Symmetric secret key generation from physical layer measurements has recently emerged as an alternative to computationally secure protocols or predistribution techniques for key sharing. This new approach takes advantage of the intrinsic properties of radio channels (e.g. reciprocity and spatial decorrelation) by quantizing measurable channel characteristics into a suite of codewords. However, noise or other transmission artifacts can degrade the link reciprocity and lead to bit errors between the users' codewords, which are strongly related to the quantization thresholds. In this paper, we present a framework for quantization in the context of key generation. In order to find a trade-off between reciprocity (an indicator of robustness) and codeword diversity (an indicator of randomness) under a fixed-length constraint, we investigate the non-uniform quantization and we show the interest of optimizing the quantization thresholds as a function of the signal-to-noise ratio.