The derivation of the effective temperature of a star is a critical first step in order to perform a detailed spectroscopic analysis. Spectroscopic methods suffer from systematic errors related to model simplifications. Photometric methods may be more robust, but are exposed to distortions caused by interstellar reddening. Direct methods are difficult to apply, since fundamental data of high accuracy are hard to obtain. We explore a new approach in which the spectrum is used to characterize a star's effective temperature based on a calibration established by a small set of standard stars. We perform Principal Component Analysis on homogeneous libraries of stellar spectra, then calibrate a relationship between the principal components and the effective temperature using a set of stars with reliable effective temperatures. We find that our procedure gives excellent consistency when spectra from a homogeneous set of observations are used. Systematic offsets may appear when combining observations from different sources. Using as reference the spectra of stars with high-quality spectroscopic temperatures in the Elodie library, we define a temperature scale for FG-type disk dwarfs with an internal consistency of about 50K, in excellent agreement with temperatures from direct determinations, but distinct from widely-used scales based in the infrared flux method.