Listen to those Indonesian words. The sound is very clear and native speakers agree with me it's way better than the voice of English Google Translator. I doubt Indonesian speaker uses more advanced technology than English, so what's the difference?
2 Answers
IMO the English is better than the Indonesian, though both are 10 times better than the Afrikaans which is 100 times better than the Swahili (which is infinitely better than the Zulu). Icelandic is ignominiously the same as Swahili.
Simply asking Indonesian speakers to judge the quality of synthesis for English vs. Indonesian is, experimentally speaking, not valid evidence regarding the quality of synthesis across languages. It's not at all clear what would be a valid procedure to test comparative quality, since there are so many variables to control (subject knowledge of the details of language A vs. language B; subject exposure to other synthesis of a language such as comparing original Klatt synthesis of English with Google synthesis). It might be best to use people with absolutely no exposure to the test languages (maybe impossible in the case of English), and play Real-X:Synth-X::Real-Y:Synth-Y sets (where "real" and "synthesized" are identified), which could get at subjective judgments of similarity. Even so, there would have to be good controls over the native languages of the subject (so that Germanic speakers don't overwhelm the subject pool, or Chinese speakers).
We don't know exactly how Google does its synthesis, but it is likely that for some of the worse syntheses, they are using a text-to-IPA type filter with some kind of knowledge (or misunderstanding) of surface phonetics in the language fed into Klatt synthesis. Whereas in the better languages, they probably rely on a large corpus of sampled speech (which is the basis for some kind of diphone resynthesis). The combinatorics of vowel-consonant-vowel configurations are more complicated and probably speaker-variable in English compared to Indonesian, so it's easier to acquire an ideal corpus of Indonesian that generalizes to the ability to handle any arbitrary input (not just actual words).
Given that hundreds of languages are supported, at any given point in time the voices for different languages are from different versions of the system, different datasets, or even completely different libraries from different projects or vendors.
You happen to have found a language for which the quality is better than for English, and for some reason users are more often surprised by a quality difference that way around than the other way around.
answered Jul 15, 2017 at 17:04
