TL;DR:
Text to speech converts written words into sound in stages: clean the text, decide how to say it, then produce audio.
Rhythm, stress and pitch (prosody) are what separate a natural voice from a flat one.
An assistant can speak the first sentence while later ones are still being written, which makes answers feel immediate.
You ask a question and a voice starts answering before you expected it to. The words were written by software a fraction of a second earlier, and now they are sound coming out of your speakers. The step that turns writing into speech is called text to speech, and it is more than reading letters aloud.
The short answer
Text to speech (TTS) takes a string of written text and produces an audio waveform of someone saying it. It has to solve two separate problems. First, what are the sounds? Second, how should they be delivered: how fast, with what emphasis, rising or falling at the end?
The first problem is mostly a dictionary and rules. The second is where a voice becomes pleasant or irritating, and it is the harder one.
What happens step by step
Normalise the text. Turn "3 kg" into "three kilograms", "Dr." into "doctor", and "2026" into "twenty twenty-six".
Work out pronunciation. Map each word to its sounds. English spelling is unreliable, so this uses a pronunciation dictionary plus a model for unknown words.
Plan the delivery. Decide where to pause, which words to stress, and whether a sentence should end up (a question) or down (a statement).
Generate the sound. A trained model produces the audio waveform that matches the sounds and the planned delivery.
Play it. The audio is sent to the speakers, often in small chunks as soon as each is ready.
Walk-through: a reply being spoken while it is written
Suppose you ask how to add up a column and the assistant's written answer is three sentences long. Writing the whole answer takes a couple of seconds. If the voice waited for the final full stop before starting, you would sit in silence for that long.
Instead, the system treats the answer as a stream. The moment the first sentence is finished ("Click the cell below your numbers."), that sentence goes through the steps above and starts playing. While you hear it, sentence two is being written, and then spoken as soon as sentence one ends. The listener only notices that the answer began quickly.
This works because speech is slow compared with writing. People speak somewhere around 150 words per minute, while text can be produced much faster than that, so the writing stays ahead of the voice and the audio rarely runs dry. The catch is that the assistant cannot go back and change a sentence you have already heard.
What makes a voice sound natural
Feature | What it is | What goes wrong without it |
|---|---|---|
Stress | Which word is emphasised | "I never said she took it" changes meaning with the stressed word |
Rhythm | Pauses and speed | Machine-gun delivery with no breathing room |
Pitch | Rising and falling tone | Questions that sound like statements |
Numbers and names | Reading them correctly | "1/2" read as "one slash two" or a name mispronounced |
These features together are called prosody. Most of what people call a robotic voice is missing prosody, not wrong sounds.
What it cannot do
Know what you meant. It speaks the words it is given. If the text is wrong, the speech is confidently wrong.
Pronounce everything. Unusual names, acronyms and mixed-language words can come out badly.
Show structure well. A table or a list read aloud can be hard to follow, which is why spoken answers should be short and plain.
Let you skim. Audio is linear. You cannot glance ahead as you can with text.
Tip: If an answer involves numbers, steps or code, an assistant that can also draw or write on screen is better than speech alone. Voice is good for explanation, the screen for detail.
Privacy and control
Speaking out loud has a social side: other people in the room hear it. Check that you can lower the volume, interrupt, or turn off spoken answers when you are somewhere quiet. Also check whether the text of your question is processed on your device or elsewhere, as products differ and the answer should be in their privacy page.
For DuckyHelper, the full privacy picture is on the privacy page.
Try it yourself
Find the "read aloud" or "speak" feature on your computer and give it a short paragraph that includes a date, a unit and an abbreviation, for example "On 3 March, Dr. Lee measured 5 kg." Listen for how it handles each. Then add a question mark to a statement and hear what changes.
To hear a spoken answer arrive while you keep working, download DuckyHelper. Press fn on a Mac, or Ctrl+Shift+Space on Windows and Linux, and ask. You can try it for 10 minutes, no card.
Frequently asked questions
Why does the voice sometimes mispronounce a name?
The system guesses from spelling, and names break the usual rules. Many tools let you spell it out or give a hint.
Does a better voice mean a smarter assistant?
No. The voice only delivers words. How good the answer is depends on the writing step, not the speaking step.
Why speak sentence by sentence?
It lets the answer begin sooner. Waiting for the whole text would add a silent pause before you hear anything.
Get started
