Model v01 Tacotron2, DDC PWGAN

Thorsten Müller edited this page 2020-09-15 06:22:29 +02:00

Table of Contents

• Release Candidate 1 - "early bird"
• Content of this rc
• Audio samples
• How to use
• Known issues
• Release Candidate 2 - to come

Based on "thorsten" voice dataset v02 (normalized)

Release Candidate 1 - "early bird"

This early bird rc was published on september, 14th 2020 and is meant for first experiments. All files (configs and checkpoints on google drive.

Content of this rc

• trained on dataset v02 (normalized)
• trained on mozilla tts repo (commit d4319fe)
• tacotron2 + ddc training for 360k steps (thanks to Olaf)
• pwgan vocoder model (2,75 mio. steps) thanks to erogol
• provide audio samples
• jupyter notebook for testing
• usage of german phoneme cleaner (thanks to repodiac)
• usage of english loan words
• use compute_statistics output in vocoder training
• tbd

Audio samples

• soundcloud link

How to use

• Link to repodiacs repos (phoneme cleaner and dockerfile)
• Link to ojthiele notebook

Known issues

• Some problems with english words in german sentences
• Sometimes last character in a word is dismissed
• Output sounds little metallic

Release Candidate 2 - to come