This paper deals with careful non-real-time LPC analysis. A baseline system is first described. It uses a pitch-synchronous covariance-method analysis with a laryngograph signal providing the pitch synchrony. Work to improve the voicing decision and F0 determination and to find a better voiced excitation waveform is described. Setting a lower limit on the value of B-, is found to be useful. Buzziness in the synthesis of voiced fricatives can be reduced by adding white noise to the excitation, and regions where this should be done can be automatically detected using three parameters: total power, the ratio of high-frequency power to total power, and B^. The location of the analysis frame for covariance-method analysis is found to be unimportant. A modified autocorrelation method with a carefully placed, Manning-windowed, pitch-synchronous analysis frame is found to give results that are as good as or -better than the covariance method. Finally, a method of resynthesizing speech using a modified LPC residual is described. The method allows prosody and formant parameters to be manipulated with minima! degradation.