metadata
configs:
- config_name: default
data_files:
- split: train
path: data.zip
tags:
- audio
- voice
- portal
- piper-tts
- ljspeech
license: cc-by-4.0
Wheatley — English (en)
LJSpeech dataset of Wheatley (en).
- 827 pairs
- 44.1kHz mono 16-bit PCM WAV (original wiki quality)
Piper TTS Training (High Quality on T4 GPU)
Preprocessing (downsample to 22.05kHz)
python3 -m piper_train.preprocess \
--language en-us \
--input-dir ./wheatley_en \
--output-dir ./train_wheatley_en \
--dataset-format ljspeech \
--single-speaker \
--sample-rate 22050
Training (Kaggle T4 16GB)
python3 -m piper_train \
--dataset-dir ./train_wheatley_en \
--accelerator 'gpu' \
--devices 1 \
--quality high \
--batch-size 16 \
--validation-split 0.0 \
--num-test-examples 0 \
--max_epochs 10000 \
--resume_from_checkpoint /path/to/checkpoint.ckpt \
--checkpoint-epochs 1 \
--precision 32 \
--max-phoneme-ids 400