Speech therapy using computer based minimal consonant pair discrimination

Using concepts from neural network architecture, the authors developed a system based on traditional template matching techniques that do not require the lengthy training times typical of networks. Dynamic selection from a variety of feature generation options and a detailed selection of regions over time and frequency using an error-rate-minimization criterion allow the system to focus on aspects of speech patterns that are most effective for phoneme pair discrimination. Preliminary application indicates clinically useful performance with TBI, CHI, and stroke patients.<>