So, what is TTS? TTS stands for text to speech: software that takes written words and reads them aloud in a synthetic voice. It's the voice in your phone's screen reader, the one in your car's navigation and the narrator on many of the short videos you scroll past. This explainer covers what TTS means, how text to speech technology turns letters into sound, what it's used for, how it differs from voice cloning and speech to text, how to choose a TTS tool and where it still trips up.
What does TTS stand for? The meaning of text to speech
TTS stands for text to speech, also written text-to-speech. The term names both the technology and the feature built on it: you supply text, the system supplies audio. You'll also see it called speech synthesis (the technical name), text to voice or read aloud.
The input is always text: a sentence typed into a box, a web page, an e-book, a chat message. The output is speech, played live or saved as an audio file such as MP3 or WAV. Nobody records those words in advance. The system works out how they should sound at the moment you ask.
In everyday use, "TTS" often means one specific feature rather than the whole field:
- On a phone or computer, it's the built-in read-aloud or screen reader voice.
- On TikTok, it's the editor option that reads your on-screen captions in a synthetic voice.
- On Twitch, "TTS messages" are viewer donations or alerts that a stream tool reads out loud.
- In accessibility, it's the engine that lets a blind or low-vision person hear what's on the screen.
How text to speech technology works
Every TTS system, old or new, has to answer the same four questions. What words are actually on the page? How is each one pronounced? How should the sentence flow? And what does all that sound like as a waveform? The W3C's Speech Synthesis Markup Language 1.1 specification (2010) describes a similar chain for any synthesis processor, running from text normalization through text-to-phoneme conversion and prosody analysis to waveform production. Modern neural systems blur these steps together, but the four-stage picture is still the clearest way to see what's happening.
Step 1: Text normalization
Written text is full of things that aren't words yet. Normalization rewrites them into something speakable:
- "$4.99" becomes "four dollars and ninety-nine cents."
- In "Dr. Lee lives on Elm Dr.", the first "Dr." is "doctor" and the second is "drive."
- "1/2" could be "one half", "January second" or "the first of February", depending on context and region.
- "2026" is "twenty twenty-six" as a year but "two thousand twenty-six" as a count.
A large share of the mistakes you hear from a TTS voice start right here.
Step 2: Linguistic analysis
Next, the system works out pronunciation and structure. It converts each word into phonemes, the individual sounds of a language, using a pronunciation dictionary for common words and learned rules for everything else. It also has to choose between words that are spelled alike but said differently, such as "read" (present) and "read" (past), or "lead" the metal and "lead" the verb, by looking at the grammar around them.
Then comes prosody: which syllables get stress, where the voice rises (a question) or falls (a statement), and where it pauses. Punctuation does a lot of the work. A comma usually becomes a short break, a period a longer one.
Step 3: The acoustic model
The acoustic model takes the phonemes and the prosody plan and predicts what the speech should look like as sound features, usually a mel spectrogram: a picture of how energy across frequencies changes over time. This is where a voice's character lives, including pitch, timbre, speed and rhythm. In neural TTS, the acoustic model is a network trained on many hours of recorded speech.
Step 4: The vocoder
A spectrogram isn't audio yet. The vocoder turns those features into an actual waveform your speakers can play. Older vocoders produced the buzzy, metallic sound people associate with robot voices. Neural vocoders changed that. Google's Tacotron 2 paper (2017) paired a network that predicts mel spectrograms with a modified WaveNet vocoder, and reported a mean opinion score of 4.53, against 4.58 for professionally recorded speech in the same test.
From robotic to realistic: how TTS has changed
TTS is decades old, and the voice quality you hear today comes from several generations of technique. You'll still run into all of them.
| Generation | How it makes sound | What it sounds like | Where you still hear it |
|---|---|---|---|
| Formant synthesis | Builds speech from acoustic rules, no recordings | Clear but robotic | Older devices, retro "computer voice" effects |
| Concatenative | Stitches together short pieces of recorded speech | Natural in places, choppy at the joins | Older navigation units and phone menus |
| Statistical parametric | A model predicts speech parameters, a vocoder renders them | Smooth but muffled | Older assistants and embedded devices |
| Neural TTS | Neural networks predict spectrograms or audio directly | Close to human, with natural rhythm | Most current apps, assistants and AI voice tools |
DeepMind's WaveNet announcement (2016) summed up the old trade-off: concatenative TTS needs a very large database of recorded fragments and is hard to modify, while parametric TTS is flexible but tended to sound less natural. WaveNet modeled the raw audio waveform directly, and in DeepMind's listening tests it narrowed the gap between the best existing systems and human speech by over 50 percent, in both English and Mandarin Chinese.
