Engineering: Combining acoustic cues

• Does not capture interrelations between cues
• Cannot generate confidence scores
• Vulnerable to cascade failure

• Linear combination is not adequate
• Combine linear transformations with saturating nonlinearities
• No theoretical basis for algorithm
• No justification for optimality
• Disappointing results

• Multi Layer Perceptron (MLP) for stateless decision
• Use NICO toolkit
• Three inputs, one output, one hidden layer
• Training issues
  • Backpropagation training, fast but requires differentiable metric
  • Mean squared error metric, inadequate but differentiable
• Two hidden units found to be adequate
• Some insight from examining network weights