We propose a novel framework for synchronization in feature-based data embedding systems. The framework is tolerant to de-synchronizing errors in feature estimates, which have hitherto crippled feature-based embedding methods. The method uses a concatenated coding system comprising of an outer q-ary LDPC code and an inner insertion-deletion code to recover from both desynchronization caused by feature estimation discrepancies between the transmitter and receiver; and errors in estimated symbols arising from other channel perturbations. We illustrate the framework in a speech watermarking application employing pitch modification for dataembedding. We show that the method indeed allows recovery of watermark data even in the presence of desynchronization errors in the underlying pitch-based embedding. The resilience of the method is also demonstrated over channels employing low bit rate speech encoders.