Exciting!
With Mandarin, have you solved whatever issue(s) was/were causing the tone mistakes?
FWIW, I think Welsh definitely needs some love to help bridge the gap between the end of level 3 and the easiest of the advanced content.
1 Like
We think so, but that’s part of why I want to be cautious about when we’ll be able to publish Mandarin! I hope very much it will be by the end of January, but time will tell…
With the Welsh, I’m keen to replace Level 3 entirely - I think we’ve got some much better stuff now - and then I’m hoping that our new approach to listening is going to help bridge exactly that gap to a more C1 level of comprehension 
2 Likes
Hi Aran
Is there any plan to incorporate AI conversation partners into the courses. I understand their limits (especially with lesser spoken languages) and some peoples feelings towards speaking to AI as ultimately the aim of learning a language is to speak with real People. But … After finishing the SSI French course, because I don’t know anyone who speaks French, I’ve used an AI language platform to practice speaking and it’s been a massive help. I managed about an hour a day but the advantage of AI is that I can chat whenever I have 4 mins throughout the day. Combined with listening to French podcasts I find this the best follow up to the SSI course for someone like me who isn’t living in the country where the language spoken.
1 Like
This is a tricky and interesting line of thought. We’ve looked at some of the models that are already available for Welsh, and we think that structured in the right way, there is definitely some opportunity to give people interactional practice when they don’t have access to speakers. It’s a training wheels kind of deal, I think - you’ll always benefit from moving to real conversations asap - but as a filler when that isn’t possible, or to help build confidence, there is clear potential.
Having said that, we want to finish the work we’re doing on realtime assessment and listening exercises (as well as revamping the design and launching the schools interface) and increasing our capacity to build in new languages before we take on integrating any AI functionality.
1 Like
Duolingo tried implementing AI speech recognition to train pronunciation, it led to complains like “my dad is native German speaker and the app marked his answer as incorrect”. False negatives are very annoying. By far I feel there’s some frustration in knowing that organic speaker would understand the same line AI’s speech recognition failed on, especially on unusual voices. But models are getting better. Gender/age diversity in training data can be weak point, talked to robot trained mostly on male voices, it could only understand me when I talked in unnaturally low voice.
1 Like
@elizaveta Yes, there are definitely bugs in the Duolingo AI voice recognition. There’s a word in German that it always flags my pronunciation as wrong, no matter how I say it, and I’m sure it’s pretty close.
Other times, I hesitate in the middle of a sentence and it gives me the “All perfect” response. I hadn’t even finished saying it!
Then I tried a Basque AI chat partner, to have more of a chance to speak a little Basque each day, but the microphone option just produced gibberish, forcing me to type what I wanted to say - NOT what I wanted to practise!
So I wouldn’t be rushing to try to implement something until it’s a little better, but great potential for the future.
4 Likes
Spanish on Duo has those issues too. From what I gathered, experiences vary widely, some voices are easily recognized by system, some people apparently have voices that ai has trouble with, probably not because of pronunciation per se. Which means it has potential to become access barrier if made a mandatory part of the process. I can imagine sci fi where voice minority is actual minority that has trouble accessing banks, education etc.
1 Like
There’s interesting branch of possibilities here. Models already can be pretty good at singing, and neurologically it’s often easier to start to sing than start to talk, especially if there are some speech disorders (that’s why people sing to babies and why special kids sometimes actually have first songs as their earliest speech). Theoretically, borg takeover could also transform say something into sing something at some point.
1 Like
One thing we certainly won’t be doing will be making voice interaction a mandatory part of the process 
Have you seen this video @elizaveta? It’s been around for a while but still amusing:
8 Likes
I was reading this conversation and just thinking of that. 
3 Likes
1 Like
God, this is hilarious
2 Likes
Is there any update on the first new versions (he said impatiently)?
2 Likes
Yes - but it’s a mixed bag - we’ve got them coming through quickly now to the new app - but it’s still a slow headache to get them released in the current app (where we have problems with getting the sync right).
I believe that we’re on the verge of releasing Italian and Portuguese in the current app, with French on its way soon after - but we’ve got about 8 released in the new app - if you’d like to play around with them, on the understanding that you’re in a live development app that may explode at any point on a regular, daily basis, drop me an email
3 Likes
Got to say, that’s really tempting. 
2 Likes