next up previous contents
Next: Examples of existing applications. Up: Overview of speech recognition Previous: ASR performance.

State-of-art in ASR error rates.

The following brief list exhibits the word error rates of representatives of some important categories [Rabiner, 1994] for vocabulary dependent recognizers. It is noteworthy that the evaluations [Rabiner, 1994] were applied in laboratory conditions and the performance in real-world conditions is usually much worse. The sentence error rate for the continuous speech tasks is much higher as well which makes these systems unsuitable for real-world applications. For large vocabulary tasks, it is customary to place rather strict restrictions to the grammar and syntax to keep the word matching procedure computationally feasible.
1.
Isolated words, speaker independent, 129 airline words, word error rate 2.9%
2.
Connected words, speaker dependent, airline reservations, finite state grammar with perplexity 4, 129 airline words, word error rate 0.1%
3.
Continuous speech, speaker independent, airline travel information system (ATIS), finite state grammar with perplexity 12, 1800 words, word error rate 4.0%
4.
Continuous speech, speaker independent, speech read from the Wall Street Journal, finite state grammar with perplexity 200, 20 000 words, word error rate 13.0%


Mikko Kurimo
11/7/1997