Upload README.md with huggingface_hub

README.md

@@ -16,4 +16,34 @@ license: cc-by-4.0
 
 LJSpeech dataset of Wheatley (en).
 - **827** pairs
-- 16kHz mono 16-bit PCM WAV
+- 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
+```