From 603662f152cead6cfc65ee3f7c085262bf9614e4 Mon Sep 17 00:00:00 2001 From: Jingyong Hou Date: Thu, 11 Nov 2021 09:55:29 +0800 Subject: [PATCH] add a simple intro to MDTC --- kws/model/mdtc.py | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/kws/model/mdtc.py b/kws/model/mdtc.py index 4ba1e75..db26493 100644 --- a/kws/model/mdtc.py +++ b/kws/model/mdtc.py @@ -167,6 +167,15 @@ class TCNStack(nn.Module): class MDTC(nn.Module): + """Multi-scale Depthwise Temporal Convolution (MDTC). + In MDTC, stacked depthwise one-dimensional (1-D) convolution with + dilated connections is adopted to efficiently model long-range + dependency of speech. With a large receptive field while + keeping a small number of model parameters, the structure + can model temporal context of speech effectively. It aslo + extracts multi-scale features from different hidden layers + of MDTC with different receptive fields. + """ def __init__( self, stack_num: int,