Smart home devices have brought transformative benefits in security, convenience, and entertainment. However, their increasing ubiquity also raises serious concerns around technology-motivated domestic abuse, where the perpetrators exploit these smart devices to control over intimate partners. Recent findings indicate that such abuse is more effective and less risky for perpetrators, while being greatly harmful and challenging for victims. This paper presents a hybrid deep learning-based approach to detect early signs of domestic violence through speech emotion recognition using audio data captured via smart microphones. We leveraged both the Toronto Emotional Speech Set (TESS) and Surrey Audio-Visual Expressed Emotion (SAVEE) to train a CNN-LSTM hybrid model. This approach captured both spatial and temporal features of speech to represent emotional signals based on Mel-Frequency Cepstral Coefficients (MFCCs). Our model aims to detect critical emotional states such as fear, sadness, anger, and disgust as potential indicators of abuse in smart homes. By enhancing emotion detection accuracy of existing systems, the proposed model contributes to proactive and privacy-conscious domestic violence mitigation strategies, advocating safer smart connected homes.