Context-Aware Emergency Sound Recognition System for Enhancing Situational Awareness in Hearing-Impaired Individuals

People with hearing impairments can miss environmental cues that signal immediate danger, such as emergency sirens, vehicle accidents, and gunshots. This difficulty in detecting critical environmental sounds can increase their safety risks, especially in busy or unpredictable environments. Most existing assistive sound recognition systems operate in controlled indoor settings, treat all sirens as a single undifferentiated category, and often do not consider the user’s surrounding environment or the relative risk associated with a detected event. These gaps limit the ability of hearing-impaired individuals to respond appropriately during real-world emergencies when situational context is important. We present a context-aware emergency sound recognition machine learning framework that combines sound classification with environmental awareness to provide more informative and personalized risk-graded alerts. The proposed system consists of three components. An emergency sound detection model, an environment identification model, and a risk identification matrix. The Emergency Sound Detection Model is a Convolutional Neural Network trained on 128-band Mel spectrograms extracted from audio standardized to 3-second clips at 22,050 Hz. This model recognizes five emergency sound categories(ambulance, police, firetruck, gunshot, vehicle accident) and a general "other" class. To improve robustness to real-world noise, the model uses gain, time, and pitch shifting, background-noise injection, and SpecAugment-style time and frequency masking with class weighting and learning rate scheduling applied during training. The second component, the Environment Identification Model, classifies the user's surroundings into indoor, outdoor, or in-transport categories using a lightweight classifier trained on embeddings from YAMNet, which is a model pretrained on YouTube AudioSet. The outputs of both models feed into a structured Risk Identification Matrix, which is the third component of the system. This Matrix is manually constructed from public safety guidelines (e.g., FEMA, CISA) that map sound type and environment combinations to a High/Medium/Low risk level. Unlike prior systems that issue undifferentiated alerts, our system communicates this risk level through both text and American Sign Language video feedback. The Emergency Sound Detection Model achieved 94% classification accuracy and a macro-averaged F1 score of 0.92, outperforming classical baselines (SVM: 87%, Random Forest: 85%, KNN: 78%) and a non-augmented CNN baseline (88%). The Environment Identification Model using YAMNet embeddings with a dense classifier reached 89% accuracy, exceeding a Mel-spectrogram CNN baseline (82%) and matching a soft-voting ensemble of both approaches. A real-world validation study with 10 participants, including 6 hearing-impaired individuals, produced a 4.6/5 overall satisfaction rating, with all 6 hearing-impaired participants reporting improved situational awareness. The results demonstrate that context-aware sound recognition can extend conventional emergency sound detection by incorporating information about where an event occurs and translating model outputs into risk-aware accessibility feedback. Future work will focus on efficient real-time deployment on smartphones, expanded sound categories, personalized alerting, and broader evaluation with hearing-impaired users in real-world environments.