Text-to-Speech and Synthetic Voices: How Computers Learned to Talk

Not long ago, a computer voice was easy to spot. It was flat, choppy, and oddly paced, with every word pronounced in the same monotone. Today, many synthetic voices are hard to tell apart from a real person, complete with natural rhythm, emphasis, and even the small pauses people make when they speak.

The technology behind this is called text-to-speech. It powers everything from screen readers and satnavs to phone menus and AI voice agents. This article explains how it works, how it developed, and what makes a synthetic voice sound natural, or not.

What Is Text-to-Speech?

Text-to-speech (TTS), also called speech synthesis, is technology that turns written text into spoken audio. You give it a sentence, and it produces a recording of that sentence being spoken in a computer-generated voice.

It's the reverse of speech recognition. Speech recognition turns voices into text; text-to-speech turns text into voices. Together, they allow computers to hold spoken conversations.

A Brief History

Early machines

One of the first electronic speech synthesizers was the Voder, built at Bell Laboratories and demonstrated at the 1939 New York World's Fair. It didn't read text. A trained operator "played" it using keys and a foot pedal to produce speech sounds, and it took months of practice to make it say anything intelligible.

Rule-based voices

From the 1960s onward, researchers developed systems that generated speech using rules about how human speech sounds are formed, known as formant synthesis. These voices were robotic but clear and could say anything. Work by researcher Dennis Klatt in this period produced some of the best-known synthetic voices, including systems that underpinned the voice used for many years by the physicist Stephen Hawking.

Stitching recordings together

In the 1990s and 2000s, concatenative synthesis became common. Developers recorded a real person reading many hours of material, cut the recordings into tiny pieces of sound, and then stitched those pieces together to form new sentences. The results sounded more human, but the joins between pieces could be noticeable, and the voice couldn't easily change its tone or emotion.

The neural network breakthrough

The major breakthrough came in the mid-2010s. In 2016, DeepMind introduced WaveNet, a neural network that generated speech audio directly, producing far more natural voices than previous methods. Soon after, systems like Google's Tacotron showed that neural networks could learn to turn text into speech from examples, without hand-written rules.

Since then, neural TTS has become the standard. Modern voices can sound warm, expressive, and conversational, and they can be generated quickly enough to use in live conversations.

How Text-to-Speech Works

Although modern systems are often built as large neural networks, the process can be understood in a few stages.

1. Cleaning up the text

Written text contains many things that aren't spoken exactly as they're written. The system first converts the text into words that can be pronounced. This is called text normalization. For example:

• "$25.50" becomes "twenty-five dollars and fifty cents."
• "Dr. Smith" becomes "Doctor Smith," but "Smith Dr." might become "Smith Drive."
• "3/4" could be "three quarters," "March fourth," or "April third," depending on context.
• "2026" is read as a year in one sentence and as a quantity in another.

Getting this right depends on understanding the surrounding words.

2. Working out pronunciation

Next, the system decides how each word should sound. English spelling is notoriously inconsistent ("though," "through," "tough," and "cough" all end differently), so systems use pronunciation dictionaries along with models that predict the pronunciation of unfamiliar words.

Some words are spelled the same but pronounced differently depending on meaning. "I will read it" and "I have read it," or "lead the team" and "made of lead." The system has to use context to choose correctly.

3. Planning the rhythm and tone

Natural speech isn't just a string of correctly pronounced words. It has prosody: the rise and fall of pitch, the timing of words, where emphasis falls, and where pauses come. Prosody is what tells a listener whether something is a statement or a question, what's important in a sentence, and how the speaker feels.

Compare "I didn't say she took it" with the emphasis placed on different words. The words are identical, but the meaning changes each time. Predicting natural prosody is one of the hardest parts of speech synthesis, and it's where modern neural systems have improved the most.

4. Generating the sound

Finally, the system produces the audio. Many systems first generate an intermediate representation of the sound, similar to a spectrogram, and then use a component called a vocoder to turn that into an actual audio waveform. Other modern systems generate the audio more directly. Either way, the result is a recording of the text spoken in the chosen voice.

