Japanese text to speech needs a real front-end, not a bigger model

Two problems before the acoustics

Japanese text has no spaces, so the first job is deciding where the words are; and most kanji have several readings chosen by context, so the second is deciding what they say. Both are dictionary and parser work. A model that skips them and maps characters straight to sound produces fluent, confident, wrong speech - which is worse than a foreign accent, because a listener cannot tell it is wrong.

Counters make this concrete. The same numeral is read differently depending on what is being counted, and a name written in kanji can have a reading its owner chose. No amount of acoustic quality recovers from getting either wrong.

What we measured

Transcription error on probe sentences, per candidate route: synthesise the sentence, recognise the result, compare against the input. It is a blunt instrument - a recogniser has its own failures - but it catches exactly the class of fault we cared about, because a wrong reading comes back as different characters rather than as a worse score.

The difference was not subtle, and it was concentrated where you would expect: proper nouns, counters and homographs. The route with a native Japanese grapheme-to-phoneme front-end won on all three, so that is the one that ships.

Why refuse the other languages

The same engine claims several other languages. We tested them and did not ship them: a model that claims six languages and is good at one is a support burden dressed as a feature. Every language in the catalogue has to be one we would defend in front of a native speaker, which is why the shop lists what it lists and not everything the weights allow.