What is text-to-speech? Explained with an audiobook narrator
Text-to-speech is AI that reads written text aloud in a natural-sounding voice. Think of an audiobook narrator turning a printed script into a recording.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
In one sentence: Text-to-speech is AI that reads written text out loud in a natural-sounding voice, so apps, phone lines and websites can speak to people instead of only showing words.
The everyday picture
Picture an audiobook narrator in a recording booth. They are handed a printed script. Before they start, they work out how to say tricky names, where to pause, and which words to stress.
Then they read it aloud, with rhythm and tone, and the studio saves the recording. Anyone can now listen to the book instead of reading it.
Text-to-speech (also called TTS, or speech synthesis) is that narrator, done by software. You hand it text, and it hands back audio of a voice reading that text.
Like a narrator, the text-to-speech model works out how each word sounds, plans the pace and emphasis, reads it aloud, and hands a recording to your app.
How it works, step by step
- Your text comes in (the script handed to the narrator). It might be a web page, a message for a phone menu, or a reply written by a chatbot.
- The text is prepared for reading (the narrator marking up the script). The system works out how each word should sound, broken into small speech sounds called phonemes. It also has to decide how to say things like dates, numbers and short forms, which can be read more than one way.
- It plans the delivery (the narrator deciding on pace and emphasis). This rhythm, stress and pitch is called prosody. You can also give the system "director's notes" in a markup language called SSML, to add pauses, change the speed or fix how a word is said.
- A trained model creates the voice (the narrator reading aloud). In Amazon Polly's neural voices, one neural network turns the speech sounds into a sound picture called a spectrogram, and a second one, called a vocoder, turns that picture into actual sound waves.
- The audio is delivered (the finished recording). Your app can play it straight away, stream it during a call, or save it as a file such as an MP3.
Try it · Put it in order
Put the narrator's work in the right order
Tap the steps in the order they happen.
Where the picture stops
- It does not understand the story. A narrator knows when a line is a joke or bad news. The model predicts how the text should sound, so it can stress the wrong word or read a sad line in a cheerful tone.
- It can stumble on names and jargon. Microsoft notes that synthetic voices may mispronounce niche or specialist words, names and uncommon phrases. Unlike a narrator, it will not stop and ask.
- The voice is not a person in a booth. It is generated fresh each time. Custom voices built from a real person's recordings raise questions of consent that a normal narrator contract does not.
Why it matters for your business
- Access. AWS notes that text-to-speech began as a tool for people with visual impairments. Reading pages aloud still helps them, and people who prefer to listen.
- Speed and cost. Updating a phone menu, a training module or a product video can mean editing text, not booking a new recording session.
- Customers. AWS lists customer service and call centres among the main uses: spoken answers that sound natural, in many languages.
- Risk. Microsoft's guidance is to tell people when a voice is synthetic, and it requires written permission from the person whose voice a custom voice is built on. Hiding that a voice is AI can damage trust and, in some places, break the law.
An example
Imagine a 25-person travel agency that sends booking updates by phone. Staff record a new message every time a schedule changes, which takes hours. With text-to-speech, the update is typed once and read out by a stock synthetic voice. The agency adds the right pronunciation for its destination names using SSML, and the message opens by saying it is an automated voice. A staff member listens to each new message before it goes out.
Try it · What would you do?
How should the agency launch its automated voice?
A travel agency wants a synthetic voice to phone customers with booking changes. One manager suggests building a custom voice from recordings of a popular former employee so it sounds familiar.
Pick what you would do. You can change your mind.
Frequently asked questions
Does text-to-speech sound like a real person?
Modern neural voices sound much closer to human speech than older systems, which stitched together recorded snippets. They can still mispronounce rare words or put the stress in the wrong place, so it is worth listening before you publish.
What is SSML?
SSML (Speech Synthesis Markup Language) is a set of tags you add around your text to tell the voice how to read it: where to pause, how fast to speak, and how to pronounce a word. Most major services support it, though not every tag works with every voice.
Can I make a voice that sounds like a specific person?
Some services offer custom voices trained on recordings of one speaker. Microsoft keeps this behind an approval process and requires written consent from the voice talent. Never copy someone's voice without permission.
Is text-to-speech the same as speech-to-text?
No, they work in opposite directions. Speech-to-text listens and writes; text-to-speech reads and speaks. Voice assistants often use both.
Related explainers
The DevDrift builds AI automation with a person approving what goes out: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.