I want to know something more about it. Are there any github repo or an open source project?
It is possible
The Recurrent Neural Network architectures help in building efficient NLP algorithms, which can identify semantics and their relations across long pieces of text.
With very minor tuning, they can be made generative too. So, here is an excellent article on RNNs which I highly recommend, which also talks about how an RNN was trained on Shakespere's texts and wrote one itself.
Ya Sure it is possible. It wont be efficient as human speech though (At least not yet). It all depends on how you use your data. If you use your data efficient enough, then you could be the one creating an AI which closely resembles human speech. Your idea is good. You would need help from a GPU for processor all that complex text processing. I hope I was at least a little of help. I unfortunately don't know about any open source projects.
It depends whether you include learning to produce sound waves in the task (I'm taking "how to speak" to be different from "how to write" in your question).
If the task is to learn a language, you can certainly learn from written texts, and try a generative approach (usually, after a few sentences, humans can be disappointed though).
If the task includes generating a voice response that is also learnt, then it had better hear some spoken language too. Of course, you can bypass the task of learning how to generate sound by plugging a standard text to voice module after your generator.
Yes, It is possible. Also, if you are from development background you can use JAVA Sphinx, some AI tools and API's. If you are from testing background you can use JAWS which actually reads the readable properties of HTML tags. Nowadays, there are some pdf readers, document reader softwares available you can have a look at them. You can search for Sphinx related projects but I don't know about the other projects.
Please vote and mark the solution if useful.