In　this　paper,　different　ways　of　training　codebook　containing　autoregressive　(AR)　parameter　vectors　are　discussed.　The　fundamental　goal　of　the　discussion　is　to　investigate　if　the　classical　approach　for　training　AR-codebooks　by　clustering　line　spectral　frequencies　(LSF)　can　be　improved.　To　do　this,　we　discuss　and　evaluate　the　alternatives　in　terms　of　the　de-correlated　AR-parameters　and　manifold　learning.　The　different　training　methods　are　evaluated　using　different　metrics　quantifying　the　distance　between　actual　power　spectral　density　(PSD)　and　the　estimated　PSD　from　the　AR-codebook.　The　experimental　results　show　that　the　training　on　the　de-correlated　features　can　improve　the　performance　to　some　degree　compared　to　the　traditional　LSF　training　approach　in　terms　of　the　Itakura-Saito　divergence　not　in　terms　of　the　Kullback-Leibler　divergence,　the　log-spectral　distortion　and　speech　distortion.