Zero-Shot Emotion Style Transfer
Samples and code for the paper Zero Shot Audio to Audio Emotion Transfer With Speaker Disentanglement accepted at ICASSP 2024
We perform two ablations where we remove the two CNN-BiLSTM networks operating on the F0 contour and the HuBERT tokens. The results for these cases are provided in the diagram below. Note that there are two more entries (ZEST-no-F0-enc and ZEST-no-unit-enc) in the table than what was provided in the paper.
We first show the utility of the adversarial learning in the speaker embedding module.
We show the comparison between the x-vectors(left of each group) and the EASE embeddings(right of each group) side-by-side for two different speakers in ESD dataset. The colours in the t-SNE plots are according to the 5 emotions present in the dataset. We note that x-vectors show clear emotion clusters while EASE embeddings are ignorant of the emotion information.
We next show the emotion latent information learnt by the SACE module. We show t-SNE visualization of this latent space and colour them based on the 10 speakers and 5 emotions separately.
Emotion clusters | Speaker clusters |
---|---|
We show 4 examples of the ability of the F0 predictor module to reconstruct the ground truth pitch contour.
Angry | Happy | Sad | Surprise |
---|---|---|---|
We show three examples of how the F0 conversion works in ZEST. We show three examples from three test settings - DSDT (source and reference speaker seen but different with unseen text), USS (Unseen source speaker with seen reference speaker/emotion and unseen text) and UTE (seen source speaker with unseen reference speaker/emotion and unseen text). The first row indicates the F0 contour of the source speech signal, the second row indicates pitch contour of the reference speech signal while the last row indicates the converted F0 contour.
DSDT | USS | UTE |
---|---|---|