Certain optimisations that TensorFlow recommends cause significant
performance penalties in c_speech_features. Explicitly disable these
features in the BUILD file only for c_speech_features.
The build has been restructured to produce a single deepspeech library,
rather than 3 separate deepspeech, c_speech_features and kissfft
libaries. This should make the deepspeech client and libraries
significantly easier to use and distribute.
libdeepspeech now uses the namespace DeepSpeech. The 'DeepSpeech' class
has been renamed to 'Model', 'getMfccFrames' has been renamed to
'getInputVector' and a function outside of the Model class has been added,
'audioToInputVector' to allow it to be used without initialising
TensorFlow and loading a model.
Formerly, libdeepspeech was basically a C library. Rather than returning
a context struct and having every function require this context struct,
instead provide a DeepSpeech class that handles its own destruction.
