Group Project:

Language Modeling for Conversational Speech Recognition

TEAM GOALS:

1. Creating language models that are attuned to spontaneous conversations.
   "There is more to conversational speech than Dysfluencies."

   Beyond mere word-based transcription:

   • identifying and cleaning up disfluencies.
   • transforming the output into "simplified English".

GROUP TEAM:

Roni Rosenfeld (CMU) (Project Leader)
Bill Byrne (JHU)
Rukmini Iyer (BBN)
Mark Liberman (UPenn)
Liz Shriberg (SRI)
Jack Unverferth (DoD)
Enrique Vidal (Valencia-Spain)
Rajeev Agarwal (TI)
Dimitra Vergyri (student assistant)

BASELINES/PRELIMINARY ERROR ANALYSIS

Baseline WER for swbd using the standart tirgram
( N-Best results, LWER results)

"Cheating LM" Experiments

Varying the size of the training data

Disfluency-Based Error Analusis

PROJECTS

Error Analysis

Error Correcting Parsing

Modeling Disfluencies

Exploiting Remote Domains Via Data Bleaching (Data Bleaching)

Backchannel modeling

• 'Quick and Dirty'