ekwek1/soprano-factory

Built on the Soprano ultra-lightweight TTS architecture, this training framework enables custom model fine-tuning on local hardware using LJSpeech-formatted datasets with automatic 32kHz resampling. The 600-line implementation supports adding new voices and languages while maintaining the base model's efficiency (80M parameters, <1GB memory) and extreme inference speed. Trained models integrate directly with the Soprano inference pipeline across CUDA, CPU, and MPS devices on Windows, Linux, and macOS.

212 stars.