This paper explores use of synergistically-integrated systems of microphone arrays and neural networks for robust speech recognition in variable acoustic environments, where the user must not be encumbered by microphone equipment. Existing speech recognizers work best for "high-quality close-talking speech." Performance of these recognizers is typically degraded by environmental interference and mismatch in training conditions and testing conditions. It is found that use of microphone arrays and neural network processors can elevate the recognition performance of existing speech recognizers in an adverse acoustic environment, thus avoiding the need to retrain the recognizer, a complex and tedious task. We also present results showing that a system of microphone arrays and neural networks can achieve a higher word recognition accuracy in an unmatched training/testing condition than that obtained with a retrained speech recognizer using array speech for both training and testing, i.e., a matched training/testing condition.

Microphone arrays and neural networks for robust speech recognition

The Huge Microphone Array (HMA) is a collaborative effort between Brown University and Rutgers University that started in February 1994 to design, construct, debug, and test a real‐time 512‐microphone array system and to develop algorithms for use on it. Analysis of known algorithms made it clear that signal‐processing performance of over 6 Gigaflops would be required; at the same time, there was a need for ‘‘portability,’’ i.e., fitting into a small van, that also set an upper limit to the power required. It was essential that the array be able to be used in both large and small acoustic environments. These tradeoffs and many others have led to a unique design in both hardware and software. The hardware uses 128 fast, floating‐point DSP microprocessors and is designed so that data flow is independent of data processing. This leads to an unusually simple software environment. This paper presents the full design and its justifications. Performance for a few important algorithms relative to usage of processing‐capability, response latency, and difficulty of programming is discussed. [Work supported by NSF Grant MIP‐9314625.]

The Huge Microphone Array (HMA)

This paper describes algorithms for automatic end‐point detection of microphone‐array speech signals. Microphone arrays provide a hands‐free sound pickup. The captured sound typically has a higher signal‐to‐noise ratio (SNR) than that captured with conventional microphones used at distances, such as in teleconferencing environments. However, due to multipath distortion (room reverberation) and ambient noise, the detection of starting/ending points of array speech is more difficult than that of close‐talking speech. In this paper, short‐time energy and short‐time zero‐crossing rate are computed for the original speech waveform and its high‐pass filtered and low‐pass filtered versions. These six functions are then utilized in different combinations to determine the end points. Speech data used in the experiments are collected in a hard‐walled laboratory room, having a reverberation time of approximately 0.5 s with a one‐dimensional beamforming line array. From the experiments, it is found that the high‐pass filtered signal gives a more reliable estimate of end points than does the low‐pass filtered counterpart. This result is consistent with the fact that reverberation and noise in rooms are typically more prominent at low frequencies and are relatively moderate at mid‐ and high frequencies. The detection algorithms have been integrated into a dynamic‐time‐warping‐ (DTW) based speech recognizer. Recognition performance of the system is evaluated for both array speech and for close‐talking speech.

Qiguang Lin

Papers

End‐point detection of microphone‐array speech signals