What Is Text to Speech? How AI Voices Work in 2026
What is text to speech? TTS turns written text into spoken audio. Learn how AI voices work, how they differ from old robot voices, and where they are used.
Text to speech (TTS) is software that converts written text into spoken audio. Microsoft defines it as enabling "applications, tools, or devices to convert text into human like synthesized speech." Modern AI voices use neural networks, which is why they sound far more natural than the robotic voices of earlier decades.
What is text to speech?
Text to speech, usually shortened to TTS, is also called speech synthesis. The term describes any system that takes written words and produces a spoken waveform. In web browsers, according to MDN, the entry point is window.speechSynthesis, and the voices come from the operating system's speech system. The same web page can therefore sound different on two devices.
How does text to speech work?
Most TTS systems run in two broad stages. First the system analyzes the text, then it generates the audio. The analysis stage handles three problems that are simple for people and hard for software:
- Text normalization: Numbers, abbreviations and symbols must become spoken words. "1325" can be "thirteen twenty-five", and "St." can mean Saint or Street.
- Pronunciation: Letters must become sounds. Dictionary lookup is quick and accurate but fails on unknown words. Rules handle any input but get complicated, so nearly all systems combine both.
- Prosody: The system decides pitch, duration and stress across a phrase. Prosody remains a challenge for speech synthesizers, and statements, questions and exclamations need different pitch contours.
The generation stage turns that plan into sound. Older systems joined recorded fragments, while neural systems predict the audio directly. How neural text to speech works explains that pipeline step by step.
A short history of text to speech
Speech synthesis is older than most people expect. The first rows below come from Wikipedia's history section. The WaveNet and Tacotron 2 rows come from the primary sources.
| Year | Milestone | |---|---| | 1939 | Homer Dudley of Bell Labs exhibits the Voder, a keyboard-operated voice synthesizer, at the New York World's Fair | | 1961 | John Larry Kelly Jr. and Louis Gerstman at Bell Labs synthesize speech on an IBM 7090 | | 1968 | Noriko Umeda and colleagues build the first general English text-to-speech system at Japan's Electrotechnical Laboratory | | 1980s–1990s | DECtalk, largely Dennis Klatt's work at MIT, and Bell Labs systems are dominant | | Sep 2016 | DeepMind publishes WaveNet, which models raw audio one sample at a time | | Dec 2017 | The Tacotron 2 paper first appears on arXiv |
Old TTS vs AI voices: what changed
Early commercial systems mostly used concatenative or parametric methods. Concatenative TTS relies on "a very large database of short speech fragments" recorded from one speaker, so changing the voice or the emphasis means recording a new database. Parametric systems generate speech from modeled parameters and signal processing. Google still describes its Standard voices as parametric.
Neural voices learn from recordings instead. Google says its WaveNet voices were "trained using raw audio samples of actual humans speaking", which gives them more natural emphasis and inflection.
DeepMind's 2016 listening test, scored on a 1 to 5 mean opinion score (MOS) scale, shows the gap:
| Approach | US English | Mandarin | |---|---|---| | Concatenative | 3.86 | 3.47 | | Parametric | 3.67 | 3.79 | | WaveNet | 4.21 | 4.08 | | Human speech | 4.55 | 4.21 |
These numbers come from one 2016 test. DeepMind says WaveNets "reduce the gap between the state of the art and human-level performance by over 50%." Scores from different papers use different test sets, raters and years, so do not compare them across studies.
Where is text to speech used?
- Built-in features: Browsers and operating systems expose speech synthesis to apps, using the device's own voices.
- Accessibility: Read-aloud and screen reader tools help people who cannot easily read on-screen text.
- Narration and media: Google positions its Studio voices for media narration.
- Voice agents: Google positions its Chirp 3: HD voices for conversational agents.
- Developer APIs: Azure, Amazon Polly, Google Cloud, the Gemini API and OpenAI all sell TTS through an API.
Can AI voices sound human?
In controlled tests, sometimes. Tacotron 2 reported a MOS of 4.53, compared with 4.58 for professionally recorded speech. That result comes from the authors' own listening test, so treat it as a research claim rather than an independent benchmark. Microsoft says its HD voices "automatically detect emotions in input text, and adjust speaking tone in real-time." That is also a vendor claim.
Gaps remain. In DeepMind's 2016 chart, human speech scored 4.55 against 4.21 for WaveNet, and prosody across a whole passage is still an open research problem. Script wording affects how well a voice can carry that prosody, which is why the tips for natural-sounding AI voices focus on how the text is written.
What to check before you pick a TTS tool
| Check | Why it matters | Example | |---|---|---| | Languages and voices | Coverage varies by provider | Azure lists 100+ languages and locales for its standard neural voices | | Pricing unit | Characters, credits and tokens are not comparable | Azure counts punctuation, spaces and markup; Amazon Polly charges per character of text; Gemini quotes tokens | | Commercial rights | Free plans may exclude commercial use | ElevenLabs lists a commercial license from Starter upward, not on its Free plan | | Disclosure | Some providers set rules for it | OpenAI's usage policies require clear disclosure that a TTS voice is AI-generated |
Free options exist too. Browser voices cost nothing, Amazon Polly's free tier includes 5 million Standard characters per month, and Azure's F0 tier includes 0.5 million Neural characters per month. Check whether a free plan allows commercial use before you publish anything.
Try text to speech with OndaLabs
OndaLabs, the new name of NeoSpeech, is an AI voice studio. It offers text to speech, two-speaker dialogue, speech to text, custom voices, image generation and music generation. It has 30 prebuilt voices usable in every supported language, with dedicated pages for 37 languages. You control delivery by describing it in plain words and with supported expressive tags in the script.
The Free plan costs $0 and includes 2,000 credits a month (about 2 minutes), Onda Lite and expressive tags. Onda Lite costs 1 credit per character. Onda Flash, the most expressive model, is on paid plans at 1.5 credits per character. Commercial use starts on the Starter plan at $9 a month. To hear the difference yourself, open the text to speech studio and generate a short passage.
Frequently asked questions
What does TTS stand for?
TTS stands for text to speech. The term is used as a synonym for speech synthesis, which means generating spoken audio with software. Some product pages write "text-to-speech" with a hyphen, but the meaning is the same.
What is the difference between text to speech and speech to text?
Text to speech converts written text into spoken audio. Speech to text, also called transcription, does the reverse and converts recorded speech into written words. Voice apps often combine the two, for example by transcribing a call and then reading a summary back.
Is AI text to speech free?
Some options are free. Browser voices through the Web Speech API cost nothing and need no account, although the voices depend on the device. Amazon Polly and Azure also have free tiers with monthly allowances. Check the limits before you rely on one for a project.
Can I use text to speech audio commercially?
It depends on the plan and the provider. ElevenLabs lists a commercial license from Starter upward, not on its Free plan. OndaLabs includes commercial use on Starter and above. Read the terms of the plan you use before you publish audio for a client or a product.
Can text to speech sound human?
Some systems score close to human recordings in controlled listening tests. Tacotron 2 reported 4.53 against 4.58 for professional recordings, although that was the authors' own test. Prosody across long passages is still an open problem, so listen to your output before you publish it.
Do I have to say that a voice is AI-generated?
OpenAI's usage policies require a clear disclosure to end users that a TTS voice is AI-generated, not human. Other providers set their own rules, so check the terms of any service you use. Disclosure requirements change, so read the current policy before you launch.