The latest step borrows from large language models. The VALL-E paper (2023) treated TTS as a language modeling task over audio codes and could imitate an unseen speaker from a 3-second recording. That idea, a model that has heard enough voices to generalize from a short sample, is what makes today's zero-shot voice cloning possible.
What is text to speech used for?
Accessibility
This is where TTS matters most. The World Health Organization's fact sheet on vision impairment (2026) puts the number of people worldwide with a near or distance vision impairment at 2.2 billion or more. Screen readers such as VoiceOver on Apple devices, TalkBack on Android and Narrator on Windows use TTS to read out everything on the screen, from button labels to emails. People with dyslexia and other reading difficulties use read-aloud tools to follow text by ear as well as by eye.
Reading and learning
Browsers, e-readers and reading apps can read articles and books aloud, which helps on a commute, during chores or when you're learning a language. If you read on a Kindle, our guide to Kindle text to speech goes through the options.
Navigation and smart devices
Turn-by-turn directions can't be pre-recorded for every street name on earth, so map apps synthesize them on the fly. Smart speakers, car dashboards, ticket kiosks and home appliances rely on the same idea.
Video voiceover and content
Creators use TTS to narrate explainers, product demos, slideshows and short videos, especially when they'd rather not record their own voice or want a quick draft read before hiring a voice actor. Character voices for games and comedy clips are a big category too; our guide to funny text to speech covers that side.
Customer service
Phone menus, delivery updates and support bots use TTS so a business can change what the voice says by editing text, with no new recording session.
Browse voices, play the samples, type your text and pick a language and speed. Sign in to generate and download it with credits.
TTS vs voice cloning vs speech to text
These three get mixed up all the time. The quickest way to tell them apart is to look at what goes in and what comes out.
| Text to speech (TTS) | Voice cloning | Speech to text (STT) | |
|---|---|---|---|
| Input | Text | A short voice recording, then text | Spoken audio |
| Output | Speech in a preset voice | Speech in the recorded person's voice | Written text |
| Typical uses | Narration, read-aloud, alerts | Personal voiceovers, translated versions | Captions, dictation, meeting notes |
| Consent question | Covered by the tool's terms | Needs the speaker's permission | Needs consent to record people |
Voice cloning is really a branch of TTS: the same pipeline, plus a step that captures one person's voice and uses it in place of a stock voice. Our voice cloning tutorial walks through how that works in practice. Speech to text, also called automatic speech recognition, runs the other way and turns audio into words.
How to choose a TTS tool
The right tool depends on what you're making. Work through these questions before you commit to one:
- Does it speak your language well? Test a real paragraph from your project, not the demo sentence. Listen closely to numbers, names and mixed-language phrases.
- Does it have voices that fit your audience? A calm narrator for e-learning, a warm voice for a children's story, a character voice for a game. Play samples side by side, as you can in our AI voice generator.
- Can you control the delivery? Speed is the minimum. Control over pauses, emphasis and pronunciation saves a lot of rewriting. Some tools accept SSML tags; others offer simpler sliders and fields.
- What can you do with the audio? Check the download formats and whether the terms allow commercial use.
- How does it handle long text? Long scripts are usually generated in chunks, so check that the voice stays consistent from one part to the next.
- What happens after the free tier? Read the pricing and the license before you build a series around one voice.
A practical shortcut: paste the same 150-word script into two or three tools, generate it, and listen back without looking at which file is which. Your ears will decide faster than a feature list.
Where TTS still struggles, and how to work around it
Neural TTS is good enough that most people can't pick it out in a short clip. Longer listening still exposes a few weak spots.
Names, brands and unusual words
A model can only guess at words it hasn't seen, so brand names, surnames, place names and new slang go wrong more than anything else. Spell the word the way it sounds (write "Worcester" as "Wuss-ter"), use the tool's pronunciation feature if it has one, or check an unfamiliar English word in our pronunciation guide first. In SSML, the phoneme and sub elements exist for exactly this job.
Numbers, dates and abbreviations
Normalization makes its best guess, and the guess is sometimes wrong. If it matters, write it out: "ten thirty a.m." instead of "10:30", "two to three days" instead of "2-3 days."
Emotion and emphasis
TTS reads the words, not the subtext. Sarcasm, excitement or grief can come out flat, or wrong. Shorter sentences, deliberate punctuation and a voice whose natural delivery fits the mood get you much closer than fighting a mismatched voice.
Consistency over long scripts
Across a long audiobook or course, pacing and tone can drift between chunks. Generate in sections of similar length, keep the settings identical and listen carefully at the joins.
Browse narrators, storytellers and character voices, and listen to the samples before you choose.
