For low complexity, mobile, hands-free, speaker independent connected digit recognition, a fixed-point digital signal processor based implementation is essential. In this paper, we investigate algorithms for connected-digit recognition using whole-word digit models and a background model. We show that significant improvement can be achieved by using background model adaptation, continuously adaptive separate cepstral mean subtraction for background and speech segments and discriminative training. The system achieves almost 96% digit accuracy on a 15 speaker database of speech recorded in a car. A real-time system using the Lucent's DSP1627 has also been developed. We also present the results of our experiments in reducing complexity for the fixed-point system. These include a method to approximate state-likelihood computation using a Vector Quantization based mixture selection and use of beam width pruning during Viterbi decoding.
Cite as: Raghavan, P., Gupta, S.K. (2000) Low complexity connected digit recognition for mobile applications. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 4, 390-393, doi: 10.21437/ICSLP.2000-832
@inproceedings{raghavan00_icslp,
author={Prabhu Raghavan and Sunil K. Gupta},
title={{Low complexity connected digit recognition for mobile applications}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 4, 390-393},
doi={10.21437/ICSLP.2000-832}
}