Upload README.md with huggingface_hub

README.md

@@ -15,5 +15,43 @@ license: cc-by-4.0
 # GLaDOS (Portal 2) — English (en)
 
 LJSpeech dataset of GLaDOS (Portal 2) (en).
-- **331** pairs
+- **982** pairs
 - 16kHz mono 16-bit PCM WAV
+
+## Piper TTS Training (High Quality on T4 GPU)
+
+This dataset is ready for Piper TTS training. Recommended parameters:
+
+### Preprocessing
+```sh
+python3 -m piper_train.preprocess \
+  --language en-us \
+  --input-dir ./glados_p2_en \
+  --output-dir ./train_glados_p2_en \
+  --dataset-format ljspeech \
+  --single-speaker \
+  --sample-rate 22050
+```
+
+### Training (Kaggle T4 16GB)
+```sh
+python3 -m piper_train \
+  --dataset-dir ./train_glados_p2_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
+```
+
+- **Quality**: `high` (22,050 Hz, larger voice model, better quality)
+- **Batch size**: 16 fits T4 16GB VRAM at high quality
+- **Precision**: 32 (full precision, T4 backward pass is slow with 16-mixed for VITS)
+- **Fine-tune**: from an existing Piper checkpoint (same quality/rate)
+- **Max phoneme ids**: 400 filters very long sentences to fit VRAM