Botnets are the machines that increasingly controlled by cybercriminals to perform various attacks. They use Domain Generation Algorithm (DGA) to frequently generate their illegitimate domains for preventing detection. To overcome such dynamics, existing solutions try to capture the characteristics of domain names, such that the automatically generated domains can be identified. However, those solutions are not conformed to the linguistic conventions of reading and writing. For a comprehensive understanding of strings of domain names, we present DOmain Linguistic PHonIcs detectioN (DOLPHIN), a novel method that can detect the illegitimate domain names generated by DGAs. Considering the correspondence between pronunciations and spellings, we design the DOLPHIN patterns. They are the classification of vowels and consonants in variable lengths as follow the principles of phonics. DOLPHIN recognizes strings of domain names and reconstructs them with the components of variable-length vowels and consonants following the DOLPHIN patterns. We implement the features used DOLPHIN in supervised learning methods and compare them to the fore-most method FANCI. Experimental results show that, compared to FANCI with RFs, DOLPHIN can achieve higher detection accuracy of 0.0238 in average with lower FPR without much overhead.