Robust recognition methods

The performance of currently available recognition systems is good under well-controlled conditions of use. This performance is highly dependent on the complexity and difficulty of the task in hand. Thus, on average, error rates measured in the laboratory, in speaker-independent mode, can range from 0.3% (for sequences of numbers) to 5% (for a vocabulary of 20,000 words in continuous speech), then to 8% (for spelled-out letters), and up to 55% for spontaneous telephone conversations!

Furthermore, the error rate increases dramatically when the conditions under which a system is trained and used are different (particularly with regard to noise type and level).