INTERSPEECH 2004 - ICSLP 8th International Conference on Spoken Language Processing Jeju Island, Korea October 4-8, 2004

A Distributed Speech Recognition System in Multi-User Environments

Kyu Jeong Han (1), Shrikanth Narayanan (1), Naveen Srinivasamurthy (2)

(1) University of Southern California, USA
(2) Qualcomm Incorporated, USA

A typical distributed speech recognition (DSR) system is a configuration that distributes computational burden in signal processing and pattern recognition between a mobile unit and a remote recognition engine. For this system to be robust and practically acceptable, distortions caused by erroneous data transmission should be minimized. In this paper, the effects of multiple simultaneous users in a wireless network on speech recognition are considered. Specifically, multiple access interference (MAI) is shown to be a significant factor in the recognition performance degradation of a DSR system. From simulation results, both a minimum-mean-square-error (MMSE) detector and a decorrelating filter are shown to be effective in reducing MAI and improving recognition accuracy. In a CDMA system with 6 interferers at an SNR of 7 dB in an AWGN channel, there was a 30% absolute reduction in word-error-rate (WER) for a connected digit recognition task by using an MMSE detector to combat MAI.

Full Paper

Bibliographic reference.  Han, Kyu Jeong / Narayanan, Shrikanth / Srinivasamurthy, Naveen (2004): "A distributed speech recognition system in multi-user environments", In INTERSPEECH-2004, 2121-2124.