One of the prominent features of the SPRACH Broadcast News system was the inclusion of a second information stream based on the novel modulation-filtered spectrogram (MSG) features, which afforded significant performance improvements over the baseline Perceptual Linear Prediction (PLP) features. These are discussed in the Broadcast News report (D7.2), as well as a paper published in Speech Communication .
The unprecedented amount of training data available for Broadcast News allowed us to use our MultiSPERT hardware to the full, leading to the very large MLP acoustic models described in the Broadcast News report (D7.2), as well as a submission to ICASSP-98 .
Work on latent variable models continued, resulting in a presentation at the IEEE Workshop on Neural Networks for Signal Processing , and a paper to appear in Speech Communication .
Investigations were also made in the area of model combination based on various criteria, included acoustic confidence measures, and this is discussed in the Broadcast News report (D7.2).
Theoretical basis of the multi-stream paradigm have been summarized in the deliverable D5.1.
Work on Non Linear Discriminant Analysis was applied to the
Broadcast News system. The results are shown in table 2.1.5
for the following systems:
26*9-1000-54 : 9 frames of 26 features, 1000 hidden nodes
26*9-2000-54 : 9 frames of 26 features, 2000 hidden nodes
36*9-600-54 : 9 frames of 36 features coming from NLDA analysis, 600 hidden nodes
36*9-2000-54 : 9 frames of 36 features coming from NLDA analysis, 2000 hidden nodes