You are here

Charles Fox research description

Large corpora of transcribed speech are rare and expensive to acquire, yet are of great use for developing and improving ASR systems. Of particular research interest are corpora of natural speech, such as far-field recordings of multiple speakers in noisy environments. The ESDS* database contains many thousands of hours of such recordings, but made for non-ASR purposes. We are working with one examplar corpus from this database, called Family Life to make it useable in ASR research.  In particular, the transcriptions have no timing annotations, and many of the audio files are mislabelled. Family Life is one of many corpora in ESDS and if data cleansing is possible on it, then a potentially large collection of natural speech corpora could become available from ESDS.

* ESDS = Economic and social data service