piper-train-docker

Image for training a Piper TTS voice easily.

This only supports single-gpu training currently.

Quickstart

See inside the Dockerfile for all environmental variables

Generate dataset

See the dataset format here.

TL;DR:

There is a metadata.csv at the root. It looks like this:

id | text

Where id.wav exists in the wav subfolder.

With Hugging Face

You will need a Hugging Face token and a Repo id. They use the environmental variables HF_TOKEN and HF_DATASET respectively

Without Hugging Face

You will need to mount your dataset folder to /dataset within the container

Run training

Base Checkpoint

By default, the base checkpoint is the high quality "lessac" voice. You can set the checkpoint with the CHECKPOINT variable

Using pretrained from Hugging Face

Use a file name from the checkpoint repo

For example: en/en_US/lessac/high/epoch=2218-step=838782.ckpt

Custom checkpoint

If you want to use a predefined checkpoint, you'll need to mount it in /base_checkpoints and set the CHECKPOINT variable to the checkpoint's filename

For example: lessac.ckpt (assuming /base_checkpoints/lessac.ckpt is mounted)

Results

The files end up in the /cache directory. You should mount it if you want to persist the files.

Tensorboard will be started and running on port 6006. You will need to forward it if you want to use the interface

Example command

(See 'Generate Dataset' section first)

With Hugging Face dataset

docker run --gpus all -p 6006:6006 -e HF_TOKEN=... -e HF_DATASET=... -e CUDA_VISIBLE_DEVICES=0 -v ./cache:/cache ifansnek/piper-train-docker:latest

With local dataset

docker run --gpus all -p 6006:6006 -e CUDA_VISIBLE_DEVICES=0 -v ./dataset/dataset -v ./cache:/cache ifansnek/piper-train-docker:latest

Note:

Sometimes in WSL you will need to use something like --gpus all --shm-size 32 --ipc=host for proper CUDA use if you run out of shared memory.

Additionally, you may have to use absolute bind mount paths on Windows.