@lissyx Does deepspeech 0.6 alpha versions support ARM NNAPI and support delegates for GPUs like Mali T-880? Have you or someone done any bench marking on performance improvements in terms of inference times w.r.t using ARM CPU Vs using these delegates which make use of the GPU?
I have local patches that does enable those feature, but as much as we could test:
- NNAPI gives no gain
- GPU delegation gives no gain
- Both are not happy with our current model, and the limitations are pretty hard
Some measures: https://github.com/mozilla/DeepSpeech/issues/2270
Branch (quite old): https://github.com/mozilla/DeepSpeech/compare/master...lissyx:tflite-delegates?expand=1