---
configs:
- config_name: default
  data_files:
  - split: train
    path: data.zip
tags:
- audio
- voice
- portal
- piper-tts
- ljspeech
license: cc-by-4.0
---
# Caroline — Français (fr)

LJSpeech dataset of Caroline (fr).
- **20** 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 fr \
  --input-dir ./caroline_fr \
  --output-dir ./train_caroline_fr \
  --dataset-format ljspeech \
  --single-speaker \
  --sample-rate 22050
```

### Training (Kaggle T4 16GB)
```sh
python3 -m piper_train \
  --dataset-dir ./train_caroline_fr \
  --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
```