AI AUDIO FILTER
Audiofilter's AI audio filter for text to speech turns written words into natural speech on your own machine. Pick one of 19 free voices or clone a new one from a single short recording, type your text, and press Convert. Like every audio filter on this site, the model runs inside the page — your text and your recordings never leave this browser.
SharedArrayBuffer, which requires the
Cross-Origin-Opener-Policy and Cross-Origin-Embedder-Policy
headers and a secure context (localhost or HTTPS).
If you just changed server settings, do a hard reload (Ctrl/Cmd + Shift + R).
SharedArrayBuffer and a
WebAssembly heap that can grow to 1.5 GB, and Chrome and Edge handle that
combination most predictably. Firefox and Safari have the threads and
SharedArrayBuffer too but set a lower ceiling on WebAssembly memory,
so they are worth trying — if the engine will not start, or stops partway,
switch to Chrome. Chrome on a phone or tablet works as well; the first download
and each generation simply take longer on a phone's processor.
The other four Audiofilter tools have no such requirement and run anywhere.
Text to speech and voice cloning — free, unlimited, nothing uploaded
Tap the small button to preview
First run only — everything is cached afterwards.
English only
Higher values let the model hold a sentence open a little longer before ending it. The effect is subtle — roughly a second across the whole range.
A short walkthrough recorded by Audiofilter.
Loading video… Video unavailable here. Use the X link below.Playing this video loads content from X. X's privacy policy applies.
Three steps, no sign-up
Start from one of the nineteen built-in voices, or make a new one from a single recording — drop in an audio file, or record about ten seconds straight from your microphone. There is no training step and no queue: the voice is ready as soon as the clip has been read.
Paste up to 500 characters, then set how it should be read. Expressiveness decides how much the delivery varies between runs, and the pause controls set how long the model rests between sentences and at commas — the difference between a brisk read and unhurried narration.
Audio starts playing as it is produced, so you hear the opening words before the last ones exist. Scrub through the waveform to check it, then save the result as MP3, AAC or WAV. Every conversion after the first one begins immediately, because the model is already on your machine.
No account, no quota, no watermark
One short clip is enough to clone a voice. Nothing is fine-tuned and nothing is stored — the clip is turned into a compact voice embedding in the browser, used for as long as the tab is open, and then gone.
There is no backend and no API key, and this is the one tool page that carries no advertising. The speech model itself is compiled to WebAssembly and runs inside the page, so your text and your recordings have nowhere to go even in principle.
Every built-in voice is released under CC0, which means the audio you generate is free to use commercially with nothing to credit — an unusual thing for a free text to speech tool.
Adjust expressiveness, the pause between sentences and whether commas are treated as breaths. A sentence-end threshold is available too, for holding a phrase open a fraction longer before it resolves.
Downloads are encoded on your device as well, so the finished file never round-trips through a server either. WAV is written instantly; MP3 and AAC go through a local encoder.
The model is cached the first time you convert, so later visits skip the download entirely and generation starts the moment you press Convert.
What generates the speech on this page
The voice generation here is PocketTTS, a neural speech synthesis engine run through ONNX Runtime compiled to WebAssembly. It executes inside this browser tab, so your text and any voice sample you provide stay on your own device.
Voice cloning is zero-shot: a short sample is reduced to a compact voice embedding and used to condition the synthesis. Nothing is fine-tuned, nothing is uploaded, and the embedding disappears when you close the tab.
No. There is no backend and no API call. The model runs inside this browser tab, and your text and recordings stay on your device.
Yes. There is no account, no trial, no character quota and no watermark. Because the work happens on your own machine rather than on a server, there is nothing to meter in the first place.
Yes. Every built-in voice is released under CC0 1.0, so the speech you generate is yours to use for anything, including commercial work, with no attribution required. Audio from a voice you cloned yourself is likewise yours.
The model has to be fetched once before anything can be generated. After that your browser keeps it cached and Convert responds immediately.
Six to twelve seconds of one person speaking clearly, with no background music, echo or other voices. Longer is not better — clips over fifteen seconds get trimmed.
Up to 500 characters per conversion, which is roughly eighty words or half a minute of speech. There is no limit on how many conversions you run, so longer scripts can be done a paragraph at a time.
English. The model was trained on English speech and every built-in voice is an English reader, so English is the only language it pronounces correctly. French, German, Spanish, Portuguese, Italian and Dutch will produce sound, but with English letter-to-sound rules applied to them. Chinese, Japanese and Korean do not work: the tokenizer has no pieces for those scripts, so the text is reduced to raw bytes the model has never seen. The accents listed on the voices are accents of English, not separate languages.
MP3, AAC or WAV. The encoder runs on your device too, so the finished file is never sent anywhere before you save it.
Prefer Chrome; Edge behaves the same way, and Chrome on a phone or tablet works too — the first download and each generation take longer on a phone's processor, but the tool runs. The engine needs WebAssembly threads and SharedArrayBuffer, which older browsers do not have, and a WebAssembly heap that can grow to 1.5 GB — that last part is where Firefox and Safari set a lower ceiling, so the engine may not start there, or may stop partway. Try them if you like, and switch to Chrome if it fails. This applies only to text to speech; the vocal remover, noise reducer and converter run in any recent browser.
Once the model has been cached, generating speech needs no network at all. The page itself still has to load, so you do need a connection to open it the first time in a session.
The page must be served with the COOP and COEP headers over localhost or HTTPS. If the setting was just changed, do a hard reload so the browser stops reusing its cached copy.
Need a different audio filter for a recording you already have?
Open the vocal remover → Open the noise reducer →