AI Trend Research

← AI glossary

AI glossary · Beginner · Speech & audio

TTS

TTS (Text-to-Speech, speech synthesis) is technology that turns written sentences into speech that sounds like a human voice.

About this term
Level
Beginner · Core ideas for first-time readers
Category
Basics
Added
2026-10-08
In plain words It is similar to a voice actor who takes a script and reads it aloud. However, an AI voice actor does not so much understand the meaning of sentences and add emotion the way a person does; rather, it produces plausible intonation by following patterns from the many voice recordings it learned from.

How it works

First, a text analysis step expands numbers and abbreviations into spoken words and decides pronunciation and where to pause. Next, an acoustic model uses this information to create an intermediate representation (something like a blueprint of the sound) that captures pitch, duration and timbre. Finally, a component called a vocoder turns this blueprint into a speech waveform you can actually hear. Recently, combining these steps into a single deep learning model, or splitting sound into tokens and generating them in sequence as LLMs do, has also become widely used.

Basic TTS processing flowInput sentence → Text analysis; Text analysis → Acoustic model; Acoustic model → Vocoder(Blueprint); Vocoder → Speech outputInput sentenceText analysisPronunciation andpausesAcoustic modelPitch, duration,timbreVocoderGenerate the soundwaveformSpeech outputBlueprint
Basic TTS processing flow The steps that turn text into sound, from left to right. Recently, the middle steps are often combined into a single model.

Why it matters

TTS is used wherever a 'screen you listen to' is needed, such as voice assistants, navigation, audiobooks and screen readers for blind and visually impaired people. Combined with speech recognition (ASR) and an LLM, it becomes an AI assistant you can talk to. Recent TTS has become natural enough to be hard to tell apart from a human voice, and it is now possible to specify emotion or speaking style. As small models reach usable quality, on-device TTS that runs directly on phones is also growing.

Things to know

As 'voice cloning', which can imitate a specific person's voice from just a short recording, has become possible, concerns about misuse in voice phishing and deepfakes have grown. As a result, many services add AI watermarks to generated speech or include mechanisms to confirm the person's consent. TTS can also still make awkward mistakes with proper nouns, loanwords, reading numbers and context-dependent intonation.

Example

The news covered on this site, 'Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice TTS', is a good example. It describes using Kokoro-82M, a relatively small open TTS model, as the teacher model and distilling it into a much smaller student model that produces only one voice. By reducing the number of voices in exchange for a lighter model, it shows a trend toward making TTS easier to run on phones and small devices.

Where it came up on this site

Related terms

Speech recognition Multimodal Distillation On-device AI AI watermark Token