Ask what is text to speech and the short answer is unhelpful: software that reads written text aloud. That definition covered the flat robotic voices of twenty years ago and it covers today's systems, which can hold a listener for an hour without anyone noticing. The interesting part is everything the definition hides.
Modern synthesis is very good in English. It is markedly less good in most other languages, and it falls apart in a specific, predictable place: the sentence that contains two languages at once. For anyone producing content for more than one market, that gap is the whole story.
What Actually Happens Between Text and Sound
A synthesis engine runs several stages before any audio exists.
First it normalises the text, converting anything that is not a plain word into the words a person would say. Numbers, dates, currency, abbreviations. This stage is where the entertaining failures live, because St. is either Saint or Street, Dr. is either Doctor or Drive, and 1/2 is either a half or the first of February depending on where you are standing.
Then it converts spelling into sounds, which in English is a genuine research problem rather than a lookup. Then it predicts prosody, the rhythm and pitch contour that make a sentence sound like a statement rather than a question. Finally a neural vocoder turns that specification into a waveform. The full history of speech synthesis is essentially the story of each of those stages moving from hand-written rules to learned models.
The Errors That Give It Away
Homographs remain the classic tell. Read, lead, live, bass, wind, tear, close. Each is two different words wearing the same spelling, and choosing correctly requires understanding the sentence rather than the word. Modern models get this right most of the time, and the residual failure rate is exactly high enough to be noticeable across a long script.
The second tell is emphasis. Human speakers stress the word that carries the new information, and that choice depends on what was said thirty seconds earlier. Synthesis works sentence by sentence, so it produces a technically correct reading with the weight in slightly the wrong place, over and over, which listeners experience as a vague flatness they cannot name.
Both problems have the same practical fix, which is markup. The W3C's Speech Synthesis Markup Language lets you specify pronunciation, insert pauses, set emphasis and override the engine's guess phoneme by phoneme. Any serious production workflow using text to speech software involves a pass of SSML correction, and the scripts that sound best are the ones that were edited, not merely submitted.
Why Other Languages Lag
Three reasons, and only one of them is technical.
The first is data. English has an order of magnitude more recorded, transcribed, rights-cleared speech than almost anything else, and model quality tracks that directly.
The second is variety collapse. Ask for spanish text to speech and you will be offered a small number of voices that flatten an enormous range. The intonation of Buenos Aires, Mexico City, Bogotá and Madrid differ in ways that native listeners register instantly, and vocabulary differs too. A voice that sounds neutral to a developer in California can sound distinctly foreign to the audience you are actually addressing.
The third is code switching. Real multilingual text is full of English product names, place names and technical terms embedded in another language. Engines are trained on one language at a time, so they either apply the wrong phonetics to the foreign word or switch accent mid-sentence. It is the single most audible failure mode in multilingual synthesis and there is no clean workaround beyond hand-tuning each occurrence.
Cost, and What It Hides
Cloud synthesis is billed per character, with higher-quality neural voices priced above standard ones. Anyone comparing google cloud text-to-speech pricing against a competitor will find the headline rates broadly similar and the real cost somewhere else entirely: in the number of regeneration cycles a script needs before it is usable, and in the human hours spent writing the markup that gets it there.
A short script that runs clean is nearly free. A long script full of names, numbers and two languages is not, and budgeting for it as pure compute is how projects overrun.
Where a Recorded Voice Still Wins
Synthesis has genuinely won the categories where consistency beats warmth: system prompts, navigation, accessibility, internal training, anything that changes weekly and must sound identical every time. Nobody should be paying a studio to re-record a menu.
It has not won performance. Anything that needs a character, comic timing, emotional range or a specific regional identity is still recorded, and that gap has narrowed less than the demos suggest. The same distinction shows up in screen localization, where dubbing meaning and method are built entirely around matching a performance rather than reproducing words.
The commercial argument sits underneath all of this. Audiences respond to content that sounds like it was made for them rather than adapted towards them, and the research on why consumers buy in their native language is consistent on the point. A synthetic voice with the wrong accent communicates the information and quietly undermines the reason you translated anything.