crazygiscool/ahsoka-piper-voice

Ahsoka Tano Voice (Piper TTS)

A fine-tuned Piper TTS model that clones the voice of Ahsoka Tano from Star Wars: The Clone Wars.

Files

File	Description
ahsoka-final.onnx	ONNX model for inference
ahsoka-final.json	Model config (sample rate, phoneme map, inference params)
ahsoka-final.ckpt	PyTorch Lightning checkpoint (for continued training)
dataset.jsonl	Training dataset (1019 samples)
dataset.jsonl.gz	Compressed dataset

Quick Start (Piper CLI)

Install Piper

# Clone and build from source
git clone https://github.com/rhasspy/piper.git
cd piper
./build.sh

Or use a pre-built binary from Piper Releases.

Download Model

mkdir -p ~/.local/share/piper-voices/ahsoka
cd ~/.local/share/piper-voices/ahsoka

# Download from HuggingFace
wget https://huggingface.co/crazygiscool/ahsoka-piper-voice/resolve/main/ahsoka-final.onnx
wget https://huggingface.co/crazygiscool/ahsoka-piper-voice/resolve/main/ahsoka-final.json

Synthesize Speech

# Pipe text in
echo "Hello, I am Ahsoka Tano" | piper --model ~/.local/share/piper-voices/ahsoka/ahsoka-final.onnx --output_file output.wav

# Or specify text directly
piper --model ~/.local/share/piper-voices/ahsoka/ahsoka-final.onnx --text "The force is strong with this one" --output_file output.wav

Python Usage

Install Dependencies

pip install piper-tts piper-phonemize torch soundfile

Synthesize with ONNX Model

import soundfile as sf
import piper_phonemize
import onnxruntime as ort
import numpy as np

# Load model
session = ort.InferenceSession("ahsoka-final.onnx")

# Phonemize text
text = "I am Ahsoka Tano"
phonemes = piper_phonemize.phonemize_espeak(text, "en-us")
phoneme_ids = piper_phonemize.phoneme_ids_espeak([p for seq in phonemes for p in seq])

# Add padding (required format: [BOS, 0, phoneme, 0, phoneme, ..., EOS])
ids = [0] + [p for pid in phoneme_ids for p in [0, pid]] + [0]

# Inference
x = np.array([ids], dtype=np.int64)
x_len = np.array([len(ids)], dtype=np.int64)
scales = np.array([0.667, 1.0, 0.8], dtype=np.float32)

audio = session.run(None, {
    "input": x,
    "input_lengths": x_len,
    "noise_scale": scales[:1],
    "length_scale": scales[1:2],
    "noise_w": scales[2:],
})[0]

# Save
sf.write("output.wav", audio[0, 0], 22050)

Synthesize with PyTorch Checkpoint (for development)

import torch
from piper_train.vits.lightning import VitsModel

model = VitsModel.load_from_checkpoint("ahsoka-final.ckpt", dataset=None)
model_g = model.model_g
model_g.eval()

with torch.no_grad():
    model_g.dec.remove_weight_norm()
    # (same phonemize + inference code as above)

Inference Parameters

Parameter	Default	Range	Description
noise_scale	0.667	0.0 - 1.0	Phoneme noise (variation)
length_scale	1.0	0.5 - 2.0	Speaking speed (lower = faster)
noise_w	0.8	0.0 - 1.0	Phoneme width noise

Training Details

• Base model: Piper English (en-us)
• Training data: 1019 audio samples (22050 Hz, mono, 16-bit PCM WAV)
• Fine-tuning LR: 1e-5 (20x lower than default for gentle adaptation)
• Epochs trained: ~7000
• Hardware: CPU only

Dataset Format

Each line in dataset.jsonl contains:

{"text": "Transcription text", "audio_filename": "relative/path/to/audio.wav"}

All audio must be: 22050 Hz, mono, 16-bit PCM WAV.

License

Apache 2.0