SegINR: Segment-wise Implicit Neural Representation for Sequence Alignment in Neural Text-to-Speech

TL;DR

SegINR introduces a segment-wise implicit neural representation for sequence alignment in neural TTS, eliminating the need for duration predictors and complex frame-level modeling, leading to improved speech quality and efficiency.

Contribution

It proposes a novel segment-wise INR method that models temporal dynamics and defines segment boundaries automatically, simplifying neural TTS pipeline.

Findings

Outperforms conventional methods in zero-shot adaptive TTS

Achieves higher speech quality with lower computational costs

Effectively models temporal dynamics within speech segments

Abstract

We present SegINR, a novel approach to neural Text-to-Speech (TTS) that addresses sequence alignment without relying on an auxiliary duration predictor and complex autoregressive (AR) or non-autoregressive (NAR) frame-level sequence modeling. SegINR simplifies the process by converting text sequences directly into frame-level features. It leverages an optimal text encoder to extract embeddings, transforming each into a segment of frame-level features using a conditional implicit neural representation (INR). This method, named segment-wise INR (SegINR), models temporal dynamics within each segment and autonomously defines segment boundaries, reducing computational costs. We integrate SegINR into a two-stage TTS framework, using it for semantic token prediction. Our experiments in zero-shot adaptive TTS scenarios demonstrate that SegINR outperforms conventional methods in speech quality with computational efficiency.