Fuzzy Expectation Maximization Phoneme Prediction in Diffusion Model-based Dysarthria Voice Conversion

In this paper proposed an effective fuzzy expectation-maximization phoneme prediction method in diffusion model-based dysarthria voice conversion (FEMPPDM-DVC) which is accessible to (i) training without parallel data (ii) converting a longer duration of the audio data (iii) preserves speaker identity. By integrating Fuzzy Expectation-Maximization (FEM) clustering and Diffusion model-based dysarthria voice conversion approach, the proposed method is able gradually generate normal utterances. Feature extraction is performed using Mel-frequency cepstral coefficients (MFCCs), a robust method in acoustic feature extraction in Text-to-Speech systems. Ensures the effectiveness and accuracy, through repeated parameter adjustment using the FEM clustering algorithm. The conversion network is a diffusion model-based structure, which can convert normal utterances to dysarthria utterances in the forward diffusion process. After forward diffusion process, the dysarthria utterance will go through a reverse diffusion process to convert dysarthria voice to normal utterance. Once converted, a GAN-based vocoder is applied to convert mel-frequency spectrogram to waveform. Objective evaluation is conducted on the Saarbrücken Voice Database (SVD) dataset show that FEMDDPM-DVC is able to improve the intelligibility and naturalness of dysarthria utterances in 15 kinds of dysarthria voice. The proposed method is compared with five other dysarthria voice conversion methods, show that the proposed method has the most performance among other method.