Streaming for conversations

For live uses, such as AI voice agents, speed is critical. Modern TTS systems can stream audio, starting to speak the first part of a sentence before the rest has been generated. This allows replies to begin in a fraction of a second.

Controlling How a Voice Speaks

Many TTS systems let developers fine-tune how text is spoken using a standard called SSML (Speech Synthesis Markup Language). SSML can be used to add pauses, change speaking speed or pitch, emphasize particular words, and specify how numbers, dates, or abbreviations should be read, for example telling the system to read "1234" as "one two three four" rather than "one thousand two hundred thirty-four."

Newer systems can also adjust speaking style based on simple instructions, such as sounding calmer, more upbeat, or more formal.

Types of Synthetic Voices

TTS providers typically offer several kinds of voices.

Standard voices are ready-made voices in a range of languages, accents, and genders. They're clear and reliable, though sometimes less expressive.

Premium or neural voices are more natural and expressive, with better prosody and more human-like delivery.

Custom voices are created specifically for a business or brand, often by recording a professional voice actor and training a model on those recordings. This gives the business a distinctive voice that no one else uses.

Cloned voices are created from recordings of a particular person's voice. Voice cloning raises important questions about consent and misuse, which are covered in a separate article.

Where Synthetic Voices Still Struggle

Despite huge progress, TTS isn't flawless.

Names and unusual words. Personal names, place names, brand names, and technical terms are often mispronounced, especially when they come from other languages.

Numbers and formats. Phone numbers, account numbers, dates, and prices can be read in the wrong format if the context is unclear.

Emotional fit. A voice that sounds cheerful when delivering bad news, or flat when expressing sympathy, can feel jarring.

Long passages. Over long stretches, some voices drift in tone or pacing, or repeat the same intonation patterns until they sound mechanical.

Phone audio. A voice that sounds rich and clear through good speakers can lose some of its quality over a narrowband phone line.

Where Text-to-Speech Is Used

Accessibility. Screen readers use TTS to read on-screen text aloud for people who are blind or have low vision, and TTS also helps people with reading difficulties.

Navigation. Satnavs and map apps announce directions.

Phone systems. IVR menus and automated messages often use TTS instead of studio recordings, making them quick to update.

AI voice agents. TTS gives AI systems their spoken voice during live phone conversations.

Voice assistants. Smart speakers and phone assistants respond using synthetic voices.

Content. TTS is used to create narration for videos, e-learning courses, and audiobooks.

Speech support. People who've lost the ability to speak can use TTS to communicate, sometimes with a voice built from earlier recordings of their own speech.

Choosing a Voice for Phone Calls

For businesses using synthetic voices on calls, a few practical considerations matter:

• Clarity over character. On a phone line, a clear, steady voice is more important than a distinctive one.
• Suitable pace. A slightly slower pace than everyday speech helps callers follow, especially when hearing important details.
• Accent and familiarity. Voices with accents familiar to the people being called tend to be easier to understand and more comfortable to listen to.
• Consistency. Using the same voice across a business's calls helps callers recognize it.
• Testing on real phone lines. Voices should be tested over actual phone calls, not just through computer speakers, since the audio sounds different.
• Honesty. As synthetic voices become more convincing, being clear with callers that they're hearing an AI or automated voice matters more, not less.

A Voice Without a Speaker

Text-to-speech has travelled from a machine that took months to learn to play, to voices that can speak naturally in real time. That progress has made computers far more useful, from helping people with disabilities to powering automated phone calls that sound genuinely conversational. It also means a synthetic voice can now be mistaken for a real one, which is why the way these voices are used, and how openly, matters just as much as how good they sound.

Related articles

• How an AI Voice Agent Hears, Thinks, and Speaks: The Path of One Reply
• AI Voices on Phone Calls: Disclosure, Consent, and the Law