Context : The determination of stellar effective temperature ( T _ { \mathrm { eff } } ) in F , G , and K stars using H \alpha profile fitting is a quite remarkable and powerful tool , because it practically does not depend on other atmospheric parameters and reddening .
Nevertheless , this technique is not frequently used because of the complex procedure to recover the profile of broad lines in echelle spectra .
As a consequence , tests performed on different models have sometimes provided ambiguous results .

Aims : The main aim of this work is to test the H \alpha profile fitting technique to derive stellar effective temperature .
To improve its applicability to echelle spectra and to test how well 1D + LTE models perform on a variety of F-K stars .
We also apply the technique to HARPS spectra and test with the Sun the reliability and the stability of the HARPS response over several years .

Methods : We have therefore developed a normalization method for recovering undistorted H \alpha profiles and we have first applied it to spectra acquired with the single order coudé instrument ( resolution R = 45 ~ { } 000 ) at do Pico dos Dias Observatory to avoid the problem of blaze correction .
The continuum location around H \alpha is optimized using an iterative procedure , where the identification of minute telluric features is performed .
A set of spectra was acquired with the MUSICOS echelle spectrograph ( R = 40 ~ { } 000 ) to independently validate the normalization method .
The accuracy of the method and of the 1D + LTE model is determined using coudé/HARPS/MUSICOS spectra of the Sun and only coudé spectra of a sample of 10 Gaia Benchmark Stars with effective temperature determined from interferometric measurements .
HARPS spectra ( R = 100 ~ { } 000 ) are used to determine the effective temperature of 26 stars in common with the coudé data set by the same procedure .

Results : We find that a proper choice of spectral windows of fits plus the identification of telluric features allow a very careful normalization of the spectra and produce reliable H \alpha profiles .
We also find that the most used solar atlases can not be used as templates for H \alpha temperature diagnostics without renormalization .
The comparison with the Sun shows that the effective temperatures derived by us with H \alpha profiles from 1D + LTE models underestimate the solar effective temperature by 28 K. A very good agreement is found with the interferometric benchmarks and with the Infrared Flux Method determination , that shows a shallow dependency on metallicity according to the relation T _ { \mathrm { eff } } = T _ { \mathrm { eff } } ^ { H \alpha } -159 [ Fe/H ] + 28 K within the metallicity range -0.7 to +0.45 dex .
The comparison with Infrared Flux Method show a 59 K scatter dominated by photometric errors ( 52 K ) .
In order to investigate the origin of this dependency , we analyzed in the same way spectra generated by 3D models and found that they produce hotter temperatures , and that their use largely improve the agreement with the interferometric and Infrared Flux Method measurements .
Finally , we find HARPS spectra to be fully suitable for H \alpha profiles temperature diagnostics , they are perfectly compatible with the coudé spectra , and the same effective temperature for the Sun is found analyzing HARPS spectra over a time span of more than 7 years .

Conclusions :