RealtimePPGVC

Voice conversion model for real-time synthesis using PPG (Phonetic PosteriorGram) as an intermediate feature, written in Pytorch.

Implementation details

Reference

• Joint Adversarial Training of Speech Recognition and Synthesis Models for Many-to-One Voice Conversion Using Phonetic Posteriorgrams

Dataset

• CSJ Corpus (for training PPG)
• JVS Corpus (for target speech)

ASR training result

ASR training sample

Transcript: "二階から (n i k a i k a r a) ..."

The correspondence between index and phone is described here.

VC speech sample

Baseline samples (No GAN, No DAT)

https://drive.google.com/drive/folders/1Djq4dwZgJdGy4rFVArZY_kLySoxu9iSj?usp=sharing

• gen_[ID].wav: generated speech
• ref_[ID].wav: source speech
• jsut_target.wav: speech from target speaker