Tasks

Overview

The articubench benchmark provides three distinct methods for initializing the PAULE control model and generating Control Parameters (CPs).

Task Types

Acoustic-only (Copy-Synthesis)

Input: Target audio recording (human speech) and sample rate

Purpose:

• Test model’s ability to mimic human speech acoustics

• Focus on articulatory and acoustic quality

• No semantic information provided

Semantic-only

Input: Target semantic embedding vector, desired duration

Purpose:

• Generate speech from meaning alone

• Test semantic-to-articulation mapping

• Handle one-to-many relationship (multiple valid pronunciations)

• Also known as “full generation task”

Semantic-Acoustic

Input: Target audio recording with sample rate AND semantic embedding vector with a desired duration

Purpose:

• Joint optimization of acoustics and meaning

• Most complete evaluation scenario

• Balance multiple constraints