Data Alignment and Duration Modelling in VITS

The paper analyses data alignment and duration modelling in the modern end-to-end speech synthesis model VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech). The standard version of VITS utilizes the MAS (Monotonic Alignment Search) procedure to align input text/phones and corresponding speech during the training procedure; the alignment is also used to obtain phoneme durations for the stochastic duration predictor training. This study analyzes the resulting MAS alignment and compares it with a reference alignment obtained by an LSTM-based phonetic segmentation system. We also examine the performance of VITS when the reference phonetic segmentation replaces the default MAS alignment. The comparison shows that while the original VITS is still slightly preferred in terms of quality, it provides a less interpretative data alignment. The duration modelling is more transparent in the modified version, allowing better duration control and modifications. The analysis has been carried out on two Czech voices.