Fusion of dictionaries in voice creation and speech synthesis task

The accurate phonetic transcription is very important for different fields of speech technologies. In speech synthesis, it is the benchmark for the voice segmentation, and therefore one of the crucial points for the synthesized speech pronunciation quality. In ASR the availability of matching phonetic transcription allows a higher recognition precision. Use of different dictionaries could improve the phonetic transcription since it allows a better word coverage, but the “direct” dictionary merging presents incompatibility problems. Dictionary fusion is the method that automatically learns the dictionary-to-dictionary transformation rules. The results presented in this paper show that fusion significantly improves the compatibility between the dictionaries (about 32-83% of improvement), and allows reducing the number of “severe” pronunciation errors in comparison with the graphemeto-phoneme (g2p) techniques.