How does text-to-speech work, and what is neural TTS?

The three stages of a TTS system

1. Text normalization. The system rewrites text into words that can be spoken. "Dr. Lee paid $25 on 3/4" must become something like "Doctor Lee paid twenty-five dollars on March fourth" (or "the third of April", depending on the locale). Many mispronunciations happen here, not in the voice.
2. Linguistic analysis. Words are turned into phonemes, the sound units of a language, and the system predicts stress, phrasing and intonation. Words spelled the same but said differently, such as "read" in present and past tense, or Chinese characters with more than one reading, need context to resolve.
3. Waveform generation. The sounds and their timing are turned into audio samples. Older systems glued together recorded fragments (concatenative synthesis) or used a formant model, which sounds robotic. Neural systems predict the audio with a network.

A classic neural design, Tacotron 2, showed the modern split clearly: one network maps text to a mel spectrogram (a time-frequency picture of speech), and a second network, the vocoder, turns that picture into a waveform. In its authors' listening tests, it scored close to professional recordings. Newer systems often merge the stages or generate compressed audio tokens with a language-model style network, which lets them copy a speaker's voice from a short sample.

Neural TTS vs the system voices in your browser

Web pages can speak through the Web Speech API without any model of their own. The speechSynthesis object reads a SpeechSynthesisUtterance aloud with one of the voices the platform provides.

Each voice reports a localService flag. The specification says it is true for a local synthesizer and false for a remote service; a remote voice may need the network and may add delay.

const u = new SpeechSynthesisUtterance("Your order has shipped.");
u.lang = "en-GB";
u.rate = 0.9;   // 1 is normal; spec allows 0.1 to 10
u.pitch = 1.0;  // 0 to 2, 1 is the voice default
u.volume = 1.0; // 0 to 1
const voices = speechSynthesis.getVoices();
u.voice = voices.find(v => v.lang === "en-GB") || null;
speechSynthesis.speak(u);

Voices often load after the page starts, so getVoices() can return an empty list at first; listen for the voiceschanged event. For example, cv.cm's text-to-speech tool uses exactly these built-in voices with rate and pitch controls. It is a quick way to listen to a draft, not a neural voice generator, and the voice list depends on your device.

SSML: controlling pauses and pronunciation

Speech Synthesis Markup Language (SSML), a W3C standard, is XML that tells a TTS engine how to read text. Most cloud TTS services accept it. The Web Speech specification allows SSML in an utterance, but engines that do not support some tags must strip them and read the remaining text, so browser support is inconsistent.

<speak>
  Your code is <say-as interpret-as="characters">A7K</say-as>.
  <break time="500ms"/>
  Please call <sub alias="World Health Organization">WHO</sub>
  before <prosody rate="90%" pitch="+2st">Friday at noon</prosody>.
  The word is <phoneme alphabet="ipa" ph="təˈmɑːtoʊ">tomato</phoneme>,
  and it is <emphasis level="strong">not</emphasis> optional.
</speak>

If a service ignores SSML, you can still steer it with plain text: commas and full stops for pauses, spelled-out numbers, and phonetic respellings of hard names.

Writing text that sounds good when spoken

• Use short sentences. Long clauses make the voice run out of natural breath points.
• Write numbers the way they should be said when it matters: "two thousand twenty-six" or "twenty twenty-six".
• Expand abbreviations and units the listener needs to hear: "km/h" may be read letter by letter.
• Set the correct language and region. An en-US voice reading a Spanish name, or a zh-CN voice reading English, will use the wrong sound system.
• Listen to the whole result once before publishing. Check names, numbers and words with two pronunciations.

Voice cloning: consent and law

Modern models can imitate a voice from a short recording. That is useful for dubbing your own videos or giving a voice back to someone who lost it, and it is also used for fraud, such as calls that imitate a relative or a company director. Treat any voice as personal data that belongs to the speaker.

Reputable cloning services ask the speaker to record a consent statement before a voice is created. If a tool lets you clone any voice from any clip without such a check, do not use it for anything you publish.