Avoiding file I/O for manifests

[WilliamVenner]

Is it possible to use NeMo models (specifically speaker recognition) without having to set up manifest files?

I don't see why we can't set up these manifests in-memory or using code. If I already have the data in memory, why should I need to write said data to a file in order to perform inference?

https://github.com/NVIDIA/NeMo/blob/stable/examples/speaker_tasks/recognition/speaker_identification_infer.py#L38-L39