A robust word boundary detection algorithm with application to speech recognition
Hany Agaiby, Tom James Moir · 2002
A new robust word boundary detection algorithm is described that performs well under a variety of noise conditions including competing talkers. The algorithm uses the direction of the signal as the main criterion to differentiate between wanted-speech and background noise. A 'viewing zone' is assumed within which a speech source is considered desired-speech and signals coming from outside this zone are considered noise. The algorithm uses the time delay between signals received at two microphones to estimate the direction of the dominant signal. This estimate together with an estimate of the coherence function between the two signals as well as measures of the signal energy are used to determine word boundaries. Two state-of-the-art speech recognisers were used to evaluated the performance of the algorithm. For each recogniser, the recognition accuracy is measured with manually labelled noisy speech and compared when speech is automatically processed using the proposed algorithm. The results showed that the algorithm performs as well as manual labelling under signal-to-noise ratios as low as 0 dB.