Text-to-Speech (TTS)

Text-to-Speech literally means “text to speech.” It refers to a computer program that reads written words aloud. You enter a sentence, and the program outputs an audio file. The abbreviation for this is TTS. Early versions sounded tinny and choppy because they stitched together individual recorded sound fragments. Today’s systems instead generate the sound from scratch, and they capture stress, speaking pace, and pauses remarkably well.

Why machines should learn to speak

Speech is the most convenient way to take in information when your eyes are busy. While driving, cooking, or exercising, having something read aloud is more practical than looking at a screen. That’s precisely why TTS is built into navigation devices, voice assistants, and podcast apps.

The technology is especially important for people with visual impairments. A so-called screen reader reads them the entire screen content, including menus, emails, and web pages. People with dyslexia also benefit from it. For them, TTS isn’t a convenience—it’s what makes using a computer possible at all.

TTS became economically interesting once the voices started sounding real. Recording an audiobook costs a narrator’s fee and studio time. A synthetic voice produces the same amount of text in minutes and for mere cents. Publishers, advertising agencies, and video platforms now use this on a large scale. At the same time, professional voice actors fear for their livelihoods.

From letter to sound wave

The first step is cleaning up the text. “5/12/2024” must become “the twelfth of May, twenty twenty-four.” Abbreviations like “Dr.” or “mph” also need to be spelled out in full. This step is called text normalization, and while unspectacular, it’s prone to errors.

After that, the system translates the letters into sounds. This isn’t trivial in German, because the same sequence of letters can sound different depending on context. “Wachstube” can mean either a guardroom or a tube of wax, depending on its meaning. The program therefore has to take the meaning of the sentence into account. Modern models learn to handle such cases from huge amounts of text and audio examples.

In the final step, the actual sound wave is generated. A neural network—a computational model trained on examples—produces thousands of numerical values per second for this purpose. These values describe how the air moves during speech. In doing so, the model also determines the melody of the sentence. A question rises at the end, a statement falls. It’s precisely this nuance that distinguishes natural-sounding systems from robotic ones.

Synthetic voices in everyday life and their dark side

Most commonly, you encounter TTS on your phone. Voice assistants use it to answer, navigation apps use it to announce the next turn, and translation apps use it to speak foreign languages aloud. On video platforms, many explainer videos are voiced with an artificial voice. Often you only notice it when a proper name is stressed incorrectly.

In the news, the term usually comes up alongside the keyword voice cloning. Just a few seconds of recording are enough for a system to imitate a particular voice. Scammers use this for calls in which a relative supposedly in distress needs money. The EU’s Artificial Intelligence Act therefore requires that artificially generated voices be labeled as such.

TTS is frequently confused with the reverse process. Speech recognition, known as speech-to-text, turns spoken words into written text. It’s used in dictation functions and automatic subtitles. A voice assistant needs both: first listening, then answering.