Results of Voice Cloning Experiments by blue-fish.
Real-Time-Voice-Cloning
These experiments try to reproduce the results of the paper "Transfer
Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis."
-
RTVC-1: Results with pretrained models
-
RTVC-2: Additional training for vocoder model
-
RTVC-3: Synthesizer trained on a limited dataset
-
RTVC-4: Synthesizer trained on dataset with all silence trimmed out
-
RTVC-5: Use VCTK dataset to train synthesizer and vocoder
-
RTVC-6: Tacotron 1 synthesizer (VCTK)
-
RTVC-7: Tacotron 1 synthesizer (LibriSpeech)
A better voice clone can be made by finetuning a multispeaker Tacotron
model to the target voice, which this experiment demonstrates.
-
RTVC-FT-1: Voice clones with finetuned RTVC-7 model