Formant bandwidths are severely underestimated in LPC analysis of voiced segments with low F1/F0 ratio. Parameter quantization errors aggravate the situation and cause steady nasal sounds and high vowels to have amplitude booms. Spectral smoothing using a lag window was used, as a possible remedy. It is found, however. that indiscriminate smoothing reduces the crispness of the synthesized speech. Lag windowing is introduced, only for frames with low F1/F0. Detection of such frames is made using a pattern classification approach after a preliminary reduced order LPC analysis stage. Estimates of F1 and F2 are found by a closed form solution of the 4th order inverse filter polynomial. A Fisher classifier is used to detect the condition of low F1/F0, If a frame is classified as having low F1/F0 the autocorrelation vector is multiplied by a binomial window. Results show improved quality over both indiscriminate spectral smoothing and the standard autocorrelation methods.