Artificial Intelligence for Lyric Generator based on Bi-Directional LSTM and Transformers

In the endeavor of fusing music composition with AI, generation of lyrics has been a persistent issue of concern. Incredible music is a perfect blend of melody and significant lyrics. In this exertion, distinct techniques and strategies have been introduced. The venture of lyric generation using deep learning techniques has been advancing over a decade. One such project in this field is our AI LYRIC WRITER. This project is a comparative study that evaluates the effectiveness of various AI models in generating lyrics. The models are trained on a dataset available on Kaggle, featuring the lyrics by artist Taylor Swift. We have experimented with different models. One of these models is an integration of Bi-directional LSTM and N gram language processing. A bi-directional LSTM architecture processes a sequence of words, both in forward and in backward direction, to understand and remember them better. A CuDNN LSTM layer, which is beneficial for efficient training, is also incorporated into the model. A dropout layer is added to the model which acts as a safety net to prevent overfitting. In an attempt to improve the accuracy, we explored various pre-trained models such as GPT-2, T-5 and XLNET. Although GPT-2 and T-5 were fruitful in learning contexts and patterns, their accuracy was compromised. XLNet proved to be successful in predicting the lyrics aligned with the vocabulary adapted by the artist and reproducing the precise genre employed by the artist. The training data has driven to nearly perfect accuracy in the model's performance.