Multi-speaker articulatory trajectory formation based on speaker-independent articulatory HMMs

摘要

Inter-speaker variability in the speech spectrum domain has been modeled using speaker-adaptive training (SAT), in which speaker-independent phoneme-specific hidden Markov models (HMMs) were used along with a speaker-adaptive matrix. In this paper, multi-speaker articulatory trajectory formation based on this method is presented. Both speaker-independent and speaker-specific features are statistically separated from a multi-speaker articulatory database, which consists of the mid-sagittal motion data of the lips, incisor, and tongue measured with an electro-magnetic articulographic (EMA) system. We evaluated the proposed method in terms of the RMS error between the measured and estimated articulatory parameters. When multi-speaker models of articulatory parameters with two speaker-adaptive matrices for each speaker were used, the average RMS error of articulatory parameters was 1.29 mm and showed no statistically significant difference from that for speaker-dependent models (1.22 mm). For comparison, multi-speaker models of the conventional speech spectrum were also constructed using a multi-speaker spectrum database, which consists of speech data simultaneously recorded during the articulatory measurements. The average spectral distance between the vocal-tract and estimated spectrum from two-matrix models was 4.19 dB and showed a statistically significant difference from that for speaker-dependent models (3.97 dB). These results indicate that modeling of inter-speaker variability in the articulatory parameter domain with a small number of matrices for each speaker almost perfectly approximates the speaker dependency of articulation and is better than that in the speech spectrum domain.