Pitch Prediction from MFCC Vectors Using Support Vector Regression

Mel-frequency cepstral coefficients (MFCC) are proved to be the effective feature for speech recognition and speaker recognition, while pitch frequency is also one of the favorite prosodic features. This work manages to bridge them with hidden Markov model (HMM) and support vector regression (SVR). A set of speaker-independent HMMs is used to align the training data, so that the parameters of the local SVR models can be trained from the pooled data with the same label. To evaluate the accuracy of the regression functions, MFCC streams are extracted from the test data, and the pitch frequencies are predicted from them with the trained mapping functions. The comparison between the predicted pitch contour and the reference one from PRAAT proves the strong bind of pitch frequency to MFCC vectors. And the HMM and the SVR are the proper models to characterize the connection.