Multilingual speech recognition at Dragon Systems

The paper reports on some Dragon Systems' experiments with multilingual large vocabulary speech recognition, both for its discrete-word product DragonDictate/sup (R/) for Windows Version 1.0 and for its speaker-independent continuous speech research systems. The experiments in discrete word recognition involve English, French, German, Italian, and Spanish. The tests show significant but not overwhelming, difference between the languages, with French being the hardest language to recognize and Italian being the easiest. The continuous speech experiments involve the Ricardo and CallHome corpora of conversational telephone speech, and show such high word error rates that no language-specific differences emerge. However, experiments with English Switchboard and CallHorne recognition indicate that improved recognition technology and larger amounts of training data can improve accuracy substantially. The authors therefore expect that future multilingual LVCSR experiments will be more illuminating.