Is anyone else mystified by some of the pronunciations that text to speech comes up with in the japanese version of skritter?

An example is 盆暗. Skritter correctly shows the kana as ぼんくら, but text to speech comes up with ぼんあん. The 読み is correct, but these 読み aren't right for this word.

My Japanese wife is bemused by some of the pronunciations that text to speech comes up with.

I think that some potential solutions to this could be (in order of preference):

- record native speakers so that fewer words default to text to speech
- turn text to speech off, or allow an option for it to be turned off
- go through the entire database and check that the text to speech is using the correct 読み

I really hope that skritter can sort this text to speech issue out. It makes me start to think that this is maybe not the best application for learning Japanese. I also find it difficult to recommend skritter to other japanese learners, which I used to do enthusiastically.

I am using the latest version of skritter, as downloaded from the App Store yesterday.

Some pronunciations that were just plain wrong from 20 minutes of practice today were:

染み染み
快気祝い
遺伝子組み換え
全粒粉

The latter caused hilarity from my wife. It was pronounced as ぜんつぶこう which is some weird mixture of 音読み and 訓読み

This surely can't be the standard of product that a commercial organisation wants to be associated with.

I brought this issue up last year (in an e-mail, not a blog), and it was identified as being due to errors in an automated speech-recognition program (which would be quite complex, considering how on- and kun-yomi are not used in a consistent manner). This was to have been removed in an upgrade, but is still there.

(Aside: the words you list aren't in the lists I have. I would like to expand my lists, especially in the realm of science. What list did you get them from?).

This is being fixed! The TTS (text to speech) mispronunciations are due to the TTS engine itself. We were about to disable this entirely due to how many errors there are, and only use the human recorded audio we have, but then realized a solution. It will inevitably take a bit of time to curate this, though.

We've switched the way the TTS works-- it was initially reading from the kana before, and not the word itself, but was mispronouncing diphthongs like しゃ or しょ, and was breaking them apart as しや or しよ. We've switched it to read from the word itself, but it's not so good all the time and mispronounces some words, especially words that have an alternate reading, less common words, or compound words that get voiced.

The solution is to blacklist any words being mispronounced by the TTS, and have them treated differently. It would be extremely helpful if everyone could send in a feedback report for any more words they spot the TTS mispronouncing. If you just say "TTS" in the report, we'll know what to do!

In the scenario where it's possible to switch a word to read from the reading that doesn't have a diphthong that was being mispronounced, we will revert it that direction, opposed to disabling the TTS for the word.

It will take time, but we think this is a good solution as recording audio for every word is not feasible (due to the file size), and that disabling the TTS, the only other option, is not as good as manually ensuring each word is being pronounced correctly.

If only the TTS engine didn't mispronounce diphthongs when grabbing the reading from kana, this would be a non issue!

After the next update of the app, all the words that have been found/reported so far as being mispronounced will be fixed, however there will likely be more words lingering around in the next update until those too can be caught and fixed. Each update will become more revised, and eventually they will all be found!