speech synthesis (TTS)

in English: text-to-speech, speech synthesis

basicwidespread

Speech synthesis is a technology that converts written text into machine-generated speech resembling a human voice.

In plain terms

Imagine you have a tireless reader: whatever text you give them, they read it aloud immediately, even at 2 a.m. Speech synthesis does exactly this, only a program, not a human, stands behind the voice.

In the past, these machine voices were choppy and robotic. Today's AI-based systems stress words, pause, and can even convey emotion, so they are often hard to tell apart from a live announcer.

Good for many things: it reads navigation instructions aloud, creates audiobooks, helps the visually impaired, and also provides the voice for AI agents working in telephone customer service. It differs from a recording in that it does not play pre-recorded sentences but can pronounce any new text.

How it works

The system first prepares the text: it expands abbreviations and writes out numbers and dates (so '2026' becomes 'two thousand and twenty-six').
It then determines how to pronounce words, that is, it converts letters into speech sounds.
A neural network plans the melody of a sentence: where the emphasis falls, where the pauses go, and how fast the tempo is.
Another model generates actual sound waves from this "score," which you hear through the speaker. In the latest systems, these steps are no longer sharply separated: often a single large model performs the entire task from text to audio.
With modern systems, you can also specify the voice's character (male, female, cheerful, serious), and some can even imitate a specific person's voice from a few seconds of sample.

An everyday example

When the car navigation announces, 'turn right in three hundred meters,' you are hearing speech synthesis. The same happens when your phone reads out a message or when you start a news portal article with the 'listen' button.

Who is affected

everyoneIn navigation, voice assistants, telephone customer service, and read-out articles, you hear machine voices day in, day out, often without noticing.

parent, student, teacherThe read-aloud function supports students with reading difficulties or visual impairments and helps practice pronunciation in language learning. → Education & Family

marketerNarration for ads, videos, or podcasts can be created in minutes without a studio or announcer, even in multiple languages. → Marketing

entrepreneur, executiveVoice-based customer service agents can receive calls around the clock, but the customer must know they are speaking to a machine.

A common misconception

MythSynthetic voices can always be identified because they sound robotic and monotone.

In realityThe voice of today's systems is often deceptively human; on a short recording, most listeners cannot distinguish it from live speech. Therefore, you should be suspicious of the situation, not your ears.

Limits and risks

A person's voice can be cloned from a few seconds of audio, so if an 'acquaintance' urgently asks for money by phone, call them back on their usual number.
With rare names, foreign words, abbreviations, and ambiguous sentences, the machine can still make mistakes in pronunciation or stress.
The quality and variety of Hungarian-language voices generally lag behind English, though the gap is narrowing quickly.
Unauthorized imitation of another person's voice raises legal and ethical issues and directly affects the livelihoods of actors and voice-over artists.

Similar, but not the same

speech recognition – Speech recognition works in the opposite direction: it converts spoken words into written text.
voice cloning – Voice cloning replicates a specific person's voice, while speech synthesis is a broader term and mostly works with ready-made, 'nobody-owned' voices.
voice assistant – The voice assistant understands your question and answers it; speech synthesis is only the final step that speaks the answer aloud.

Related concepts

Broader

generative AI, speech technology, synthetic media

Narrower

voice cloning, neural speech synthesis, Machine dubbing

On the edge

speech recognition, voice assistant, deepfake, voice-based agent, music generation

A brief history

One of the pioneers of machine speech was the Hungarian Farkas Kempelen, who described his mechanical speaking machine in a book published in 1791. The era of modern, natural-sounding systems began in 2016, when DeepMind introduced the WaveNet model based on neural networks.

Where things stand

By the end of 2026, speech synthesis is everyday technology: it is present in phones, cars, customer service, and content production, and the best systems speak in real time with emotions. Attention has shifted to preventing abuse. Under the transparency rules of the EU AI Act (Article 50), artificially generated audio must be marked in a machine-detectable way, and it must be disclosed that a deceptively realistic forgery (deepfake) is artificial. These rules apply from August 2026, partly with transitional periods.

Status: 2026 November

Also known as

speech synthesis · text-to-speech · TTS · text-to-speech reader · machine speech · text-to-speech · speech synthesis · AI-hang

Sources

van den Oord et al.: WaveNet: A Generative Model for Raw Audio (2016)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.