The project wants to build up a dataset to train neural networks for speech recognition software. The first goal for every language is collecting 12000 hours. They have reached this, but they only release a dataset twice a year, that's why you still see 1100 hours at this download.
Other languages are looking good too. Germany has almost reached 600 h, French almost 500 h and the website was localized to many other small and big languages in the last year.
The main factor to build a good dataset is A having a diverse group of donors and B having enough sentences. So if you want to help you can either donate your voice, validate other voices or collect more sentences to record. All three tasks are equally important to create a good dataset.
1 comment
[ 2.4 ms ] story [ 16.7 ms ] threadOther languages are looking good too. Germany has almost reached 600 h, French almost 500 h and the website was localized to many other small and big languages in the last year.
The main factor to build a good dataset is A having a diverse group of donors and B having enough sentences. So if you want to help you can either donate your voice, validate other voices or collect more sentences to record. All three tasks are equally important to create a good dataset.
https://voice.mozilla.org