ONNX with spk_emb input (any blend/style at run time); voices.json = embedding tables + trio preset; Kopa blends removed

v3: Polish Matcha-TTS + fine-tuned HiFi-GAN vocoder, ONNX with baked voices