A novel spoken document retrieval system using Auto Associative Neural Network based keyword spotting

This paper formulates a novel approach to spoken document information retrieval for instinctive speech corpora. The conventional method for this problem is to make use of an Automatic Speech Recognizer (ASR) integrated with the typical information retrieval method. However, ASRs tend to produce transcripts of spontaneous speech with momentous word error rate, which is a negative aspect of standard retrieval system. To prevail over such a constraint, we propose a method for spoken document retrieval based on spoken keyword spotting using Auto Associative Neural Networks (AANN). The proposed work concerns the exploit of the distribution capturing capability of an auto associative neural network for spoken keyword detection. It involves sliding a frame-based keyword template along the audio documents and by means of confidence score acquired from the normalized squared error of AANN to competently search for a match. This work provides a new spoken keyword spotting algorithm based spoken documents clustering. The experimental results recommend that the proposed method is promising for retrieving relevant documents of a spoken query as a key.