I like the idea of following conversations as a learning tool, but this has been done by a lot of software products. May I suggest that you use real conversations? Find several native speakers and ask if you could record their actual conversations. You may have to do this over several weeks to get enough elements of current speech. Make transcripts of the conversations to use in your software. Make absolutely certain the audio is available with your software. You should probably provide the ability to listen to both fluent speech and individual words. You may have to have someone else pronounce the individual words.
As each lesson progresses, you may just want to showcase certain items (not the whole conversation, but provide the whole thing for ear training exercises), like subjects, then objects, and adjectives, then verbs. I believe we all learn things a little bit differently, so this approach may not work for everyone. But, I think helping someone listen for key elements is always going to work. You may want to let the user drive (let him/her decide what elements to listen for and learn).
