The Sillwood Technologies System for the VoiceMOS Challenge 2022

TL;DR

This paper presents a system for the VoiceMOS Challenge 2022 that fine-tunes pre-trained models with stochastic weight averaging and influence functions to improve voice quality assessment, achieving top-10 rankings.

Contribution

The novel approach combines fine-tuning, stochastic weight averaging, and influence functions to enhance generalization and data quality in VoiceMOS systems.

Findings

Ranked 5th in main track

Ranked 7th in out-of-domain track

Improved performance through data quality filtering

Abstract

In this paper we describe our entry for the VoiceMOS Challenge 2022 for both the main and out-of-domain (OOD) track of the competition. Our system is based on finetuning pre-trained self-supervised waveform prediction models, while improving its generalisation ability through stochastic weight averaging. Further, we use influence functions to identity possible low-quality data within the training set to further increase our model's performance for the OOD track. Our system ranked 5th and joint 7th for the main track and OOD track, respectively.