An Improved Template-Based Approach to Keyword Spotting Applied to the Spoken Content of User Generated Video Blogs

Noise robust keyword spotting for user generated video blogs

“…The template matching-based KWS introduced in [8,9] depends on the theory and observation that measuring the DTW distances of sliding template over an utterance results in a minimal consecutive values at the position of the word for a wider number of frames than non keyword regions. This leads to difference between the histograms formed from template comparison against utterances that contain the keyword and utterances that do not contain the keyword.…”

“…Hence, while one approach would be to analyze the output text of an automatic speech recognition (ASR) system, the accuracy of this approach for spontaneous speech is generally quite low (around 40% error rates) [5,7] compared to cleaner, more structured speech such as broadcast news [6]. To overcome this challenge, a keyword spotting (KWS) approach can be used [8][9][10][11][12][13].…”

“…To overcome the challenge associated with the requirement for a large amount of labeled training data associated with a Hidden Markov Model (HMM) approach, template matching approaches that avoid training have been successfully used [8,9,[11][12][13][14]. The focus of this paper is on a segmental-based Dynamic Time Warping (DTW) approach, which continually compares a spoken word template with segments of an utterance, with matching performed in the speech feature (Mel Frequency Cepstral Coefficient (MFCC)) domain [8,9,13,14].…”

“…The focus of this paper is on a segmental-based Dynamic Time Warping (DTW) approach, which continually compares a spoken word template with segments of an utterance, with matching performed in the speech feature (Mel Frequency Cepstral Coefficient (MFCC)) domain [8,9,13,14].…”

“…A speaker independent approach was proposed in [12], however it requires training to generate posteriograms and suffers from a performance drop when the training and test environments differ [16]. Alternatively, the distance histogram analysis based template matching introduced in [8,9], which does not require training, is selected as the basis of this work and will be investigated in this paper.…”