A Hybrid Voice Cloning for Inclusive Education in Low-Resource Environments

Muhammad Mohtad Younus¹Arshad Iqbal^1*Esha e Noor Durrani¹Naveed Ahmad²Mohamad Ladan²

• ¹Pak-Austria Fachhochschule Institute of Applied Sciences and Technology, Haripur, Pakistan
• ²Prince Sultan University, Riyadh, Saudi Arabia

Voice cloning is the process of replicating a target speaker's vocal characteristics to generate synthesized speech that closely mimics the original. Modern voice cloning systems can now produce natural, expressive speech from limited data. A Generative Adversarial Network (GAN) and a Variational Autoencoder (VAE) based voice cloning system is proposed to enhance voice cloning characteristics using a Generalized End-to-End (GE2E) speaker encoder alongside a text-to-spectrogram synthesizer built from Tacotron technology, combined with WaveRNN-based model modifications through the addition of gating layers. It is possible to efficiently transform brief utterances into customized high-quality speech synthesis while achieving near real-time performance through the proposed approach. The proposed pipeline undergoes evaluation using three datasets, including LibriSpeech, VCTK, and local corpora, which exhibit better spectral fidelity as measured by Mel Cepstral Distortion (MCD) and subjective quality as assessed by Mean Opinion Score (MOS). This work also emphasizes the relevance of hybrid voice cloning for inclusive education in low-resource environments. In addition, it highlights responsible deployment approaches by discussing ethical implications and proposing watermarking, consent-based data collection, and automatic detection measures. The system demonstrates robust functionality when working across different languages and speaker conditions, making it suitable for assistive technology, educational applications, media production, and interactive conversational agents.