A comparative study of generative models for child voice conversion

2025-12-12

AI Summary

This study addresses the challenge of inadequate acoustic modeling of child voice characteristics in adult-to-child voice conversion (VC). It presents the first systematic evaluation of four generative modeling paradigms—diffusion models, normalizing flows, variational autoencoders (VAEs), and generative adversarial networks (GANs)—for this task. To bridge the acoustic gap between adult and child speech in fundamental frequency (F0) and formant distributions, we propose a lightweight frequency-domain pitch-and-formant warping post-processing technique. Evaluated on a newly constructed child dubbing corpus via combined objective and subjective metrics, all models benefit significantly from the proposed post-processing: average Mel-cepstral distortion (MCD) decreases by 38%, mean opinion score (MOS) for naturalness improves by 2.1 points, and target speaker similarity increases markedly. This work establishes a reproducible technical pipeline and benchmark for generative VC in child speech synthesis.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Generative Adversarial Networks (GANs) for Vision

Application Category

• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Generative models are a popular choice for adult-to-adult voice conversion (VC) because of their efficient way of modelling unlabelled data. To this point their usefulness in producing children speech and in particular adult to child VC has not been investigated. For adult to child VC, four generative models are compared: diffusion model, flow based model, variational autoencoders, and generative adversarial network. Results show that although converted speech outputs produce by those models appear plausible, they exhibit insufficient similarity with the target speaker characteristics. We introduce an efficient frequency warping technique that can be applied to the output of models, and which shows significant reduction of the mismatch between adult and child. The output of all the models are evaluated using both objective and subjective measures. In particular we compare specific speaker pairing using a unique corpus collected for dubbing of children speech.

Problem

• Investigates adult-to-child voice conversion using generative models
• Compares four generative models for converting adult to child speech
• Introduces frequency warping to improve similarity to target child voice

Innovation

• Frequency warping technique reduces adult-child mismatch
• Four generative models compared for child voice conversion
• Objective and subjective evaluation of converted speech outputs