A comparison of human and computer recognition accuracy for children's speech

Several studies have shown that automatic speech recognition error rates are greater for children’s speech than for adult’s speech. Investigations have demonstrated that word recognition error rates increase as age decreases, and that recognition performance for children’s speech is more sensitive to bandwidth reduction, compared with adult speech. This paper presents the results of experiments to measure human recognition performance for children’s speech. The paper compares human and machine recognition performance on the same children’s speech data. It is shown that human recognition performance for children’s speech exhibits similar effects of age and bandwidth to those observed for automatic systems. The results suggest that effects of age and bandwidth on automatic speech recognition accuracy are due to properties of children’s speech rather than artifacts of the technology