Unsupervised Audio Scene Analysis

The Problem: This work addresses the problem of robustly monitoring a scene using a microphone or microphone array. Rather than concentrating on human speech, we are attempting to model the general sounds that occur in a particular scene and to look for long-term contexts of those sounds and to look for anomalous sounds. Motivation: While many company’s are expending great effort in the field of automatic speech recognition(ASR), little attention is being paid to general audio and long-term modeling of audio in general. Even an ASR system which could give a complete transcription of the words heard in an environment would lack vital information. e.g., who was talking, when they were talking, what was the tone of the conversation, did someone slam the door, did someone use a harsh expletive, when do these conversations tend to happen, etc. In fact, this information is potentially of great value even without the transcription of what was said. Previous Work: Dan Ellis has done significant work in decomposing audio stream into separate audio elements resulting from independent sources[1]. He has produced a summary of the work done in this area[2]. Approach: Rather than concentrating on the low-level segmentation of separate sources, which generally runs about three orders of magnitude slower than real-time, we are going to rely on our microphones being in a sparse environ-ment. This will allow us to use a rather simple segmentation procedure and concentrate on the high-level issues. The high-level issues we are going to approach are clustering the segmented audio into different audio sources, learnin