Fig. 3From: Performance vs. hardware requirements in state-of-the-art automatic speech recognitionFeature types and architectures used in Kaldi-based ASR systems. Feature type #1 (upper left) is used in TDNN models (center) and feature type #2 (middle left) is used in CNN-TDNN models (right). The output blocks (bottom left) that use cross-entropy and chain loss functions are used in both architecturesBack to article page