A Problem Finding Optimal Number of Sentences and Possible Solutions

The aim of this project is building a speech corpus which will be used for DL and integration with a speech engine. The speech engine will then be integrated with a screen reader for the visually impaired. For the speech corpus, we had to record as many sentences as possible with a balanced set of letters and pronunciation. We start with Albanian - there are 36 letters in Albanian, some are rarely used such as Zh and Xh. In order to extract sentences with a balanced set of letters we worked with a bank of e-books in Albanian, and in total 500 e-books. We built a PDF converter that went through all of the books and extracted all the sentences. We then had to find a way to remove sentences in order to create a balanced set of sentences with specific letters.