Chapter 14 · Images, sound and video
Text to speech
TTS models turn text into sound by predicting audio features from tokens.
Chapter 14 · Images, sound and video
TTS models turn text into sound by predicting audio features from tokens.
Start with written words.
Spelling is a poor guide to sound, so the first job is always to work out which sounds a word really contains.
“Make” has four letters and three sounds. The first step turns spelling into phonemes, the basic sounds of speech, using learned rules and a dictionary.
Type something short. The timeline shows each sound, how long it lasts and the pitch line above it.
“make the boat go” becomes 10 sounds in 1.53 seconds. The solid blocks are the stressed sounds: longer, and higher in pitch.
A teaching toy: a small spelling table that will mispronounce unusual words, which is exactly why real systems use a dictionary plus a learned model. Reading aloud uses your browser's own voice.
Why can a model not simply read letters aloud?
What is prosody?
What does the vocoder do?