System overview

The Castsearch demo is a system for retrieval of relevant stories from broadcast news. The system analyses data using audio processing, speech recognition and text mining techniques to find clips relevant for textual queries. The demo is implemented using hourly podcasts from CNN gathered during the last year.

The system includes an audio analysis part that separates jingles from speech parts and finds changes between speakers to produce a number clips containing one speaker in one context as described in [1] and [2].

The resulting clips are transcribed using the SPHINX-4 large vocabulary automatic speech recognition engine to produce a text document for each speech clip, to yield text input for clustering using non-negative matrix factorization (NMF).

The NMF is used to find semantic topics that are used to evaluate the performance for topic detection. Based on these topics we show that a novel query expansion can be performed to return more intelligent search results. We show that the query expansion helps overcome errors of the automatic transcription.

  1. K. W. Jørgensen, L. L. Mølgaard, L. K. Hansen, Unsupervised Speaker Change Detection for Broadcast News Segmentation, Eusipco, 2006
  2. L. L. Mølgaard, K. W. Jørgensen, L. K. Hansen, Castsearch - Context Based Spoken Document Retrieval, ICASSP, 2007