Pairwise evaluation of accent similarity in speech synthesis

Motivated by the social and moral imperative for more inclusive speech technology, the community has witnessed a growing interest in systems capable of generating high-fidelity accent, across various speech generation tasks. In ZeroShot Text-to-Speech (ZS-TTS), accent hallucination/mismatch, where the generated speech deviates from the reference speech in accent, is reported and addressed in [1]. In Accented TTS, numerous attempts have been made to generate high-fidelity accent based on pre-defined accent variety labels or intensity levels [2, 3]. In Accent Conversion (AC), numerous attempts have been made to map speech from foreign to native accent, preserving content and speaker information while removing the foreign accent in the source [4]. However, how to evaluate accent similarity in speech is under-researched and lacks consensus.