Sensor linearization is an important aspect for enhancing the efficiencies of measuring systems. Conventional methods use additional circuits and/or software models to achieve linearization. Direct interface technique obviates the requirement of intermediate electronic circuits, including linearization circuits, between sensors and embedded systems. Sensor linearization strategies need to be explored in order to obtain maximized performances from directly interfaced sensing systems. In this work, performances of both hardware and software-based strategies, for linearization of directly interfaced thermistor sensors are evaluated. Experimental results show that, hardware-based linearization approach can yield a maximum linearized output range i.e. from 0 to ~70 oC with < 1% FSS nonlinearity error (NLE) with shunting method using only single pin (1P_Shunt). Sensitivities in both 1P_Shunt and 2P_Shunt cases are found comparable. Embedded system with higher timer speed (f = 42 MHz) is required to minimize quantization errors. In 1P_Shunt, the linearized range is independent of the β-values ranging from 3012 K to 3924 K. With Artificial Neural Network (ANN) based linearization approach, a linearized range up to 100 oC and beyond can be achieved. A shallow network having optimum architecture (1-5-1) with bayesian regularization and log-sigmoid, as activation function is found sufficient to yield < 1% FSS nonlinearity error.