Piper is a fantastic tool for providing very realistic sounding TTS while also being efficient and fast. There are some publicly available pre-trained models, but if you’re here, that means you want to create your own voice. This guide aims to walk you through the steps.

This tangential issue in StackOverflow shows actually how to define an own ONNX Session outside Piper TTS, and then pass it to the PiperVoice instance. This could solve the issue with multiprocessing (can't pickle the ONNX session)
Also, explains that ONNX already manages its own threads already. Maybe it's not worth to make it work in a separate Process.

A fast and local neural text-to-speech engine that embeds espeak-ng for phonemization.
Development has moved here. Check if this is what I'm using.