Mordukhovich Subdifferential Optimization Framework for Multi-Criteria Voice Cloning of Pathological Speech

Rytis Maskeliūnas¹, Robertas Damaševičius^1,*, Audrius Kulikajevas¹, Kipras Pribuišis², Nora Ulozaitė-Stanienė², Virgilijus Uloza²
1 Center of Real Time Computer Systems, Kaunas University of Technology, Kaunas, 51423, Lithuania
2 Department of Otorhinolaryngology, Academy of Medicine, Lithuanian University of Health Sciences, Kaunas, 50161, Lithuania
* Corresponding Author: Robertas Damaševičius. Email: email
(This article belongs to the Special Issue: Artificial Intelligence Models in Healthcare: Challenges, Methods, and Applications)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2025.072790

Received 03 September 2025; Accepted 05 November 2025; Published online 02 December 2025

Download PDF

Abstract

This study introduces a novel voice cloning framework driven by Mordukhovich Subdifferential Optimization (MSO) to address the complex multi-objective challenges of pathological speech synthesis in under-resourced Lithuanian language with unique phonemes not present in most pre-trained models. Unlike existing voice synthesis models that often optimize for a single objective or are restricted to major languages, our approach explicitly balances four competing criteria: speech naturalness, speaker similarity, computational efficiency, and adaptability to pathological voice patterns. We evaluate four model configurations combining Lithuanian and English encoders, synthesizers, and vocoders. The hybrid model (English encoder, Lithuanian synthesizer, English vocoder), optimized via MSO, achieved the highest Mean Opinion Score (MOS) of 4.3 and demonstrated superior intelligibility and speaker fidelity. The results confirm that MSO enables effective navigation of trade-offs in multilingual pathological voice cloning, offering a scalable path toward high-quality voice restoration in clinical speech applications. This work represents the first integration of Mordukhovich optimization into pathological TTS, setting a new benchmark for speech synthesis under clinical and linguistic constraints.