An efficient codebook driven voiced excitation coding method producing natural sounding speech is proposed. It can be incorporated as an essential part in a complete speech coder working at low bit rates. The interpulse correlation of such a coding scheme is investigated and exploited using linear predictive vector quantization and finite state vector quantization (FSVQ). A new and more robust FSVQ method that is able to more efficiently exploit this correlation is proposed. The new method dynamically combines two memoryless vector quantizers. This dynamic combination method is not restricted to pulse codebooks but can be employed for any coding scheme with intervector correlation.