SPEECH SYNTHESIS DEVICE, SPEECH SYNTHESIS METHOD, SPEECH SYNTHESIS PROGRAM, SPEECH SYNTHESIS MODEL LEARNING DEVICE, SPEECH SYNTHESIS MODEL LEARNING METHOD, AND SPEECH SYNTHESIS MODEL LEARNING PROGRAM

摘要

The purpose of the invention is to prevent degradation in speech and unnatural phoneme duration. A speech synthesis device according to an embodiment comprises a storage unit, a creation unit, a determination unit, a generation unit, and a waveform generation unit. The storage unit stores, as statistical model information, an output distribution of acoustic characteristic parameters including pitch characteristic parameters, and a duration distribution by time parameters in each state of a statistical model having a plurality of states. The creation unit creates a statistical model series from the statistical model information and context information that corresponds to an input text. The determination unit determines the number of pitch waveforms for each state by employing the duration based on the duration distribution in each state of each statistical model in the statistical model series, and pitch information based on the output distribution of the pitch characteristic parameters. The generation unit generates an output distribution sequence of the acoustic characteristic parameters on the basis of the number of pitch waveforms, and generates acoustic characteristic parameters on the basis of the output distribution sequence. The waveform generation unit generates a speech waveform from the generated acoustic characteristic parameters.