Evaluation of Spoken Language Recognition Technology Using Broadcast Speech: Performance and Challenges
| Presented by: |
| ||
|---|---|---|---|
| Author(s): |
| ||
Spoken Language Recognition (SLR) technology has remarkably improved in the last years, partly thanks to NIST Language Recognition Evaluations (LRE), which have become standard benchmarks for testing new approaches. NIST evaluations focus on narrow-band conversational telephone speech and deal with some specific target languages. Recent efforts to expand the scope of SLR technology assessment include the Albayzin 2008 and 2010 LRE, which deal with wide-band TV broadcast
speech. In this work, a SLR system based on state-of-the-art approaches is developed and evaluated on the Albayzin 2008 and 2010 LRE datasets, looking to identify those conditions that make the task challenging and eventually to guide the design of future evaluations using the same kind of data. We present and analyse system performance under different conditions, regarding: (1) the set of target languages (including details about the confusion of languages with each other) and (2) the amount of data available to estimate models; and (3) the presence of background noise.
