next up previous contents
Next: The 1997 CU-CON Evaluation Up: Task 7.2: Technical Description Previous: Task 7.2: Technical Description

The 1997 Hub-4E Acoustic and Language Model Data

The LDC and NIST have provided acoustic and language model training data to sites that participate in the Hub-4 broadcast news benchmark tests. The acoustic training data consists of approximately 104 hours of data recorded from a variety of television and radio programmes  [1]. The acoustic data is manually segmented into homogeneous regions termed ``evaluation focus conditions''. This was done to support the 1996 ``partitioned evaluation'' (PE) paradigm [2]. These focus conditions are as follows:
F0: Baseline broadcast speech
F1: Spontaneous broadcast speech
F2: Speech over telephone channels
F3: Speech in the presence of background music
F4: Speech under degraded acoustical conditions
F5: Speech from non-native speakers

Segments that do not fall within the specification for the focus conditions presented above are labelled FX. More details of the focus conditions can be found in [3]. The focus condition labels were used to select telephone bandwidth data for acoustic model adaptation. Language model training data consists of broadcast transcripts. This covers the period from January 1992 to April 1996, and contains approximately 132 million words.


next up previous contents
Next: The 1997 CU-CON Evaluation Up: Task 7.2: Technical Description Previous: Task 7.2: Technical Description
Christophe Ris
1998-11-10