Robust syntax free speech recognition
S. Boll, Jack E. Porter, L. Bahler · 2003
The use of a connected-speech recognizer configured to operate without a syntax specification and without training is discussed. The recognizer uses a rapid adaptation technique to match the unknown speech to a set of speaker-pooled word templates. A series of experiments are described in which 20 variable-length words are recognized in 15 minutes of conversational speech spoken by six males and two females. Results are given in terms of the percent of words correctly recognized at a false alarm rate of ten false alarms per hour. The recognizer found an average of 61% of the words in clean speech where the average signal-to-noise ratio is greater than 20 dB. When Gaussian noise is digitally added to achieve a signal-to-noise ratio of 10 dB, only an average of 28% of the words were found. When alternative methods of noise suppression are used, from 34% to 44% of the words are recognized.>