Please see previous post:
Essentially what’s happening with this bug, after much investigation, is that the libdeepspeech.so file is compiled with certain instructions (AVX) which only newer Intel/AMD chips support. Older CPU’s lacking this instruction set can’t load the libdeepspeech.so file and throw an exception, which manifests as this import error.
I’ve seen a lot of posts of frustrated people trying to get past this bug, and the recommendations is that the libdeepspeech.so file is not being found correctly. However, I verified that in 0.9.3 this isn’t the case: libdeepspeech.so is found correctly, it just can’t load and initialize.
The solution to this should be that the loading of the _impl file be wrapped up in a try-except block and then a helpful error message should be printed explaining that the failure is likely due to an old CPU/Missing instruction set. However, looking at the _impl.py file it’s clear this is autogenerated by swig, and the entry point from the swig file to put in this try-catch block is unclear.
However, if the DeepSpeech team puts in the investment of doing a proper error early on in the launch process then the results will be less bug reports like this. In the future it would be awesome if the right binaries can be swapped in depending on system state. But from what I understand this problem is coming from the TensorFlow so it might be harder to do.