What is text to speech (TTS)?

From characters to a voice

A TTS system does several jobs in sequence before any sound exists. It first normalises the text, deciding that 3.5 is three point five, that Dr before a name is doctor and that 2026 is a year. It then works out pronunciation, including words spelled the same but said differently. Next comes prosody, the pattern of pitch, stress and pauses that makes a sentence sound like a question, a list or a warning. Only then is audio produced. Most errors people notice in synthetic speech come from the early steps, a misread number or a stressed wrong syllable, rather than from the sound of the voice itself.

How the technology developed

DeepMind's 2016 WaveNet paper set out the two approaches that came before. Concatenative systems recombined fragments from a large database recorded by a single speaker, which sounded clear but meant fresh recording sessions for any new voice or emotion. Parametric systems generated speech from model parameters through signal processing vocoders, which was flexible but sounded artificial. WaveNet modelled the raw waveform directly and, in blind listening tests for American English and Mandarin, reduced the gap with human performance by over 50 percent. Google now describes its Chirp 3: HD voices as its latest generation, powered by generative models that deliver realism and emotional resonance.

Controls that shape a read

The W3C Speech Synthesis Markup Language, SSML, is the long standing way to direct a synthetic voice: tags for pauses, emphasis, pronunciation, and how to say dates, numbers and abbreviations. Support varies by voice. Google's documentation lists SSML support for Chirp 3: HD voices as a preview, alongside pace control from 0.25 to 2 times normal speed, markup for short and long pauses, and custom pronunciations written in IPA or X-SAMPA phonetic notation. The more direction a voice accepts, the less a script has to be bent around its habits.

When it matters

TTS matters wherever speech must be produced faster, more often or in more languages than recording allows. A training video that changes every quarter can be re-voiced by editing text. A promo can be released in several languages without booking several artists. It keeps a series consistent, because the same voice reads the fortieth episode exactly as it read the first. Accessibility depends on it too, since screen readers turn interface text into speech for blind users. The trade offs are nuance and trust. A synthetic voice can mispronounce a local place name, and a voice presented as a real person raises disclosure questions that ordinary narration does not.

Common mistakes

Writing for the eye instead of the ear is the biggest. Long sentences with nested clauses read fine and sound breathless. Symbols and abbreviations cause the next set of problems, so write numbers, units and brand names the way they should be spoken. Choosing a voice by how impressive it sounds, rather than by the job, is a third: an excited promotional read over step by step instructions makes viewers doubt the steps. Speeding a voice up to fit a long line into a short scene is a fourth, since cutting words sounds better than a faster read. Finally, many people publish without listening to the whole track once at normal speed, which is where a single wrong stress on a product name gets caught.

Where it shows up in the product

The video builder here voices narration with Google Cloud Chirp3-HD voices. The generator chooses from four styles, each in two genders: energetic, Puck or Laomedeia; instructional, Charon or Erinome; warm, Umbriel or Sulafat; neutral, Iapetus or Aoede. A tutorial is always given instructional rather than energetic. Thirty one languages have a mapped locale; Malay and Filipino fall back to older Wavenet voices, and Persian, though suggested, has no mapping and gets the default American English voice. In the voice dialog the whole live Google catalogue is searchable per scene, with speed from 0.85x to 1.25x. Each synthesised clip costs one credit. The password reset example uses the Erinome voice.

Questions people ask