No abstract

OATAO is an open access repository that collects the work of Toulouse researchers and makes it freely available over the web where possible. ABSTRACT A significant aging of world population is foreseen for the next decades. Thus, developing technologies to empower the independency and assist the elderly are becoming of great interest. In this framework, the IMMED project investigates tele-monitoring technologies to support doctors in the diagnostic and follow-up of dementia illnesses such as Alzheimer. Specifically, water sounds are very useful to track and identify abnormal behaviors form everyday activities (e.g. hygiene, household, cooking, etc.). In this work, we propose a double-stage system to detect this type of sound events. In the first stage, the audio stream is segmented with a simple but effective algorithm based on the Spectral Cover feature. The second stage improves the system precision by classifying the segmented streams into water/non-water sound events using Gammatone Cepstral Coefficients and Support Vector Machines. Experimental results reveal the potential of the combined system, yielding a F-measure higher than 80%.

This work is focused on the automatic recognition of environmental noise sources that affect humans' health and quality of life, namely industrial, aircraft, railway and road traffic. However, the recognition of the latter, which have the largest influence on citizens' daily lives, is still an open issue. Therefore, although considering all the aforementioned noise sources, this paper especially focuses on improving the recognition of road noise events by taking advantage of the perceived noise differences along the road vehicle pass-by (which may be divided into different phases: approaching, passing and receding).To that effect, a hierarchical classification scheme that considers these phases independently has been implemented. The proposed classification scheme yields an averaged classification accuracy of 92.5%, which is, in absolute terms, 3% higher than the baseline (a traditional flat classification scheme without hierarchical structure). In particular, it outperforms the baseline in the classification of light and heavy vehicles, yielding a classification accuracy 7% and 4% higher, respectively. Finally, listening tests are performed to compare the system performance with human recognition ability. The results reveal that, although an expert human listener can achieve higher recognition accuracy than the proposed system, the latter outperforms the non-trained listener in 10% in average.

Acoustic environments are typically composed of multiple sound sources of different typologies, making them especially complex to model and parameterize. To develop an automatic acoustic environment recognition system, this work proposes a spectro-temporal signal parameterization technique inspired by human perception. The proposed parameters are derived from the analysis of the autocorrelation function of narrow-band signals (NB-ACF) obtained from an auditory gammatone filter bank. Five features related to acoustic phenomena are extracted from the NB-ACF to parameterize sound signals. Experiments are conducted on a 4 h sound database (composed of 15 different acoustic environments) employing four different machine learning techniques: K-nearest neighbor, Gaussian mixture models, neural networks and support vector machines. The averaged recognition rates increase by 4.5% when using the proposed NB-ACF features instead of the popular Mel frequency cepstral coefficients. The improvement is even greater (5.6%) when the features are tested in acoustic environments from unknown locations. These results are attributed to the better modeling of the acoustic environments thanks to the complementarity of the spectro-temporal features derived from NB-ACF analysis.

No abstract