Auditext

How text to speech works

Between the text you type and the voice you hear there are four jobs: work out what the text says, how each word sounds, how the sentence should move, and finally the sound wave itself. Type a sentence and step through them.

1 · Text normalisation: what does it say?

A speech engine can’t pronounce “$4.50” or “Dr.” — it first rewrites the text into the words a person would actually say. This is harder than it looks: “Dr.” is “Doctor” before a name but “Drive” after one, and “3/14” is a date in the US but nonsense in the UK.

Doctor Lee paid four dollars and fifty cents on March fourteenth, twenty twenty-five at nine thirty a m — was it really twenty-five percent off?

  • Currency $4.50 → four dollars and fifty cents
  • Date (US order assumed) 3/14/2025 → March fourteenth, twenty twenty-five
  • Time 9:30 am → nine thirty a m
  • Percent 25% → twenty-five percent
  • Abbreviation Dr. → Doctor

From rules to learned voices

Formant synthesis (1970s–90s) built speech from acoustic rules — the voice Stephen Hawking used came from this family. Intelligible, but unmistakably machine-like.

Concatenative synthesis (1990s–2010s) recorded one voice actor for many hours, chopped the recording into tiny units and re-assembled them. It sounded human word by word, but joins and flat intonation gave it away.

Neural synthesis (2016 onwards) replaced the pieces with learned models. Systems like Google’s Tacotron predicted a spectrogram from text, and a neural vocoder turned it into sound. Current models go further: they take a written direction (“whisper this”) and even switch languages mid-sentence, because they have learned what those things sound like rather than following rules.

What this means for your scripts

  • Normalisation guesses. If a number, date or abbreviation matters, write it out.
  • Pronunciation guesses too. Respell unusual names phonetically.
  • Prosody follows punctuation. Commas, full stops and question marks are your cheapest direction.
  • For precise control, use pauses and emphasis — see the SSML guide — or give the voice a delivery direction in the AI voice generator.

Questions people ask

How does text to speech work?

A TTS system first normalises the text (expanding numbers, dates and abbreviations into words), then converts words into phonemes, predicts prosody — the timing, stress and pitch melody — and finally generates the audio waveform. Modern neural systems learn the last three steps together from recordings of real speech.

Why did old text to speech sound robotic?

Early systems either generated sound from simplified acoustic rules (formant synthesis) or stitched together small recorded fragments (concatenative synthesis). Both produced audible joins and flat, rule-based melody. Neural models such as DeepMind’s WaveNet (2016) and later end-to-end systems predict the waveform directly, which removed most of those artefacts.

What is prosody?

Prosody is the music of speech: how long each sound lasts, where the stress falls, how loud it is, and how pitch rises and falls across a sentence. The same words with different prosody can be a statement, a question or sarcasm.

Why does TTS mispronounce names?

Names often do not follow a language’s spelling rules and may be rare in training data, so the model guesses from similar-looking words. Spell the name phonetically in your script ("Siobhan" as "Shiv-awn") or use an SSML phoneme tag where supported.

Can AI voices express emotion?

Yes. Current models can follow natural-language directions like "excited" or "whispering" because they learned how those states change timing, pitch and voice quality from real recordings. The emotion is performed, not felt, and results vary by voice and language.

More from the studio