<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2021">
  <Title>Initial Study on Automatic Identification of Speaker Role in Broadcast News Speech</Title>
  <Section position="4" start_page="0" end_page="81" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> The most related previous work is (Barzilay et al., 2000), in which Barzilay et al. used BoosTexter and the maximum entropy model to classify each speaker's role in an English broadcast news corpus. Three classes are used, anchor, journalist, and guest speaker, which are very similar to the role categories in our study. Lexical features (key words), context features, duration, and explicit speaker introduction are used as features. For the three-way classification task, they reported accuracy of about 80% compared to the chance of 35%. They have investigated using both the reference transcripts and speech recognition output. Our study differs from theirs in that we use one generative modeling approach (HMM), as well as the conditional maximum entropy method. We also evaluate the contextual role information for classification. In addition, our experiments are conducted using a different language, Mandarin broadcast news. There may be inherent difference across languages and news sources.</Paragraph>
    <Paragraph position="1"> Another task related to our study is anchor segmentation. Huang et al. (Huang et al., 1999) used a recognition model for a particular anchor and a background model to identify anchor segments. They reported very promising results for the task of determining whether 2Even though this is a baseline (or chance performance), it is not very meaningful since there is no information provided in this output.</Paragraph>
    <Paragraph position="2">  or not a particular anchor is talking. However, this method is not generalizable to multiple anchors, nor is it to reporters or other guest speakers. Speaker role detection is also related to speaker segmentation and clustering (also called speaker diarization), which was a benchmark test in the NIST Rich Transcription evaluations in the past few years (for example, NIST RT-04F http://www.nist.gov/speech/tests/rt/rt2004/fall/). Most of the speaker diarization systems only use acoustic information; however, in recent studies textual sources have also been utilized to help improve speaker clustering results, such as (Canseco et al., 2005). The goal of speaker diarization is to identify speaker change and group the same speakers together. It is different from our task since we determine the role of a speaker rather than speaker identity. In this initial study, instead of using automatic speaker segmentation and clustering results, we use the manual speaker segments but without any speaker identity information.</Paragraph>
  </Section>
class="xml-element"></Paper>