Convert kaldi feature extraction and nnet3 models into Tensorflow Lite models. Currently aimed at converting kaldi's x-vector models and diarization pipelines to tensorflow models.
Added a StatsPooling layer that mimics Kaldi's
and ; used in x-vector models.
Implementation is a bit overkill since we really need to just compute
mean and stddev across the whole time-axis at inference time. By setting
reduce_timeaxis=True when creating the layer, all conditional branches
will be ignored making the computation graph lean.
Added unit tests and reference data generated using kaldi.
Added a
and ; used in x-vector models.
StatsPooling
layer that mimics Kaldi'sImplementation is a bit overkill since we really need to just compute mean and stddev across the whole time-axis at inference time. By setting
reduce_timeaxis=True
when creating the layer, all conditional branches will be ignored making the computation graph lean.Added unit tests and reference data generated using kaldi.