Spoken language can include sensitive topics including profanity, insults, political and offensive speech. In order to engage in contextually appropriate conversations, it is essential for voice services such as Alexa, Google Assistant, Siri, etc. to detect sensitive topics in the conversations and react appropriately. A simple approach to detect sensitive topics is to use regular expression or keyword based rules. However, keyword based rules have several drawbacks: (1) coverage (recall) depends on the exhaustiveness of the keywords, and (2) rules do not scale and generalize well even for minor variations of the keywords.Machine learning (ML) approaches provide the potential benefit of generalization, but require large volumes of training data, which is difficult to obtain for sparse data problems. This paper describes: (1) a ML based solution that uses training data(2.1M dataset), obtained from synthetic generation and semi-supervised learning techniques, to detect sensitive content in spoken language; and (2) the results of evaluating its performance on several million test instances of live utterances. The results show that our ML models have very high precision (>>90%). Moreover, in spite of relying on synthetic training data, the ML models are able to generalize beyond the training data to identify significantly higher amounts (∼2x for Logistic Regression, and∼4x-6x for a Neural Network models such as Bi-LSTM and CNN) of the test stream as sensitive in comparison to a baseline approach using the training data (∼1Million examples) as rules. We are able to train our models with very few manual annotations. The percentage share of sensitive examples in our training dataset from synthetic generation using templates and manual annotations are 98.04% and 1.96%, respectively. The percentage share of non-sensitive examples in our training dataset from synthetic generation using templates, automated labeling via semi-supervised techniques, and manual annotations are 15.35%, 83.75%, and 0.90%, respectively. The neural network models (Bi-LSTM and CNN) also use lower memory footprint (22.5% lower than baseline and 80% lower than Logistic Regression) while giving improved accuracy.