Technology
How do AI voices work?
Satnavs, screen readers and AI assistants all read text aloud. Within a few years, synthetic voices went from robotic to hard to tell from a recording.
In short. Text-to-speech first works out what to say, expanding numbers and abbreviations and choosing pronunciations, then predicts the sound as a spectrogram, a picture of frequencies over time, and finally turns that picture into audio with a vocoder. Older systems stitched recorded snippets together or generated buzzy speech from a model. Neural systems changed that: in 2017, listeners rated one at 4.53 out of 5 for naturalness, close to the 4.58 of professional recordings. Natural voices raise new questions about consent and disclosure.
From text to sound, in steps
- Work out what to say. “Dr. Smith, 10:30” has to become “Doctor Smith, ten thirty”, and every word needs a pronunciation.
- Predict the sound. A model turns the sounds and the rhythm of the sentence into a spectrogram, the same kind of picture that speech recognition starts from, in reverse.
- Make the waveform. A vocoder turns the spectrogram into the thousands of numbers a second that a speaker plays.
The older ways
For decades, synthetic voices came in two kinds. Concatenative systems cut recordings of a real speaker into small pieces and stitched them together, which could sound natural but bumpy at the joins. Parametric systems generated speech from a statistical model, smoothly but with a buzzy, muffled quality.
The neural turn
In 2016, WaveNet generated raw audio one sample at a time, each sample predicted from all the ones before it, tens of thousands of samples per second of audio. Listeners rated its speech significantly more natural than the best parametric and concatenative systems, and one model could speak with many different voices.
A year later, Tacotron 2 combined a network that turns text into a spectrogram with a WaveNet-style vocoder. In listening tests, people scored its speech 4.53 on a five-point scale for naturalness, against 4.58 for professionally recorded speech. That score, the mean opinion score, is simply the average rating listeners give.
What still gives it away
- Names and rare words, especially from other languages.
- Numbers, dates and units read the wrong way: is 1/2 a half or the first of February?
- Emphasis and emotion, which depend on meaning the system may not grasp.
- Long passages, where a slightly wrong rhythm becomes tiring.
Voices, consent and disclosure
A voice that sounds human raises two questions. Whose voice is it? A voice built from someone’s recordings needs that person’s consent. And do you know you are talking to an AI? In the EU, the AI Act brings transparency duties for AI that talks with people and for content it generates: the EU AI Act and AI assistants.
How Lychnos Ring uses it
Lychnos Ring answers with its voice: the first line appears on its see-through screen, and the rest is read out by a synthetic voice, from its small speaker when you raise it to your ear or in earbuds that support Bluetooth LE Audio. It is in development: the AI ring.
Sources, read 4 November 2026: van den Oord and colleagues, WaveNet: a generative model for raw audio, 2016; Shen and colleagues, natural TTS synthesis by conditioning WaveNet on mel spectrogram predictions, 2017 (both arXiv). General information.
Questions
How does text-to-speech work?
It works in steps: the text is cleaned up and turned into sounds to say, a model predicts what those sounds should look like as a picture of frequencies over time, and a second model, the vocoder, turns that picture into an audio waveform.
Can listeners tell AI voices from real ones?
Often not easily. In a 2017 study, listeners rated a neural system’s speech 4.53 out of 5 for naturalness, against 4.58 for professionally recorded speech. Names, numbers, emphasis and emotion still trip systems up.