Speech and word detection algorithms for hands-free applications

This paper describes a robust speech detection algorithm for speech-activated hands-free applications. The system consists of three techniques: (1) noise suppression with efficient implementation, (2) robust endpoint detection and (3) speech verification using garbage modeling and confidence measure. With efficient implementation, noise suppression improves the SNR by roughly 10-20 dB. The endpoint detection uses the technique described in [1] with improvement for non-stationary noise. Garbage modeling and confidence measure are used to handle out-of-vocabulary (OOV) words and background pulse noise.