How Machines Speak
Inside the TTS Pipeline
Text is a script, not a recording
Raw text is a terrible input for speech. 'I got home at 5:30 on 3/14 and paid $1,204' must first be normalized into what a human would actually say — 'I got home at five thirty on March fourteenth and paid twelve hundred and four dollars'. That frontend — numbers, abbreviations, dates, currency — is unglamorous rule-and-model machinery, and it's where most production TTS bugs live.
Then phonemization: letters to sounds. 'Read' is 'reed' or 'red' depending on tense the system must infer; 'Live' rhymes with 'give' or 'five' depending on part of speech. The frontend decides pronunciation before the model ever makes a waveform.
insight
When a TTS system says something bizarre, the acoustic model usually pronounced faithfully what the frontend handed it. Debug text-in before you blame the neural net.