Search for Keywords and Vocal Elements in Audio Recordings

摘要

This paper deals with search for keywords and non-verbal vocal elements in audio recordings. An efficient detection of specific words or sounds embedded in continuous speech is based on isolated word recognition approaches. The mel-frequency cepstral coefficients and more combinations of predictive coefficients and autocorrelation coefficients were evaluated. A keyword or key sound slides along the stored speech and in each of its positions a distance (i.e., similarity) to the corresponding speech segment is computed. We found an efficient distance measure for non-verbal sound search. The average detection rates achieved 93 percent in keyword search and 74 percent in non-verbal sound search. A system developed for automatic search in audio files is presented.