Speech Summarization using Essence Vector Modeling

Speech is the most natural and effective method of communication between human beings. It is not easy to review and retrieve the information present in the spoken document instantly when the speech is recorded and stored as an audio signal. Speech summarization extracts important information and saves time for reviewing speech documents and improves the efficiency of document retrieval. Representation learning and paragraph embedding methods have emerged as new active research subjects because of their excellent performance in many applications. In this research, we have implemented an unsupervised paragraph embedding method called Essence vector (EV) modeling for Speech Summarization, which strives at distilling the most characteristic information from a paragraph and also includes the general background information to produce a more informative low-dimensional vector representation of the input. Speech is converted into text and provided as a high dimensional vector input to the Essence Vector (EV) model which in turn summarizes the input in a low dimensional space. We have enhanced the performance of the EV model by using LSTM, a state-of-the-art neural network architecture, to handle the imperfect and erroneous Speech conversions made by the speech to text module and also added Attention mechanism to the EV model to obtain a better and relevant summary. The proposed method is evaluated using ROUGE metrics against a summarization corpus.