Extracting context from environmental audio

Computational auditory scene recognition

“…In 1998, Brian Clarkson and his colleagues proposed a system for extracting context from environmental audio [Clarkson98b]. The classification scheme consisted of the normalized power spectrum as a feature, and hidden Markov models as a classifier.…”

SoundButton: design of a low power wearable audio classification system

“…To date, the potential of wearable context recognition based on sound has been studied in some detail by two research groups: Auditory scene analysis focused on detecting distinct auditory events and classifying them has been done by MIT's Media Lab [4,5]. The Audio Research Group at Tampere University, Finland, works on auditory scene recognition, which focuses on recognizing the context or environment, instead of analyzing discrete sound events [13].…”

Classifying User Environment for Mobile Applications using Linear Autoencoding of Ambient Audio

“…Clarkson and Pentland studied user context awareness using audio, video, and other sensory streams [2] [3], [4], [5], [6] in the context of a system designed to extract personal life patterns from sensory data. This system employed featurelevel fusion and HMM clustering techniques to learn common scenarios in everyday life.…”

“…The data were resampled at 16 kHz, 16-bit mono for use in this experiment. To help ensure the predominance of ambient sound in the training and test sets, we calculated the mean power of each recording and selected for further study only those segments that were quieter than average (this amounts to the inverse of the event detection procedure described in [2]). After applying this procedure, the data were divided into a training set and a test set; 80% of the segments were selected for training; the remainder for testing.…”

“…We used seven different training configurations to test the effect of varying numbers of parameters. Specifically, we trained models with 2,4,8,12,16,20, and 24 hidden units or gaussians. Each autoencoder was initialized with random weights between -0.05 and 0.05 and trained on the whitened features for 100 iterations using a batch-mode backpropagation algorithm with an adaptive learning rate initialized at 0.05, a momentum term of 0.045, and batch shuffling.…”