Attention-Constrained Inference for Robust Decoder-Only Text-to-Speech

TL;DR

A novel inference method without altering the training process, named Attention-Constrained Inference (ACI), is proposed to facilitate monotonic synthesis of text-to-speech models.

Abstract

Recent popular decoder-only text-to-speech models are known for their ability of generating natural-sounding speech. However, such models sometimes suffer from word skipping and repeating due to the lack of explicit monotonic alignment constraints. In this paper, we notice from the attention maps that some particular attention heads of the decoder-only model indicate the alignments between speech and text. We call the attention maps of those heads Alignment-Emerged Attention Maps (AEAMs). Based on this discovery, we propose a novel inference method without altering the training process, named Attention-Constrained Inference (ACI), to facilitate monotonic synthesis. It first identifies AEAMs using the Attention Sweeping algorithm and then applies constraining masks on AEAMs. Our experimental results on decoder-only TTS model VALL-E show that the WER of synthesized speech is reduced by up to 20.5% relatively with ACI while the naturalness and speaker similarity are comparable.

Figures (4)

• Figure 1: The decoder-only architecture model contains various attention modules with different functionalities that are not always diagonal. We selected a sentence for inference and obtained the attention maps for each head of every layer. In the figure, one head from each decoder layer is randomly selected for illustration.
• Figure 2: Illustration of real attention alignments. (a), (c) and (d) are attention heatmaps from the same AEAM of the same model, with different inputs. (a) is for a text-speech pair directly picked from the test set. With the same inputs, (b) shows the reference alignment produced by the Viterbi algorithm. (c) and (d) illustrate the AEAM when some text tokens are repeated or skipped in synthesis.
• Figure 3: An example of AEAM attention values and DP values, and corresponding CMasks. Here $\bm{x}$ is the input text tokens, and $\bm{y}$ is the speech tokens. (a) presents an example of the attention value pattern of a normalized AEAM when generating $y_{t+1}$, with the maximum value of each row highlighted. The values are normalized so that each row sums up to $1$. The constraining mask (CMask) for the newly generated token $y_t$ is now needed. (b) shows the CMask obtained via the trivial argmax strategy, which is prone to causing instability, such as the attention center being falling back. Based on attention values in (a), subfigure (c) illustrates the DP values $d$ of each state. (d) is the CMask obtained via the DP strategy, which shows stability in maintaining the monotonicity.
• Figure 4: Calculate $\mathcal{C}_\mathrm{E}$ and $\mathcal{C}_\mathrm{A}$ for each attention map of 4 models, and plot them as 2-D points. The separation line is set at $\tau=1$.