How text-to-speech works
Between the text you paste and the voice you hear, several steps happen one after another. Here they are in order, without formulas, along with what Readotte does at each one.
What are the main steps of text-to-speech?
The work can be summed up in four steps: prepare the text, convert it into the sounds of the language (phonemes), predict what those sounds should be like, then produce the audio. Each step fixes what the previous one cannot handle.
In older systems, these steps were separate, hand-written programs. In neural systems, several of them are learned together by a single model, which gives a smoother voice.
What is text normalization for?
Normalization turns what is written into what should be said: “12” becomes “twelve”, “Dr.” becomes “doctor”, a date becomes a string of words. Without it, the voice would stumble over every number, acronym or abbreviation.
It is also the step where the system decides what not to read. In Readotte, symbols, emoji and icons are skipped, and in a PDF, page numbers, repeated headers and footers, small-print captions and citation marks like [12] are removed before reading.
What is a phoneme, and why is it needed?
A phoneme is a basic sound of a language, like the “sh” in “ship” or the “th” in “think”. Spelling does not always tell you how to say a word, so the system first translates it into a sequence of phonemes.
This is where the traps of English get resolved: “read” in the present and the past, “lead” the metal and “lead” the verb, silent letters as in “knight”. Some recent models learn this conversion directly from the text instead of going through a separate pronunciation dictionary.
What do the acoustic model and the vocoder do?
The acoustic model decides how each phoneme sounds in context: its length, its pitch, its loudness, the melody of the sentence. It produces a kind of score for the sound, often a spectrogram, which you cannot hear yet.
The vocoder turns that score into a sound wave, the one your speaker plays. In neural voices, both roles are played by neural networks, sometimes merged into a single model; the details vary from one model to another.
How does Readotte split a text into sentences?
Readotte cuts the text at end-of-sentence punctuation, while handling the cases that are misleading: an abbreviation such as “p. 12” or “etc.” does not end the sentence. Line breaks in the middle of a sentence, common when you copy from a PDF, are joined back together.
A heading or list item without a final period is read as a separate sentence, with a falling intonation, instead of being glued to the next paragraph. For a PDF, the layout is first rebuilt in reading order, as explained in our guide to reading a PDF aloud.
Why read sentence by sentence rather than the whole text at once?
Because the sentence is the right unit for the ear and for the interface: the voice keeps a natural intonation from one period to the next, and listening starts without waiting for the whole document to be converted. The current sentence is highlighted, and so is the word being spoken, at an estimated position.
This splitting also lets you tap a sentence to jump to it, move to the previous or next one with the arrow keys, and come back later to the exact sentence where you stopped.
Why does a voice still make mistakes sometimes?
Because every step can guess wrong: an unknown acronym, a foreign name, a word whose pronunciation depends on its meaning. The model reproduces what it has learned, and a rare case can slip through.
If a word keeps being mispronounced, switching voices can help, since each model has its own strengths. The page on neural voices explains where Readotte's voices come from and how to choose one; for English texts specifically, see English text-to-speech.
Listen to your document
PDF with selectable text, 25 MB max. Scanned PDFs (images only) are not read yet.
Free, no account: about 10 minutes of listening a day. The text is extracted and read by our servers in Europe.
Frequently asked questions
Does Readotte's text-to-speech go through a third-party service?
No. The text is spoken by our own servers in Europe using open models; no third-party text-to-speech API is used. Sentence audio is cached for 30 days at most.
Does Readotte read emoji and symbols?
No, symbols, emoji and icons are skipped so they do not interrupt the reading. The text around them is read normally.
Can I correct how a word is pronounced?
Not directly. You can switch voices or edit the pasted text, for example by spelling an acronym the way it is said, before starting playback with turn text into speech.