Principled Coarse-Graining (PCG) accelerates speech token generation by verifying proposals at the level of Acoustic Similarity Groups (ASGs) derived from the target model’s embedding space. This approach increases acceptance and throughput while maintaining intelligibility and speaker similarity.