A distributed speech recognition system in multi-user environments

Kyu Jeong Han, Shrikanth S. Narayanan, Naveen SrinivasamurthyInterspeech 2004 · DOI

Introduction

Distributed Speech Recognition (DSR) systems are designed to offload the heavy computational burden of speech recognition from mobile devices to remote servers. While this architecture enables sophisticated voice interfaces on low-power hardware, it introduces a new vulnerability: the wireless channel. Most research in DSR assumes a point-to-point connection, but real-world networks are crowded. This paper investigates Multiple Access Interference (MAI)—the interference caused by multiple users sharing a single wireless channel—and its detrimental effects on speech recognition accuracy.

Methodology

The researchers utilized a standard DSR framework based on the ETSI Aurora standard. In this setup, the client device performs feature extraction (Mel-Cepstrum parameterization) and compression, while the server handles error detection, decompression, and final recognition.

The DSR architecture separates lightweight feature extraction from heavy recognition tasks.

To address the noise introduced by multiple users in a Code Division Multiple Access (CDMA) environment, the study integrated Multi-User Detection (MUD) techniques at the server side. Two linear filtering schemes were evaluated for their ability to suppress MAI:

  • Minimum-Mean-Square-Error (MMSE) Detector: A filter that minimizes the mean square error between the original and received signals.
  • De-correlating (DC) Filter: A simpler approach that uses the inverse of the channel's cross-correlation matrix to separate user signals.

Experimental Setup

The system was tested using the TIDigits database, comprising over 12,000 utterances for training and testing. The wireless environment was simulated as a synchronous CDMA system using Gold sequences of length 63. Performance was measured primarily through the Word Error Rate (WER) across varying Signal-to-Noise Ratios (SNR) and numbers of interfering users.

Results and Discussion

The experiments revealed that MAI is a significant degrader of ASR performance. Even in high SNR conditions where background noise is minimal, the presence of other users creates a performance bottleneck. For instance, without mitigation, the system achieved a baseline accuracy of 96.69% in a clean environment, but the WER climbed toward 50% as the number of users increased to 15.

Performance comparison showing the significant reduction in WER achieved by MMSE and DC filters in the presence of 6 interferers.

The application of MUD filters yielded dramatic improvements. In a scenario with 6 interferers at 7 dB SNR, the baseline WER of approximately 40-50% was reduced by an absolute 30% using either the MMSE or DC filters. Notably, the simpler DC filter performed nearly as well as the MMSE detector, suggesting that robust interference suppression can be achieved with low computational overhead at the server.

Conclusion

This research demonstrates that Multiple Access Interference is a critical factor in the performance of mobile DSR systems. By implementing low-complexity multi-user detection algorithms, it is possible to significantly recover recognition accuracy in crowded wireless environments. Future work will explore these effects in asynchronous systems and different multiplexing environments like TDMA.

Summary updated

BibTeX
@inproceedings{han2004distributed,
  author = {Kyu Jeong Han and Shrikanth S. Narayanan and Naveen Srinivasamurthy},
  title = {{A distributed speech recognition system in multi-user environments}},
  booktitle = {Interspeech},
  year = {2004},
  doi = {10.21437/interspeech.2004-644},
}
Figure