Aletheia/tools/tts/GPU_TRAINING.md

Aletheia Russian Piper/VITS training

The selected deployment format is a single-speaker Piper/VITS model exported to ONNX. Piper's current official training interface consumes wav-file|text CSV rows, supports Russian through the ru espeak-ng voice, and exports a checkpoint with python3 -m piper.train.export_onnx.

GPU host prerequisites

• Windows or Linux with an NVIDIA CUDA GPU and a working nvidia-smi.
• Git, Python 3, build-essential, CMake, and Ninja.
• A checkout of https://github.com/OHF-Voice/piper1-gpl with the [train] dependencies installed and build_monotonic_align.sh completed.
• The generated Aletheia dataset directory containing metadata.csv and wav/.

Scratch training

git clone https://github.com/OHF-Voice/piper1-gpl.git
cd piper1-gpl
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -e '.[train]'
./build_monotonic_align.sh
python3 setup.py build_ext --inplace

/path/to/Aletheia/tools/tts/train_piper.sh \
  /path/to/tts-dataset \
  /path/to/tts-training \
  "$PWD" \
  2000

No --ckpt_path is passed: the Aletheia acoustic model is initialized from scratch. This is intentionally slower than Piper's recommended checkpoint fine-tuning. The exported deliverables are aletheia_ru.onnx, aletheia_ru.onnx.json, and artifacts.json with byte counts and SHA-256 hashes.

Before accepting the model, synthesize a held-out Russian validation list at multiple length_scale, noise_scale, and noise_scale_w settings, measure real-time factor on the target phone, and listen for skipped words, unstable stress, clicks, and repeated phonemes. A successful export alone is not a quality gate.

The epoch-500 book-reading profile selected by listening comparison is stored in aletheia_book_profile.json. Keep its duration, noise, and sentence-pause values together when evaluating the ONNX model or integrating it into the reader.

On native Windows, first run the one-batch CUDA check:

.\train_piper_windows.ps1 `
  -DatasetDir C:\path\to\tts-dataset `
  -OutputDir C:\path\to\tts-smoke `
  -PythonExe C:\path\to\.venv\Scripts\python.exe `
  -SmokeTest

The default Windows batch size is 4 so the smoke test can establish actual memory use before increasing it.