Text-to-Speech

Text-to-Speech is software that reads written text aloud. You enter a sentence, and a voice comes out of the speaker speaking that sentence. The name, translated plainly, means “text to speech,” abbreviated TTS. In the past, such voices sounded tinny and choppy because individual sounds were bluntly strung together. Today, programs generate the sound themselves and hit pauses, emphasis, and speaking tempo remarkably well. For many listeners, the difference from a real human is barely audible anymore.

From Reading Aid to Standard Interface

Originally, reading text aloud was primarily an aid for blind and visually impaired people. A so-called screen reader reads out what’s on the screen to them. The more natural a voice sounds, the longer one can listen to it without getting tired. What started as an aid for a small group has thus become a technology for everyone.

Text-to-Speech became economically interesting because it makes voice recordings cheap. A professional voice actor in a studio costs three-digit amounts per hour and needs a new appointment for corrections. A synthetic voice produces the same amount of text in seconds and can be changed as often as needed. Audiobook publishers, advertising agencies, and news sites now use this on a large scale.

This also creates a problem. If any voice can be recreated from just a few minutes of recording, fake voice messages become easy to produce. There have already been fraud cases in which callers imitated the voice of a supervisor to trigger bank transfers. That’s why the EU’s AI Act requires that artificially generated voices be labeled as such.

The Path from Letter to Sound Wave

In the first step, the system analyzes the text. It decides how numbers, abbreviations, and foreign words are pronounced. “1990” can be a year or an amount, and you can clearly hear the difference. From the text, a phonetic transcription is created — a chain of speech sounds with information on emphasis and pauses.

In the second step, a neural network predicts how this chain of sounds should sound. A neural network is a computational model that has learned patterns from many examples, in this case from hundreds of hours of real speech recordings. It first generates a kind of sound map that records, for each moment, which pitches are present and how strongly. A second network, the vocoder, then turns this map into the actual sound wave.

The comparison to a musician works well here. The first network writes the notes along with performance instructions, the second plays them on the instrument. To imitate a specific person, a voice profile is added. Often just a few minutes of recording are enough for the system to adopt the tone and speaking habits. This process is called voice cloning.

Where Synthetic Voices Speak Today

You most often encounter Text-to-Speech in voice assistants on phones and in navigation devices. Announcements at train stations and on trains are also often no longer recorded but generated in the moment. This is practical because delays and platform changes can’t be recorded in advance. Video platforms also offer automatic dubbing for content in foreign languages.

In business news, the term usually comes up in connection with companies like ElevenLabs, OpenAI, or Google. Such providers sell access to their voices as a service, billed by character or audio minute. At the same time, legal disputes are underway over who owns a recreated voice. Voice actors are demanding compensation when their voice was used as a template.

A common misconception is confusing Text-to-Speech with speech recognition. Speech recognition, known as Speech-to-Text, works exactly the other way around, turning spoken words into written text. Voice assistants contain both technologies, connected by a language model in between. Only this chain of listening, understanding, and responding results in a conversation with a machine.