next up previous contents
Next: Forecast for future Up: Current results obtained Previous: Results with clear

Results with degraded acoustic training

In this case, for each S/N condition, the HMMs are trained using auditory data degraded with an additive white noise of the same intensity as that used during the test session. Figure C.9 shows the results of this experiment. Scores are presented as in figure C.8 . The AV score in the left area (1) is highly influenced by the noisy data from the auditory vector so that the resulting recognition rate is lower than the V score. In the center area (2), AV scores are higher than both A and V scores. In the right area (3), AV scores are similar or even lower than A scores. These last results confirm a well known fact that HMMs have to be trained with (even artificially) degraded data to improve the recognition scores when the test conditions are degraded.

Figure C.9 : Test results obtained from the audio-visual (AV) and audio recognizer (A). System has been trained for each test condition.

Esprit Project 8579/MIAMI (Schomaker et al., '95)
Thu May 18 16:00:17 MET DST 1995