Judging the quality

When you decide your computer has done enough crunching, you can check the sound quality to see if it matches your voice. In piper1-gpl/lightning_logs/ open the folder with the highest version_XX-number and open the folder checkpoints. Take note of the filename of the checkpoint-file and let’s start the exporting to the voicemodel.

python -m piper.train.export_onnx ^
  --checkpoint ./lightning_logs/version_XX/checkpoints/current.ckpt ^
  --output-file ../dataset/myvoice.onnx

Replace the version_XX and checkpoint filename, save the script as e.g. export.bat in your piper1-gpl directory. This script exports to the dataset-directory, where you can also find the config.json file. Make a copy of that file and rename it to myvoice.onnx.json where myvoice is the voicename you chose earlier for the export-script.

You can test the quality with the command

python -m piper -m ../dataset/myvoice -- 'This will play on your speakers.'

Where myvoice is the voicename you chose earlier.

If you’re happy with the sound-quality, great, you can export the file to N1MM. Instructions on how to do this, are in the online N1MM-manual at Manual–Setup > Audio Setup Window.

If you’re not happy, you can let the computer refine the model by letting it crunch more epochs. For this, copy the checkpoint-file from the version_XX folder to your /dataset/ckpt-folder, change the name of the checkpoint file in the training-script to the one of this, new checkpoint, and run the script again. Rinse and repeat till your happy.

Thank you Kari, OH2XX for your initial work, Tom N1MM and the whole N1MM development-team for taking this to N1MM, and Thorsten Müller for his videos on using Piper and the team at Rhasppy for developing and maintaining Piper.