A novel linear prediction based parameter estimation algorithm is developed for analyzing deep level transient spectroscopy (DLTS) signals. The algorithm performs significantly better than a current linear prediction based algorithm used in DLTS because it accurately accounts for the effects of noise and any underlying baseline constant. The algorithm is developed for any digitized isothermal capacitance transient. It does not rely on overmodeling or require baseline nulling hardware. The superior performance of the algorithm is verified on synthesized, as well as challenging actual DLTS signals. It is shown to consistently extend the linear regions and resolve closely spaced activation energies on Arrhenius plots.