audio the other way.
How long does a conversion take?
Roughly 18 seconds of computing for every second of audio, so a five-second clip takes about a minute and a half. That is measured, not estimated — the work runs on your own processor rather than on a server, and this model is heavy.
How long a clip can I convert?
Up to 10 seconds, and a short clip is the comfortable choice. That comes straight from the figure above: five seconds is about a minute and a half of waiting and ten seconds is about three minutes. The cap is where patience runs out, not a paid tier.
How long should the reference clip be?
Five to fifteen seconds of clean speech from one person is plenty. Longer does not help: the reference encoder costs about 6 milliseconds regardless, and it distils the whole clip down to 256 numbers either way.
Why does the output sound like my room?
Because a voice sample carries the room with it. The converter reproduces what it hears in the reference, including reverb, hiss and microphone character. Record the reference somewhere quiet, or clean it with the noise reducer first.
What sample rate does the output use?
22 050 Hz mono, which is the rate the model works at internally. It is downloaded as a 32-bit float WAV so that quiet passages keep their detail.
Can I clone anyone's voice with this?
Technically it will convert to whatever reference you give it, which is exactly why the choice matters. Cloning a real person without their agreement can infringe rights of publicity and, in several jurisdictions, specific voice-likeness laws — copyright is not the only question. Use your own voice, or one you have permission to use.
How big is the download, and does it repeat?
About 102 MB, once. The browser caches it, so later visits start converting immediately. It arrives as five pieces because a static host caps individual files at 25 MiB.
Cleaning a recording rather than re-voicing it?
Open the noise reducer → Open text to speech →