Now available for download: SSCS2008 Workshop Proceedings
and the SSCS2008 Workshop Report
Talks at SSCS 2008 were recorded and indexed using a combination of word-level features derived from speech recognition transcripts and presentation metadata, including the text content of the slides.
To search through the talks, visit ISLA-TV at the University of Amsterdam. Or access individual talks below.
Speech-based methods in the video
search mix:
In a large-scale, commercial application speech recognition and language technology serve to support video search. Speaker: T. Davis Watch the video
Hybrid word-subword decoding for
spoken term detection A hybrid recognition system
directly produces lattices containing both words and
subwords. Using multigram models and searching for
in-vocabulary and out-of-vocabulary terms in separate
steps makes possible performance gains on a spoken
term detection task. Speaker: M. Fapso Watch
the video
Fast Approximate Spoken Term
Detection from Sequence of Phonemes A phoneme
recognition approach to spoken term detection is used
to achieve a smaller index size and faster detection
speed. Recognizer error is compensated with a
probabilistic model based on word pronunciation and
the recognizer's phoneme confusion matrix. Speaker:
M. Fapso
Watch the video
Cluster-based Model Fusion for
Spontaneous Speech Retrieval Training: Topics
(queries) in the collection are clustered. The best
weighting scheme for combination of retrieval models
is determined for each cluster. Test topics are
classified into the topic cluster and retrieval is
performed using the corresponding weighting scheme.
Speaker: D. Inkpen
Watch the video
Combination of Multiple Speech
Transcription Methods for Vocabulary Independent
Search: Two algorithms are presented that combine
speech transcripts generated using different word and
sub-word speech recognition methods. The approach
tackles the challenge that out-of-vocabulary terms
present in a spoken term detection task. Speaker J.
Mamou
Watch the video
Advances in the Fraunhofer IAIS
Audiomining System Performance improvements in
vocabulary-independent spoken term detection are
achieved by improved acoustic models and a hybrid word
and syllable search system. In the improved system, it
is no longer necessary to accommodate recognition
error by allowing the match between query word and
syllable transcript to be inexact. Speaker: D.
Schneider
Watch the video
Using Term Clouds to Represent
Segment-Level Semantic Content of Podcasts:
Structured surrogates, which prove useful for the
semantic representation of spoken audio in the user
interface, are created automatically. TextTiling
techniques applied to speech transcripts generate
divide the audio into topical segments and each
segment is represented by a mini-term-cloud derived
from the speech transcript. Speaker M. Tsagkias
Watch the video
