| 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) | |
| ```sh | |
| 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) | |
| ```sh | |
| 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 | |
| ``` | |