From ef9b68b5105fc26d7a281a6bbcdfc3e6d0d5bbfe Mon Sep 17 00:00:00 2001
From: Dongji Gao
Date: Mon, 25 Sep 2023 14:36:18 -0400
Subject: [PATCH] Update README.md by adding explanation of arc penalties
---
egs/librispeech/WSASR/README.md | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/egs/librispeech/WSASR/README.md b/egs/librispeech/WSASR/README.md
index 4fb5f52de..b7cbae304 100644
--- a/egs/librispeech/WSASR/README.md
+++ b/egs/librispeech/WSASR/README.md
@@ -33,6 +33,13 @@ We modify $G(\mathbf{y})$ by adding self-loop arcs into each state and bypass ar
+We incorporate the penalty strategy and apply different configurations for the self-loop arc and bypass arc. The penalties are set as
+
+ $\lambda_{1_{i}} = \beta_{1} * \tau_{1}^{i},\quad \lambda_{2_{i}} = \beta_{2} * \tau_{2}^{i}$
+
+for the $i$-th training epoch. $\beta$ is the initial penalty that encourages the model to rely more on the given transcript at the start of training.
+It decays exponentially by a factor of $\tau \in (0, 1)$, gradually encouraging the model to align speech with $\star$ when getting confused.
+
After composing the modified WFST $G_{\text{otc}}(\mathbf{y})$ with $L$ and $T$, the OTC training graph is shown in this figure: