Online Text to Speech Voice Generator

1. 1

   Enter text

   Type or paste what you want read aloud, or import a .txt or .md file.

2. 2

   Choose an engine

   Instant uses your system voices and plays immediately. Studio runs a neural voice on your device and can be downloaded.

3. 3

   Pick a voice

   Choose from 28 Studio voices, or from whatever voices your operating system provides.

4. 4

   Set the pace

   Adjust the speaking rate if the default reads too fast or too slow.

5. 5

   Generate

   Studio synthesizes the audio on your device; progress is shown per section.

6. 6

   Download

   Save the result as WAV, MP3, M4A, OGG or FLAC.

Two engines sit behind one text box, because the two things people want from text to speech pull in opposite directions.

• Instant — uses the voices already installed on your computer or phone. It starts speaking the moment you press play, downloads nothing, and supports whatever languages your system does. It cannot be saved to a file.
• Studio — runs a neural voice model on your device. It takes a moment longer and produces a real waveform, which means it can be downloaded as an audio file and used in a video, a podcast or a presentation.
• One text box — the text, the speaking rate and the voice list are shared, so switching engines does not mean starting again.
• Nothing uploaded — both engines run locally. The text you paste and the audio produced stay on the device, which is not true of the online tools that send your text to a server to be spoken.

This is the part most free text-to-speech tools hold back, and the reason the Studio engine exists.

• WAV — the exact output of the model, uncompressed, at 24 kHz. Choose it when the audio is going into a video or audio editor and will be processed again.
• MP3 — the smallest widely compatible file, and the right default for sharing, uploading or embedding. Encoded on your device.
• M4A — AAC in an MP4 container — the Apple-native choice, and well supported in video tooling.
• OGG and FLAC — Opus in OGG for the web at small sizes, and FLAC when you want a lossless file that is meaningfully smaller than WAV.
• Why Instant cannot be saved — system voices play straight to the audio output and expose no stream to record. That is a limitation of the operating system, not of this tool, and it is why the neural engine is here.

The Studio engine ships 28 voices. They differ in more than pitch — each was trained separately, and some are noticeably clearer than others.

• American English — 20 voices, 11 female and 9 male. These are the exact match for the pronunciation dictionary and are the best choice for most text.
• British English — 8 voices. They speak in a British accent using American vowel choices, because the dictionary behind them is American. The difference is audible on words like 'bath' and 'dance'.
• Choosing one — the voices are listed best-first. If a voice sounds thin or unstable on your text, move up the list rather than adjusting the rate.
• Other languages — switch to the Instant engine, which offers whatever your operating system provides — usually a much wider set of languages.

This is where text-to-speech tools most often give themselves away, so it is worth being specific about what happens.

• Whole numbers — 1,234 is read as 'one thousand two hundred thirty four', not as four separate digits.
• Years — 'in 1999' becomes 'nineteen ninety nine', while '1999 items' is read as a quantity. The surrounding words decide.
• Money — $1,234.56 becomes 'one thousand two hundred thirty four dollars and fifty six cents'. Pounds, pence and euros are handled the same way.
• Decimals and versions — 1.5 is 'one point five'. 2.0 is 'two', because that is how it is said aloud.
• Ordinals and ranges — 3rd, 21st and 102nd are read as ordinals, and a range like 128-192 is read as two numbers rather than being skipped.
• Symbols — 25% is 'twenty five percent', & is 'and', @ is 'at'. Acronyms such as ISO, PDF and UTF-8 are spelled out letter by letter rather than being pronounced as words.

Speech quality is mostly a dictionary problem, not a model problem. A perfect voice saying the wrong sounds is still wrong.

• A dictionary of 183,562 words — each entry gives the exact sounds and which syllable carries the stress.
• Word endings by rule — 'rendered' is not in the dictionary but 'render' is, and the rules for -s, -ed and -ing build the rest. This is why inflected forms sound right.
• Words that change with their job — 'use' is said differently as a noun and a verb, and so are 'record', 'content', 'present' and about seven hundred others. The surrounding sentence decides.
• Context between words — 'the' before a vowel is not the same sound as 'the' before a consonant, and neither is 'to'. Getting this wrong makes every sentence subtly foreign while every word is individually right.
• Unknown words — names and coinages go to a small neural model that predicts pronunciation from spelling, so an unfamiliar word is spoken rather than skipped.

Text-to-speech is unusual among conversion tools: the thing you paste is often the thing you would least like to hand to a stranger.

• Nothing is uploaded — the voice model downloads to your device once. After that, it all runs on your device, not our servers.
• It works offline — once the model is cached the tool keeps working with no network at all.
• No quota tracking — there is nothing on a server to count, because there is nothing on a server.
• The trade — the model has to be downloaded once, and synthesis uses your processor rather than someone else's. That is the cost of the text not going anywhere.

The work happens on your device, so the honest answer to 'how fast is it' depends on the device. Here is what to expect.

• First use — the Studio engine downloads about 93 MB — the voice model, the pronunciation dictionary and one voice. It is cached afterwards.
• After that — conversion starts immediately. Long text is split into sections and several are synthesized at once across your CPU cores.
• Free — 5,000 characters per conversion, one at a time. That is roughly six minutes of speech, and more than most free tools allow.
• Pro — no character limit, up to 25 documents at once, delivered as a single archive.
• What is never limited — every voice, every speaking rate and every output format including MP3 are available on the free plan. The paid axis is scale, not quality.

Synthesis runs on your device across several CPU cores at once, so speed scales with your hardware and nothing waits on a server.

Most of what makes synthesized speech sound good is decided before you press generate.

• Punctuate properly — commas and full stops are pauses. Text without them is delivered as one breathless run.
• Break paragraphs — a blank line becomes a longer pause, which is what separates sections of a script.
• Write numbers as words when it matters — 'twenty twenty four' if you want a year read that way in a context the tool would otherwise read as a quantity.
• Respell awkward names — phonetic spelling in the input is the fastest fix for a mispronounced name, and it always wins.
• Slow down slightly for narration — a rate just below 1× reads as more considered, which suits explanation. Leave it at 1× for dialogue.
• Test with a short extract first — generate one paragraph, listen, adjust, then run the whole document.

It is a better tool for some jobs than others, and the honest boundaries are worth stating.

• Emotional performance — the model reads clearly and evenly. It does not act, and it will not deliver a joke or a dramatic line the way a person would.
• Languages other than English — the Studio dictionary is American English. Use the Instant engine for other languages, or a service built for them.
• Cloning a specific person's voice — not supported, deliberately. The 28 voices are the voices.
• Very large documents on a low-powered device — a book-length text on an old phone is not a good match. Split it, or run it on a laptop.

• Speech to text
• PDF to text
• Image to text
• Image converter
• Text splitter
• Case converter
• Audio converter
• Audio merger
• Audio cutter
• Audio volume
• Audio editor
• Replace video audio