Us, too.

I think one of the big things that’s actually going to mitigate that, too, is to get the expectations right.

That’s one of the things that we’re going to experiment with, is using transfer learning to indeed transfer some knowledge, say from American English accent to a Australian English accent. We’ve also been trying to, for languages that are not so closely related, see how far we can go. But ...

This sounds like a valuable data set. Aren’t there larger data sets with complete information about, say, conversations? For example, we may have just the audio data set transcriptions from people presenting on TV. Is there a larger data set that may be purely text, containing only conversations?

Our software is open, the models are open. They can choose to create a particular data set in whatever language they want to attack. They can use our software to create a speech recognition engine in those languages, sort of trying to break down the silos in terms of in ...

Using these techniques that we’re researching now, we should be able to create the early models for other languages outside of English. That’s basically what we’re doing. Basically, we’re breaking down the silos that exist for speech recognition right now, and then actually empowering communities to build and create speech ...

The interesting thing about that is that, while we’re building up corpora, data sets for other languages, the initial phase where we’re building out the data sets in other languages, these other languages will be under-resourced languages for a bit of time.

Again, we started in English, so we’re working right now on a model, which is open source under Mozilla public license, to do speech recognition in English. We’re going to do other languages, too. Actually, we just got an NSF grant, National Science foundation grant, to create speech recognition models ...

We have an algorithm out there that’s open source. What we’re doing, too, which is not done for most other engines, is we’re training models.

We’re sort of the other end of things. One of the things we’re doing is called Deep Speech, which is our speech recognition engine. The idea is that we have implemented an algorithm, which is using some deep learning techniques, to actually create a speech recognition engine.

I’m Kelly, Manager of the machine learning team. We work on things like machine learning, in terms of speech recognition, speech synthesis, translation, automatic translation, that type of stuff. We are consumers of the data sets that Common Voice is producing.