Introduction - Methods for Mining and Summarizing Text Conversations - page 5

conversation types such as blogs, instant messaging and discussion forums are becoming widely
popular.

100

90

80

70

60

50

40

30

20

10

0

Figure 1.2: Popularity of various online conversational activities.

1.1.2 THE SPEECH TECHNOLOGY REVOLUTION
The rise of text conversations is not only due to the phenomenal adoption of novel Web-based
social media; an extremely rapid progress in speech technology is also playing a role. The time is
ripe for systems that can automatically transcribe meetings, phone conversations and other spoken
interactions. In this topic, when we discuss summarization of spoken conversations, we assume the
presence of a transcript, which can be either manually written or the output of an automatic speech
recognition (ASR) system. That is, we do not consider speech summarizers that work directly and ASR

solely off of the speech signal [ Penn and Zhu , Forthcoming ]. While some speech corpora contain
manual transcripts of meetings, lectures or phone conversations, it is more realistic to expect that
a deployed speech summarizer would be running on ASR output. And it is owing to the huge
advances in ASR that speech summarization has become a feasible and popular area of research in
recent years.
A survey of ASR advances is far beyond the scope of this topic; Jurafsky and Martin [ 2008 ]
give a clear, accessible overview of modern techniques and historical trends. The biggest advance-
ment began in the 1970s and was popularized in the 1980s with work on statistical speech recogni-
tion approaches using Hidden Markov Models (HMMs). Statistical HMM-based systems remain