onnx export produces garbled noise

[miranda]

I followed all the instructions for the fine tuned training and got a good audio result from the test process. After stopping training and doing the commands to export the model, it is pure noise and garbled robotic vocal sounds.

Not sure what else to say. The training worked, the onnx export does not. I'm using a Ubuntu 22 container.

[miranda]

It seems the onnx export actually works but only with the python piper-tts package installed with pip. The garbled trash sounds only happen when using a piper binary... which is strange because the standard onnx models that are available for download all work with both the binary version and the python package.

[synesthesiam]

What version of PyTorch did you use to export the onnx model?

[miranda]

pytorch-lightning        1.7.7
torch                    1.13.1+rocm5.2
torchmetrics             0.11.4

[naggie]

I think I reproduced something like this using a prepackaged onnx with https://piper.ttstool.com/ -- try core [high] and get it to say air mode. It's not deterministic but a lot of time time it is a garbled mess. Does this sound similar?