next up previous
Next: Discussion Up: Listening Test Previous: Data Reduction

Results

Absolute word identification accuracy was moderately high overall (an average of 72.5 percent of the words correct). However, listeners identified words correctly more often for the adult talker (F[1,19] = 138.47, p <.01) with 77.5 percent correct responses for the adult versus 67.4 percent correct for the child talker.

For the comparison of signal processing methods, we examined the difference between the percent words correct in the F0-shifted tokens and the original token. Overall, the signal processing was responsible for an increase of 3.7 in the percentage of errors. There was no significant main effect of processing METHOD (F[1,19] = 0.62, p =0.44). The main effect of F0 shift DIRECTION was significant (F[1,19] = 24.86, p <.01) with increased F0 resulting in an increase of 5.6 in percentage of errors, while decreasing F0 raised the percentage of errors by only 1.8. DIRECTION interacted significantly with METHOD (F[1,19]=5.37, p =.032). This interaction is illustrated in Figure 2 which shows that the PSOLA method performed better than HYBRID when F0 was shifted down and worse than HYBRID when F0 was increased.


Figure 2: Difference in percentage of identification errors for direction of F0 shift as a function of processing method.

The main effect of TALKER was marginally significant (F[1,19] = 4.41, p =.049) with the child's speech being less distorted by F0 shift than the adult's (a change in percentage error of 2.3 for the child and 5.0 for the adult). The interaction METHOD and TALKER was also significant (F[1,19] = 5.06, p = .037). Means for this interaction are plotted in Figure 3 showing that the PSOLA method is again associated with both the best and worst performance.


Figure 3: Difference in percentage of identification errors for each talker as a function of processing method.

No other interactions involving METHOD were significant, although the interaction of TALKER by DIRECTION of F0 shift was significant (F[1,19] = 12.82, p < .01) due to a disproportionately larger number of errors for the adult when her F0 was increased.



next up previous
Next: Discussion Up: Listening Test Previous: Data Reduction



Tim Bunnell
Tue Dec 5 10:53:09 EST 1995