A Voice Cloning Approach Utilizing Dual Attention Mechanism and Mixed Logistic Distribution

Although current voice cloning technologies have achieved significant advancements in speech clarity and naturalness, they still exhibit insufficient robustness when processing low-quality data and when operating in noisy environments, thereby limiting their application scope. To address existing challenges, this paper presents a novel framework combining a foundational TTS system with a voice conversion module. The proposed method synergistically integrates a conditional variational autoencoder (cVAE) with a Flow-based model. It features a dual attention mechanism structurally tailored for speech, refining focus on both channel and spatial dimensions of time-frequency features. Moreover, it supplants the conventional Gaussian assumption with a Mixture of Logistic (MoL) distribution, significantly enhancing the expressivity for modeling complex asymmetric and multi-modal spectrograms. Experiments conducted on the Multilingual LibriSpeech and CommonVoice datasets demonstrate superior performance over baseline models in terms of timbre similarity, speech clarity, and robustness in noisy conditions. Specifically, on Multilingual LibriSpeech, the proposed model achieves a WER of 6.50%, an MCD of 4.94 dB, and a MOS of 4.68, marking improvements of 1.92%, 0.71 dB, and 0.42 over OpenVoice, respectively.