Query by example keyword spotting in streams of audio

Finding occurrences of utterances whether are single words or phrases in streams of audio is very useful for fast retrieval of audio files as in querying by content in big collection of audio files. I is also useful for monitoring audio signals for security reasons for example searching for alert signals in the fight against terrorism. To achieve such goals we need to use Feature extraction techniques that are speaker independent and for detecting occurrences we propose to use approximate string matching techniques normally used when searching for substrings in long DNA strings allowing errors. We adapted feature extraction methods used in speech processing to take the problem from speech space to string search space. We use the Edit distance, the Levenshtein distance, the Longest Common Subsequence (LCS) distance, time-warped LCS distance, and Dynamic Time Warping (DTW) as a reference. We evaluated our system with a collection of audio files and validated it manually getting encouraging results.