Unose-Voice-v0.5: Empirical Neural Voice Clone

Unose-Voice-v0.5 is the foundational empirical checkpoint for Patrick's neural voice clone, trained under the Igunosa Labs computational systems research initiative.

Model Summary

Architecture: VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) + HiFi-GAN Vocoder.
Export Format: Optimized ONNX Runtime graph (60.5 MB).
Sampling Rate: 22,050 Hz.
Speaker: Patrick (Igunosa Labs).
Accent: English (Nigerian / West African inflection & cadence).
Training Schedule: 250 adaptation epochs from en_US-lessac-medium base checkpoint to Epoch 2414.

Empirical Metrics (Final Convergence)

Generator Loss ($L_{gen}$): 35.37
Discriminator Loss ($L_{disc}$): 1.86 (steady adversarial equilibrium)
Validation Loss ($L_{val}$): 38.58
Inference Latency: ~25ms Real-Time Factor (RTF 0.04) on ARM64 CPU.

Usage with Piper TTS

echo "Hello, welcome to Igunosa Labs." | piper \
  --model Unose-Voice-v0.5.onnx \
  --config Unose-Voice-v0.5.onnx.json \
  --output_file output.wav

Lab & Research

Developed at Igunosa Labs as part of the Agentic Experience Thesis: Applied Systems Research.

GitHub: https://github.com/Igunosa-Labs/unose-voice

Downloads last month: -