Sequence-to-sequence Singing Synthesis Using the Feed-forward Transformer

TL;DR

This paper introduces a feed-forward Transformer-based sequence-to-sequence singing synthesizer that eliminates the need for pre-aligned training data, enabling faster inference and reducing exposure bias.

Contribution

It presents a novel feed-forward Transformer model for singing synthesis that leverages approximate initial alignments and refines them without autoregressive decoding.

Findings

Faster inference compared to autoregressive models

Effective refinement of initial alignments using self-attention

Importance of duration model accuracy for synthesis quality

Abstract

We propose a sequence-to-sequence singing synthesizer, which avoids the need for training data with pre-aligned phonetic and acoustic features. Rather than the more common approach of a content-based attention mechanism combined with an autoregressive decoder, we use a different mechanism suitable for feed-forward synthesis. Given that phonetic timings in singing are highly constrained by the musical score, we derive an approximate initial alignment with the help of a simple duration model. Then, using a decoder based on a feed-forward variant of the Transformer model, a series of self-attention and convolutional layers refines the result of the initial alignment to reach the target acoustic features. Advantages of this approach include faster inference and avoiding the exposure bias issues that affect autoregressive models trained by teacher forcing. We evaluate the effectiveness of this model compared to an autoregressive baseline, the importance of self-attention, and the importance of the accuracy of the duration model.