Is it possible to implement across-the-board audio fixes for the computer-generated readings? Like, 君 is very, very often mispronounced as くん instead of きみ. There are 1-2 more I’m not thinking of off the top of my head, but it would be better to implement a sort of mass update on these instead of going one-by-one for each sample sentence.
4 Likes
I submitted bug fixes for this a while back! They said they’re aware and working on fixing the TTS language model. 
Not sure when they’re pushing that update though.
I think the funniest one I saw was 先輩 being spoken in Chinese. Everything else in the sentence was spoken in Japanese…
1 Like
I tried to send the お巡り to the AI to fix it, but they kept getting sent in circles (巡り) 
2 Likes
If you report the audio as you find them, the Content team will add it to their list of stuff to fix.
If you want to include it as a list or something, if you DM me I can pass it on to them!
1 Like
Okay this took me a couple reads before I face planted on my desk. Curse you and your silly dad puns!!
1 Like
Ohhh is this okay? I submitted a bunch of requests a while back and I thought based on the response I got (we are working on the TTS bot) that flagging individual instances wouldn’t be helpful for y’all and was bothersome.
Since then I’ve just been ignoring the audio stuff when it comes up.
1 Like
I’m discussing with one of the content team members now.
Unfortunately the Report button is still the best way to go at the moment.
The reason being is we get the exact ID that is attached to the sentence, so we can track it better internally.
Gonna try get the priority of these fixes bumped internally though, so your post was not in vain!
2 Likes
Thanks, so I guess then someone still needs to identify individual sentences for correction? Sometimes I feel like I do it so many times I should be asking for a commission 
Are there general rollouts that increase accuracy or provide improvements? The audio is way better than like a year ago, it’s just that there are a few remaining points to address
4 Likes
Turns out we have been collecting and collating the reports for a while now, but we’re only now going through and actually fixing them.
So after this mass-fix, there should (hopefully) be a lot less to report.
Also, once we’ve done this fix, we’ll stay on top of the reports a lot better too.
AFAIK, there won’t be any mass replacements of the existing TTS audio, as I think only some of them are broken, and we’re using the latest/best TTS technology (Azure).
If you know of any technologies that you think would be better/more accurate than our current ones, we’d love to hear about them though!
5 Likes
This is extremely nitpicky but I’d like to ask, is it okay if I submit reports for when the errr, pacing of the TTS audio is weird? I’m not really sure how to describe it but sometimes the audio just clumps together the sounds in the slightly unnatural way.
For example in the 熱 がある examples: this sentence has weird pacing and tonal qualities. I think there should be slightly more of a gap in speech between は and 今 for:
私は今熱があります
I’ve left it alone because I assumed there’s nothing to be done about it as it might just be how the TTS works, but it seems like the Bunpro team might be more interested in detailed bug reports on the TTS than I first assumed?
1 Like
I notice that too sometimes, it’s usual alright but sometimes I can’t really use a sentence for audio practice since I can hear that the intonation is not quite right.
Should we also flag these as something that can be addressed?
1 Like
We add notes next to each report about what the issue in the audio is, so yeah I think just flag it as unnatural.
I think the one you mentioned is unnatural enough to warrant a re-generation.
1 Like
I don’t know if it’s because I’m mostly studying higher level grammar and vocab (n1 and A/B unclassified etc ) but I would guesstimate that between 1/3 ~ 1/2 of the sentences I study have errors in. For individual vocab (that doesn’t have sentences) it’s about the same. I’m afraid that there are so many errors it would take me hours to flag them all so I gave up quite a while ago. Instead I just wait for it to say the wrong word and I shout the correct pronunciation loudly over it, which is kind of a fix 
I kind of assumed that the team would be working to fix audio from the bottom up so the more advanced vocab etc won’t be fixed for a while, is that correct?
As a general note, I would say most individual vocab words that have 的 on the end are almost always pronounced まと and any word that has 々 in it produces a fascinating variety of nonsense!