The length of this sequence greatly affects the network’s predicting abilities since 5 words in a row work much better than just a single word. The network receives as input a sequence of lyrics and predicts the next word to appear. LSTMs have proven in the past to be successful in similar tasks because of their ability to remember previous data, which in our case is relevant because each lyric depends on the words (and melody) that preceded it. We implemented this using an LSTM network. 20% of the training data was used as a validation set in order to track our progress between training iterations. The melody files and lyrics for each song were given to us and the train / test sets were predefined. For the training phase, however, we used Crossed Entropy loss. However, this is quite subjective leading the evaluation of generated words to use imaginative methods. This is essentially done by generating new words for the song and attempting to be as “close” as possible to the original lyrics. In this assignment, we were tasked with creating a Recurrent Neural Network that can learn song lyrics and their melodies and then given a melody and a few words to start with, predict the rest of the song. Analysis of how the Seed and Melody Effects the Generated Lyrics.However, this is entirely subjective leading the evaluation of generated words to use imaginative methods. A Recurrent Neural Network that can learn song lyrics and their melodies and then given a melody and a few words to start with, predict the rest of the song.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |