TTS front-end
Everything that happens to text before a text-to-speech model sees it: expanding numbers, dates, abbreviations and currency into words, deciding how ambiguous spellings are pronounced, converting graphemes to phonemes, and — in languages that need it — marking stress and prosody. The acoustic model gets the credit; this stage decides what it is being asked to say.
It is the TTS-side mirror of the ASR pipeline this wiki also lacks (diarization, voice activity detection, forced alignment — growth edge 5). Both are the plumbing around a model, and this corpus grades models.
Why it is invisible in English and unavoidable in Russian
English TTS can get a long way on characters alone, which is why the models on this wiki mostly
take raw text and say nothing about a front-end. The failure modes there are ordinary — Dr. read
as doctor or drive, 1/2 as January second or one half — and a large enough model absorbs
most of them from context.
Russian does not offer that. Stress is not written, it moves between forms of the same word, and it
distinguishes words outright: з+амок (castle) against зам+ок (lock), мук+а (flour) against
м+ука (torment). Vowel reduction then depends on the stress position, so getting it wrong changes
every vowel in the word, not just one syllable's emphasis. A synthesizer with no stress input
either guesses or is handed pre-marked text — and the marking is a separate NLP problem with its
own models. ruaccent is that problem solved as a library: five ONNX models, four dictionaries,
and a morphological analyzer, to produce text a synthesizer can read.
The same shape recurs elsewhere: Hebrew and Arabic diacritization, Japanese pitch accent and reading selection for kanji, Mandarin tone sandhi and the polyphone problem. In each case a language-specific classifier stands between the text and the acoustic model, and none of it is covered by any TTS benchmark this wiki holds.
What it does to the numbers
Two consequences.
Comparability. An Elo or MOS score on tts-arena-leaderboard measures a model plus whatever front-end the harness happened to feed it. For English that is close to nothing and the comparison is fair; for a language where the front-end is doing real work, two identical acoustic models with different text pipelines are two different products. None of the boards here state what the front-end was.
Where the error budget goes. chime-6 made the equivalent point on the recognition side — swapping the enhancement front-end moved the result by 10 points absolute with the models held fixed, larger than most model-to-model gaps this wiki reports. There is no published equivalent for the TTS front-end, and there is no reason to expect its share of the error to be small in a language where a single wrong stress mark makes the wrong word.
Related
text-to-speech · ruaccent · tts-benchmarks · tts-arena-leaderboard · chime-6 · speech-audio-ai · open-weight-tts