In plain words
The announcement at the railway station, the voice that reads your messages aloud while you drive, an audiobook made without a narrator: all of these can be text-to-speech, or TTS.
Reading aloud well is harder than it sounds. The computer has to know that "4" should be said "four", that "read" sounds different in "I read yesterday" and "I will read", where to pause, and which word to stress. Then it has to produce the sound itself, tens of thousands of tiny air-pressure values every second.
Older systems stitched together recorded snippets and sounded robotic. Modern ones are neural networks trained on many hours of recorded speech, so they learn the melody of a sentence and produce smooth, natural audio. Some can copy a particular person's voice from a short sample, which is useful for people who lose their voice, and also used by scammers.
Everyday example
A newsreader with a script
A radio newsreader gets a script. Before speaking, they mark it up in their head: say "Rs 500" as "five hundred rupees", pause after the headline, lift the voice on a question. Then they speak. Text-to-speech does the same two jobs: first work out how the text should sound, then produce the sound.
Where the example breaks: The newsreader understands the news and chooses a sad or cheerful tone on purpose. A TTS model has learned only how such sentences usually sound, so it can stress the wrong word or read a sad line brightly.
How it actually works
Most modern TTS has two stages: decide how it should sound, then make the sound.
Clean up the text
Numbers, dates, short forms and symbols are expanded into words ("Dr" to "Doctor", "4" to "four"), and words are converted to the speech sounds they stand for.
Plan the melody
A model predicts timing, pitch and loudness for each sound, the rhythm and intonation of the sentence. Often this comes out as a spectrogram: a picture of which pitches carry energy over time.
Make the waveform
A second network, called a vocoder, turns that plan into actual audio samples. DeepMind's WaveNet (2016) showed that a neural network could generate raw audio one sample at a time and sound far more natural.
Voices and cloning
Trained on many speakers, a model can switch voices. Given a short clip of a new voice, some systems imitate it. This is part of Generative AI, and responsible tools ask for consent and add watermarks.
- Turning point
- September 2016: WaveNet cut the gap between the best older systems and human speech by over 50% in listening tests, for US English and Mandarin
- Audio detail
- Digital speech is often 16,000 or more samples per second, each one a number the model must produce
- Everyday uses
- Screen readers for blind and low-vision people, navigation, announcements, audiobooks, dubbing, voice assistants
- The reverse job
- Speech recognition turns voice into text
If a familiar voice calls asking for money urgently, hang up and call back on a number you already have. Voices can be copied.
Questions people ask
How does text-to-speech work?
It first works out how the text should be spoken: expanding numbers, choosing pronunciations, planning pauses and pitch. Then a neural network generates the audio itself, thousands of samples per second, in a chosen voice.
What is AI voice cloning?
Voice cloning is text-to-speech that imitates a specific person, learned from a recording of their voice. It helps people who are losing their voice keep it, and lets creators dub their own videos. It is also misused in scam calls and fake clips, so consent matters.
How can you protect your family from voice clone scams?
Agree a family code word for emergencies. If a loved one calls in panic asking for money, hang up and call them back on a number you know. Be wary of urgent requests for transfers, gift cards or crypto. The US FTC puts it simply: "Don't trust the voice."
Is text-to-speech the same as speech synthesis?
Yes, almost. Speech synthesis is the broader term for any machine-made speech. Text-to-speech is speech synthesis that starts from written text, which is the most common case.
Why do some AI voices still sound robotic?
Older or lightweight systems join pre-recorded pieces or use simpler models, which causes flat melody and odd joins. Neural voices sound much more natural but can still slip on names, mixed languages, or the emotion a sentence needs.