Pretrained models

Pretrained models come as an archive that contains all three models (speaker encoder, synthesizer, vocoder). The archive comes with the same directory structure as the repo, and you're expected to merge its contents with the root of the repository.

Initial commit (latest release) [Google drive] [Dropbox]

Please ensure the files are extracted to these locations within your local copy of the repository:

encoder\saved_models\pretrained.pt
synthesizer\saved_models\logs-pretrained\taco_pretrained\checkpoint
synthesizer\saved_models\logs-pretrained\taco_pretrained\tacotron_model.ckpt-278000.data-00000-of-00001
synthesizer\saved_models\logs-pretrained\taco_pretrained\tacotron_model.ckpt-278000.index
synthesizer\saved_models\logs-pretrained\taco_pretrained\tacotron_model.ckpt-278000.meta
vocoder\saved_models\pretrained\pretrained.pt

If desired, the below checksums can be used to verify the integrity of pretrained.zip:

md5: 30459b5c0904110dc6b1447c28095554
sha1: 8e344861a76c52f9b73929a64357dc1776b640cb
sha256: 9dee6d64b2d9d1dd45b5d2d5c94ee3412e386d4b3378724a805e5e9418c7d255

Here is some information about the models. For reference, the GPUs used for training are GTX 1080 Ti.

• Encoder: trained 1.56M steps (20 days with a single GPU) with a batch size of 64
• Synthesizer: trained 278k steps (1 week with 4 GPUs) with a batch size of 144
• Vocoder: trained 428k steps (4 days with a single GPU) with a batch size of 100