mtdc speech commands testing

[budzianowski]

Hey,

thanks for the awesome work!
I want to test the default speech commands model with tdmc. After training (obtaining 97% accuracy) and running:
./build/bin/kws_main 80 10 ../../examples/speechcommand_v1/s0/model.onnx 0c40e715_nohash_0.wav

I often got random seg fault like even though the previous runs were successful:

Kws Model Info:
        cache_dim: 64
        cache_len: 244
terminate called after throwing an instance of 'std::length_error'
  what():  vector::_M_default_append

What is the suggested batch_size parameter and how to reason about it?