Conventional procedures for spectrum coding for speech address fixed rate coding. For the variable rate case, we develop spectrum coding based on constrained entropy quantization. Our approach integrates high rate theory for Gaussian mixture modeling with lattices based on line spectrum pairs. The overall procedure utilizes a union of several lattices in order to enhance performance and to comply with source statistics. We provide experimental results in terms of SD for different conditions and compare these with high rate lower bounds. One major advantage of our coding system concerns adaptivity, one design can operate at a variety of rates without re-training.