Wireless indoor localization is a key technology for the future Internet of things (IoT) paradigm. In this paper, we perform an experimental comparative study of machine learning-based localization schemes, such as k-nearest neighbor (k-NN) and variants of support vector machine (SVM), based on the received signal strength (RSS) measurements of the ambient frequency modulation (FM) and digital video broadcasting- terrestrial (DVB-T) signals in three real testbed environments. The consideration of readily available, ambient radio signals frees the need for dedicated radio transmitters. Noise-reduction techniques such as feature selection and ensemble learning are proposed in conjunction with SVM. Our results examine the performance comparisons between SVM and k-NN, as well as the performance comparisons of SVM-based methods incorporating different noise-reduction schemes, with noisy RSS data. Insights into the performance of learning- based localization schemes working with real database collected from real environments are provided.