| 摘要 |
PURPOSE: A method and a system for learning English pronunciation using a voice recognition technique are provided to allow a user to accurately diagnose his pronunciation. CONSTITUTION: A voice signal is inputted. Characteristics are extracted from the inputted voice signal and converted into a characteristic vector, and a characteristic vector including rhythm information is extracted in a discrete cosine transform step. A sound model is trained using continuously distributed Hidden Markov model for a sound model of each phoneme. The pronunciation of a user is inputted and relatively compared with a probability value with respect to a candidate phoneme stream generated using a pseudo phoneme list to point out accuracy and errors. A probability value with respect to the input voice signal is calculated by applying the trained Hidden Markov model corresponding to the candidate phoneme stream to determine ranking of phoneme streams. A correct ranking of phoneme streams and distribution of candidate phoneme streams are examined to generate the final estimation result. Statistics of the estimation results for phonemes of the user's pronunciation are taken. |
