Real-time connected-word recognition in a noisy environment

The authors deal with connected-word recognition in high-noise environments, especially for car applications. A description of some preprocessing methods that significantly improve a given recognition system is given. A large database recorded in a car at a speed of 120 km/h (75 m.p.h.) was used for the recognition experiments. When the car database was used instead of low-noise data, the performance of the system dropped from about 1% error rate to 14.5%. Introducing a dynamic compression algorithm, a spectral weighting method, noise cancellation, and embedded training improved the performance to an error rate of about 3%