https://www.duolingo.com/sifon10

Why does duolingo pronounce "girl" as G-l when I learn English from either portuguese or Chinese?

December 27, 2017

5 Comments


https://www.duolingo.com/ycUvuSap

I was always taught to pronounce "girl" as "gööl", with the slightest hint of "r" before the "l". What do you mean by "G-l"?

December 27, 2017

https://www.duolingo.com/Peter594672

The voicing robot does fail to get it right at times. Some words for some reason are harder for it than the others. "Girl" pronounced as "gee earl" is one. The other two I come across a lot are "here" pronounced as "a chair" or "H-air", and "where" pronounced as "double you a chair".

December 27, 2017

https://www.duolingo.com/anomalousjack

It does it all the time, and becomes even more noticeable when you are doing the reverse trees and the machine is speaking in your own language. In fact, it's pretty useless when you think about it, so much so that perhaps it might be possible to pick out someone who has learned your language from Duolingo by their weird pronunciation?

December 27, 2017

https://www.duolingo.com/Oritteropo1

It happens in English from Spanish too.

I would guess that there are some stray characters either in the words that are sent to the TTS, or in the TTS word to phoneme conversion. It's pretty consistent with spaces in the input between the g and the irl and between the w and h in where.

It's interesting that it happens consistently between English courses. It does suggest shared data.

December 27, 2017

https://www.duolingo.com/JasonMey

Duolingo uses a text-to-speech program for most courses. As a result, the quality can be really hit-and-miss. It offers a number of advantages when it comes to producing the courses, and means that no phrases need to lack audio, but the quality of each takes a hit.

This is a big reason why: 1. I use multiple resources, not just Duolingo, for learning vocab and 2. I value the effort put into courses like Esperanto, which have actual recordings from a speaker instead of computer voices, even if half of the phrases don't have any audio. Better to get a good model on 50 phrases than a bad model on 100.

December 28, 2017
Learn a language in just 5 minutes a day. For free.