A Word-axis Speaker Embedding Trained with Multi-Speaker Analysis Task

In recent research, there has been an increased use of pre-trained speaker vectors as conditions for voice synthesis. While these are useful for speaker verification, they have intrinsic limitations in the range of information they cover, leading to a situation where the vocoder plays a critical role in learning and reconstructing formants. In this paper, we aim to extract time-variant speaker embeddings that are more conducive to speech synthesis and have a cognitively meaningful distribution. We propose and evaluate a method for extracting speaker embeddings along the word-axis, which is multi-task trained in a multi-speaker analysis context. Our method has shown improved performance over existing methods both in speaker verification tasks and in aligning with human perception.