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.
Apache License 2.0
20
stars
3
forks
source link
fix(dsp/layers): index range for `StatsPooling` when subsampling was falling short #11
The index range computed when subsampling is specified in the StatsPooling
layer (i.e. input_period or output_period > 1), was stopping one frame before
the end. So for 900 input frames, and output period of 300, we were getting
2 frames of output instead of 3.
The index range computed when subsampling is specified in the
StatsPooling
layer (i.e. input_period or output_period > 1), was stopping one frame before the end. So for 900 input frames, and output period of 300, we were getting 2 frames of output instead of 3.Also fixed a couple of typos.