Embedded Implementation of Distress Situation Identification through Sound Analysis
Résumé
OBJECTIVE: The development of an embedded system capable of detecting distress sounds, e.g. breaking glass or a cry for help, in a person's home and notifying relevant personnel in the case of a distress situation. METHODS: The system is based on a personal computer (PC) equipped with a sound card and microphone that is capable of performing real time analysis of sound signals. Sounds are processed through 4 modules: Sound Event Detection and Extraction, Sound/Speech Classification, Sound Recognition and Speech Recognition. Training, testing and validation of the model was performed using 2 databases - a life sound database which we created and a French adapted speech corpus (a large and structured set of texts recorded by hundreds of different French speakers). RESULTS: The system was found to be reliable for detecting and classifying sounds at signal to noise ratios of 10 decibels (dB) or more, with an error rate of 5% or less. However, it was less efficient at sound and speech recognition. The error rate for sound recognition ranged from 9% to 37% at different sound levels. For speech recognition the error rate was 22%. This comprised 6% due to distress words being picked up in a normal sentence (leading to potentially false distress alerts) and 16% due to a distress word not being recognised (resulting in potentially missed distress alerts). CONCLUSION: An embedded PC, equipped with a classical sound card and a microphone, is capable of real-time detection and analysis of sounds to detect distress situations. The system requires further refinement to improve its accuracy before it can be evaluated in real-life.