Mental health disorders such as stress and anxi- ety have become major concerns worldwide due to increasing workload, lifestyle imbalance, and psychological pressure. Early detection can enable timely intervention, but traditional assess- ment methods are subjective and resource-intensive. This paper presents a voice-based stress and anxiety detection framework using a synthetic speech dataset. Audio signals are preprocessed and transformed into a 61-dimensional acoustic feature set com- prising Mel Frequency Cepstral Coefficients (MFCC), chroma features, spectral contrast, zero-crossing rate, and RMS energy. Principal Component Analysis (PCA) reduces dimensionality while retaining 95% variance. A Support Vector Machine (SVM) with RBF kernel is trained and evaluated using five-fold stratified cross-validation, achieving a mean accuracy of 92.4%. For a test audio file, the system not only predicts the class (normal, anxiety, or stress) with confidence but also generates a detailed PDF report that includes a sliding-window anxiety trend over time, fluency score, speech rate, pitch variation, rule-based obser- vations, personalised suggestions, and an overall communication score. This interpretable output bridges the gap between raw classification and actionable feedback. The proposed method can be deployed in healthcare monitoring systems, telemedicine platforms, and intelligent mental health assessment applications. Index Terms—Stress Detection, Anxiety Detection, Voice Anal- ysis, Deep Learning, Machine Learning, MFCC, XGBoost, Ran-

dom Forest, Speech Emotion Recognition