Repeat after me: Self-supervised learning of acoustic-to-articulatory mapping by vocal imitation

TL;DR

This paper introduces a self-supervised neural model that learns to map speech acoustics to articulatory movements by vocal imitation, combining a synthesizer, forward, and inverse models trained on raw speech data.

Contribution

It presents a novel self-supervised framework for acoustic-to-articulatory mapping using a combination of neural models trained jointly from raw speech data.

Findings

Encouraging imitation performance in simulations

Effective joint training of forward and inverse models

Applicable to multiple speakers

Abstract

We propose a computational model of speech production combining a pre-trained neural articulatory synthesizer able to reproduce complex speech stimuli from a limited set of interpretable articulatory parameters, a DNN-based internal forward model predicting the sensory consequences of articulatory commands, and an internal inverse model based on a recurrent neural network recovering articulatory commands from the acoustic speech input. Both forward and inverse models are jointly trained in a self-supervised way from raw acoustic-only speech data from different speakers. The imitation simulations are evaluated objectively and subjectively and display quite encouraging performances.