NOISE ROBUST SPEECH RECOGNITION USING PROSODIC INFORMATION - DSP for In-Vehicle and Mobile Systems - page 144

Figure 9-2. An example of the prosodic features in Japanese connected digit speech for a male
speaker's utterance, “9053308” “3797298”, with 20dB SNR white noise.

In this paper, two kinds of prosodic features and their combination, P-D, P-V,
and P-DV, are investigated:

P-D:

P-V: maximum accumulated voting value

P-DV: + maximum accumulated voting value

These three kinds of prosodic features are combined with segmental features for
each frame. Therefore, three kinds of segmental-prosodic feature vectors are
built and evaluated.