What is text-to-speech?

Text-to-speech (TTS) is technology that lets a computer read written text aloud in a spoken voice. The read-aloud feature on your phone, automated announcements on phone lines and at stations, narration on videos and AI you can talk to by voice all rely on it.

Synthetic voices have become so natural that they can be hard to tell apart from a person. Understanding how text-to-speech works helps you fix mispronunciations, design better voice conversations and use voice-copying technology responsibly.

What does text-to-speech actually do?

Text-to-speech doesn’t play back a recording. It turns whatever text it’s given into speech on the spot. Say a yoga studio wants its phone line to announce, “Because of the storm, tomorrow’s morning class is cancelled.” Nobody has to record anything: the studio edits the text and the new message is ready.

Early systems recorded a person’s voice, cut it into tiny pieces and stitched those together. Today it’s common to use neural networks (systems that learn from huge numbers of examples) trained on large amounts of recorded speech paired with text, which has made the seams far less noticeable.

You’ll find text-to-speech in situations like these:

• Offering course materials or a small company’s announcements as audio as well as text
• Telling out-of-hours callers when you’re open
• Adding narration to a video straight from a script
• Reading a screen aloud so people who are blind or have low vision can use a website

How does text-to-speech work?

Working out how to read the text

The first job is deciding how the text should be read. Numbers, dates, symbols and abbreviations don’t come with instructions. “4/1 10:00” has to become “April first, ten o’clock” (or “the fourth of January”, depending on the date format). “Dr.” can mean “Doctor” or “Drive”, and “St.” can be “Saint” or “Street”.

Some words are spelled the same but pronounced differently. “Read” sounds different in “I read it yesterday” and “Please read it”, and “lead” is different in “lead the class” and “lead pipes”. The system uses the surrounding words to choose.

Once the words are settled, the system decides where the stress falls and where to pause. Stress can change meaning: “REcord” is a noun and “reCORD” is a verb.

Planning the sound and making the audio

Next, a model plans the voice: which pitch to use, how long each sound lasts and how loud it is. The plan is often a mel spectrogram, a chart of how strong the sound is at each pitch over time. Finally, a part called a vocoder turns the plan into an actual sound wave.

Some newer models treat audio as a sequence of short codes and build speech by predicting the next code, much as a large language model (LLM) predicts the next piece of text.

What makes a voice sound natural?

Whether speech sounds natural depends on two things: the voice and the way it speaks.

• Voice: the qualities that make a voice sound like a particular person, such as how high, deep or breathy it is
• Prosody: the rise and fall of pitch through a sentence, the speed, the pauses and the emphasis

Even a good voice sounds robotic with poor prosody. If a question doesn’t rise at the end, pauses come in the wrong places or key words get buried, the message becomes harder to follow.

You can often fix mispronunciations and awkward delivery like this:

Problem	What to try
Your class name or a person’s name is mispronounced	Add it to the pronunciation dictionary, or spell it the way it sounds
Numbers or dates are read the wrong way	Write them out unambiguously, for example “April 1” rather than “4/1”
Long sentences run together	Shorten sentences and add commas
Important points don’t stand out	Use a feature that lets you add pauses or emphasis

Developers can also use SSML (Speech Synthesis Markup Language), a W3C standard for marking up text with pronunciation, pauses, pitch, speed and volume. How much of it a service supports varies.

What matters when you use it in conversation?

When you talk to a voice assistant or an automated phone line, how quickly the reply starts matters a lot. A study of conversations in 10 languages found that the gap before the next person starts speaking averages about 200 milliseconds. We’re used to gaps that short, so when a reply is slow, people start to wonder whether they were heard or understood.

Most AI you can talk to chains three steps: speech recognition turns your voice into text, an LLM writes a reply, and text-to-speech speaks it. Each step takes time, so good voice experiences cut the wait like this:

• Start speaking the first sentence before the whole reply is finished (streaming)
• Keep spoken replies short, and leave long explanations to the screen or a text message
• Stop talking and listen as soon as the other person starts speaking

Some things don’t work well by ear alone. Web addresses, long numbers and tables are hard to remember after hearing them once, so send them as text too.

How should you approach voice cloning?

Text-to-speech can also produce a voice that sounds like a specific person. This is often called voice cloning. Some models and services can create a similar-sounding voice from a recording only seconds to tens of seconds long.

It has valuable uses. Someone at risk of losing their ability to speak because of illness can preserve their voice, then type what they want to say and have it spoken in that voice. Some phones include a feature built for exactly this.

But a voice belongs to the person it comes from. If you use voice cloning in your business, these steps keep you on solid ground:

1. Get written consent from the person whose voice you use, covering what it’s for, for how long and how they can withdraw.
2. Make sure listeners can tell the voice is synthetic.
3. Don’t imitate other people’s voices, especially well-known people’s, to promote anything. It can infringe their rights.
4. Read the service’s terms on how voices may be created and used.

Some major services won’t create a voice until the speaker records themselves reading a consent statement.

Be ready for scams that use cloned voices, too. The US Federal Trade Commission warns that a short audio clip can be enough to copy someone’s voice. If a caller claiming to be a family member, client or manager asks for money or confidential information, hang up and call them back on a number you already know. A familiar voice alone doesn’t prove who’s calling.

How does text-to-speech help accessibility?

Text-to-speech is an essential tool for people who find reading difficult. People who are blind or have low vision use websites with screen readers, software that reads the screen aloud. Read-aloud also helps people with dyslexia, people who struggle with small text, and anyone whose hands or eyes are busy.

To make your own site work well with screen readers:

• Set the page language, for example <html lang="en">. The W3C explains in Understanding Language of Page that this lets screen readers load the right pronunciation rules
• Give images alternative text that describes what they show
• Use headings to show how the page is organized
• Write link text that says where the link goes, rather than “click here”

It’s worth listening to your class sign-up form or online course page with a screen reader once, to check that it makes sense in the order it’s read aloud.