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
refactor(lib/layers): PLDA layer input can be either 2D or 3D tensor #10
Refactored the PLDA layer to expect either a 2D or a 3D tensor
in the shape (batch, dim) or (batch, 1, dim) respectively. The
latter is the default output shape of TDNN / StatsPooling layers
used in x-vector models. The former is sometimes preferred since
the latter contains a redundant axis which was the frame axis
before the x-vector was computed.
Added a reader for reading kaldi arrays (matrices or vectors)
in both text or binary formats
Refactored the
PLDA
layer to expect either a 2D or a 3D tensor in the shape (batch, dim) or (batch, 1, dim) respectively. The latter is the default output shape of TDNN / StatsPooling layers used in x-vector models. The former is sometimes preferred since the latter contains a redundant axis which was the frame axis before the x-vector was computed.