F0 Controllable General-Purpose Vocoder Enhanced Using Speaker Vectors

In a neural vocoder capable of generating high-quality waveform, constructing a multi-speaker vocoder by training with data from various speakers reduces the necessary amount of target speaker’s data. However, this approach often results in inferior vocoder quality compared to using data from a single speaker. In this study, we propose enhancing the quality of multi-speaker vocoders by incorporating speaker vectors into the input, based on models derived from the source-filter theorem. Evaluation experiments demonstrate improved model quality when speaker vectors are included as input.