Upload abstract/2308.07395.txt with huggingface_hub
Browse files- abstract/2308.07395.txt +12 -0
abstract/2308.07395.txt
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Text injection for automatic speech recognition (ASR), wherein unpaired
|
2 |
+
text-only data is used to supplement paired audio-text data, has shown
|
3 |
+
promising improvements for word error rate. This study examines the use of text
|
4 |
+
injection for auxiliary tasks, which are the non-ASR tasks often performed by
|
5 |
+
an E2E model. In this work, we use joint end-to-end and internal language model
|
6 |
+
training (JEIT) as our text injection algorithm to train an ASR model which
|
7 |
+
performs two auxiliary tasks. The first is capitalization, which is a
|
8 |
+
de-normalization task. The second is turn-taking prediction, which attempts to
|
9 |
+
identify whether a user has completed their conversation turn in a digital
|
10 |
+
assistant interaction. We show results demonstrating that our text injection
|
11 |
+
method boosts capitalization performance for long-tail data, and improves
|
12 |
+
turn-taking detection recall.
|