What is Text-to-speech?

Text-to-speech, or TTS, reads written text aloud in a synthetic voice. It powers audiobooks, navigation prompts, accessibility tools, video narration and the voice of any AI assistant that talks back.

How it works

Older systems stitched together tiny recorded sound fragments, which is why they sounded robotic. Current systems are neural networks trained on many hours of recorded speech. Given text, the model predicts the sound of natural speech, including rhythm, stress and intonation, and a second component turns that prediction into an audio waveform. Many newer systems generate audio directly, one small piece after another, the way a language model generates words.

Good TTS handles things a simple reader would get wrong: how to say "read" (present or past), how to read "$4.50" or "Dr.", and when a sentence should rise like a question.

What you can control

• Voice. Most services offer a library of preset voices. Some can create a custom voice from a sample of someone's speech, which should only be done with that person's permission.
• Language and accent. Coverage varies a lot by language, so test the ones you need.
• Style. Pace, pitch and sometimes emotion, such as calm, upbeat or serious.
• Streaming. Audio can start playing before the whole sentence is generated, which keeps conversation feeling quick.

Why it matters to you

TTS is the output half of voice AI. How natural it sounds, and how quickly the first sound arrives, decides whether a voice product feels pleasant or irritating. Pricing is usually per character or per minute of audio, so cost grows with how much your product talks.

An example

You paste a blog post into a TTS tool and get a podcast-style audio version in minutes. A listener who is driving can hear it, and so can a reader who prefers audio.

Things to watch

• Cloned voices can be misused for fraud and impersonation. Reputable services require consent and some mark output as synthetic, a goal shared with AI watermarking.
• Pronunciation of names, acronyms and brand terms often needs fixing by hand.

Related terms

Text-to-speech is the opposite of speech recognition. Combine both with a language model and you get a voice agent. Like other generative systems, it depends on training data, here recorded speech, which raises its own questions about consent.