A text-to-speech voice mispronounces a word for one of two different reasons

The front end decides the words before the back end makes a sound

A classic text-to-speech system runs in two stages. A front end reads the sentence and works out, word by word, how it should sound. A back end turns that decision into audio. DeepMind's own account of the pipeline behind WaveNet describes the same split: a text-analysis stage that decides, and a synthesis stage that only works from what it decided.¹ By the time synthesis starts, the pronunciation is fixed. Nothing about how natural the voice sounds can undo a wrong call upstream.

NPR collected two real examples of that upstream call going wrong, both from Siri. The Alabama city "Mobile, Ala." came out with the vowel of the adjective instead of the place, and "Des Moines" came out with its silent letters sounded out loud.² To a listener, those sound like the same glitch. They come from different parts of the front end.

Numbers, abbreviations, and dates are not words yet

Not every written form in a sentence is already a word. Dr. could be "doctor" or "drive." St. could be "saint" or "street." 2010 could be "twenty ten" or "two thousand ten," depending on whether it names a year or a quantity. Before a system can decide how these should sound, it has to expand them into the words a person would actually say. That step is text normalization, and it is genuinely ambiguous: the right expansion depends on context the written form alone doesn't settle.

Google researchers measured just how ambiguous it is with a system trained on nothing but this problem: millions of examples of written text paired with its correctly spoken form.³ It got the mapping right the overwhelming majority of the time, 99.3% overall accuracy on the English test set. On the fraction it got wrong, the errors swapped in a different meaning.

That's not a mistake you're likely to hear from a shipping voice. Production systems pair the raw model with a rule-based filter, built to catch and block outputs like these before they ship.³ A string of digits or an abbreviation needs interpretation before it is a word at all, and with nothing there to check it, that interpretation can drift far.

A name with no dictionary entry gets sounded out from scratch

A system also needs to pronounce ordinary words it has never been told how to say: a name, a brand, a borrowed word. English spelling doesn't make this easy: "ough" sounds different in "though," "through," and "rough," with no general rule for which applies.

A pronunciation system starts from a dictionary: common words paired with their correct sound, broken into phonemes, the sound units a word is built from (the difference between "pat" and "bat" is a single phoneme). Google engineers break "google" into its phoneme string this way in a 2015 paper.⁴ A dictionary only covers words someone added. For everything else, grapheme-to-phoneme conversion, or G2P, predicts the phoneme string from the letters, generalizing from spelling patterns it has seen: "invoked anytime a word is not in the static dictionary."⁴ That's why an uncommon name is where G2P is most likely to guess wrong: usually there's no entry to fall back on, and the spelling doesn't reliably say how the word sounds.

How often does the guess fail? On dictionary words the system had never seen in training, a strong neural predictor got 25.8% wrong, roughly one in four, falling to 21.3% with a statistical backstop.⁴ That's on ordinary English words. NPR found the same gap in the world: Siri's "Des Moines," sounding out silent letters that place name doesn't pronounce.² There's no meaning to weigh, no pronunciation chosen badly. The dictionary is, as one researcher who helped build Siri put it, "the first port of call."² For "Des Moines," it wasn't there, and the right sound was never produced.

Choosing between two correct sounds needs the rest of the sentence

Some written forms have more than one correct spoken answer built in, and no dictionary entry or spelling rule can choose between them. "I will lead the team" takes a different vowel than "the pipe is made of lead." "She plays bass" is a fish or an instrument. "He read it yesterday" and "he will read it tomorrow" say the same four letters two ways, by tense. These are homographs: one spelling, two or more real pronunciations, decided by sense or part of speech in that sentence.⁵ A homograph seen in isolation cannot be resolved by spelling alone. Both pronunciations are equally valid answers to "how do you say this word." The only source of the right answer is meaning, and it shows up only once the rest of the sentence is read.

"Mobile" is the opposite case. Siri has this everyday word in its dictionary, so unlike "Des Moines" the right sound was there. The mistake is that "mobile" has two correct spoken forms, one for the adjective and one for the Alabama city, and choosing between them needs the word right after it, "Ala." Siri picked the adjective's sound for a city name.²

Homograph disambiguation is a studied, named problem with its own accuracy figures. Researchers at Google built a labeled benchmark of 163 English homographs, drawn from roughly a hundred Wikipedia sentences apiece, and measured how well different approaches chose the right pronunciation, from ignoring the sentence to combining rules with machine learning.⁵ A separate team at Amazon reached 99.1% on the same benchmark using contextual language-model features and no hand-written rules, confirming the ceiling from a different direction.⁶

The benchmark measures a known homograph already located in the sentence, not how often one turns up unflagged in ordinary speech, or how these systems do outside the labeled list. Even inside the curated set, some cases resist every method: in one sentence from Google's error analysis, "Smith has played Trophy matches for the county from 1993 to present," every system tested still chose the verb's pronunciation over the correct noun.⁵

What neural TTS fixed, and the sentence no system can resolve

Everything so far describes the classic pipeline: normalization, then G2P, then homograph disambiguation, each finished before a synthesizer runs. Modern neural systems don't keep the boxes so clean. Google's Tacotron 2, published in 2018, sounds close to a human reader and maps a sentence's letters almost straight to audio in one network, folding much of the front end into a single model trained end to end.⁷ It is still trained on already-normalized text: the researchers write "16" as "sixteen" before the model ever sees it, so normalization has just moved outside the network.⁷

Tacotron 2 still mispronounces. On a 100-sentence stress test, it got 6 sentences wrong and skipped a word on 1. The researchers call occasional mispronunciation the single biggest reason listeners preferred the human recording.⁷ On out-of-domain headlines, names are the recurring trouble spot, the same shape of failure as G2P: a model trained mostly on ordinary vocabulary has, in effect, no dictionary entry for a name it rarely saw. The same test turned up a third kind of error: 23 of the 100 sentences had unnatural prosody, the wrong syllable stressed, an odd rhythm. That isn't a word-choice error. It's a failure in what used to be the back end: the model misjudging pace and stress rather than choosing the wrong word, in an architecture where front and back aren't separate boxes anymore.⁷ The claim that a mispronunciation is almost always a front-end mistake holds for the classic pipeline. It holds less cleanly once normalization, G2P, and synthesis train together.

Vendors patch what's left, word by word. Amazon's Polly lets a developer upload a pronunciation lexicon mapping a written form to a replacement spelling or phoneme string, applied before synthesis starts: "W3C" becomes "World Wide Web Consortium."⁸ It fixes the word someone added, not the mechanism that got it wrong.

Under all of it sits a floor no version of this pipeline gets past. "I read it" cannot be resolved from the sentence alone. Whether it's present or past depends on what came before it, and a system that sees only this sentence knows exactly as much as someone who walks in partway through and hears only that line: nothing. That's the point where a bigger dictionary, more data, or more rules stop helping. The sentence itself doesn't contain the answer.