ReVoice Results & Samples

Abstract

Total laryngectomy, a necessary procedure for advanced laryngeal cancer, results in the permanent loss of natural voice, profoundly affecting communication and quality of life for laryngectomees. Existing standard-of-care voice restoration methods face significant limitations, often requiring expensive and invasive surgical procedure (e.g., tracheoesophageal puncture, TEP), which are not suitable for all laryngectomees, or producing suboptimal speech quality (e.g., esophageal speech, ES, and electrolarynx, EL). We introduce ReVoice, a non-invasive and affordable wearable system that invents a novel AI-powered active acoustic sensing method to restore natural and intelligible voice for laryngectomees. ReVoice uses a neck-worn bone conduction actuator to gently probe the vocal tract with sound waves and capture the modulated acoustic signals representing silent articulatory movements using microphones on a neckband and eyeglass frame. These captured acoustic signals are then analyzed by an end-to-end generative AI pipeline to directly synthesize highly intelligible speech with natural temporal dynamics. We evaluated ReVoice with 26 participants, including 11 laryngectomees, which demonstrated its promising ability to generate natural and intelligible speech for continuously spoken, unseen sentences, offering significant improvements over existing solutions in terms of form factor obtrusiveness, social acceptance, speech quality and evaluation robustness. ReVoice combines active acoustic sensing with cutting-edge generative AI, demonstrating great potential for an accessible, affordable, and high-performance voice restoration solution for laryngectomees.