Autodirective Microphone Systems for Natural Communication With Speech Recognizers

AUTODIRECTIVE MICROPHONE SYSTEMS FOR NATURAL COMMUNICATION WITH SPEECH RECOGNIZERS J. L. Flanagan and R. Mammone CAIP Center, Rutgers University, New Brunswick, New Jersey G. W. Elko AT&T Bell Laboratories, Murray Hill, New Jersey Abstract Two technological advances support new sophis- tication in sound capture; namely, high-quality low- cost electret microphones and high-speed economical signal processors. Combined with new understand- ing in acoustic beamforming, these technologies per- mit spatially-selective transduction of speech signals several octaves in bandwidth. Spatial selectivity mk- igates the effects of noise and reverberation, and digi- tal processing provides the capability for speech- seeking, autodirective performance. This report out- lines the principles of autodirective beamforming for acoustic arrays, and it describes two experimental im- plementations. It also summarizes the direction and emphasis of continuing research. Introduction In many applications of automatic speech recog- nition, it is desirable for the talker to have hands and eyes free for concurrent tasks. Typical examples in- clude parcel sorting, product assembly and inspec- tion, voice dialing for cellular telephones, and data plotting and manipulation in a situation room. The user frequently needs to move around in the workspace, which often is noisy and reverberant, while issuing commands to the speech recognizer. Electrical tethers, close-talking microphones and body-worn sound equipment represent undesirable encumbrances. Ideally, one would like an acoustic system able to capture high-quality sound from natu- ral conversational exchanges in the work space. Speech-seeking autodirective microphone arrays enable unencumbered freedom of movement, while providing sound pickup quality approaching that of close-talking microphones. Low-cost high-quality electret microphones, in combination with economi- cal signal processing, permit sophisticated beam- forming and dynamic beam positioning for tracking a moving talker. Multiple beam formation permits "track while scan" performance, similar to phased- array navigational radars, so that multiple sound sources can be monitored and algorithmic decisions made about the signals [1,2]. Beamforming has been found to be more useful than adaptive noise filtering for sound pickup in noisy, reverberant enclosures [3]. This report mentions the acoustic principles in- volved in dynamic beamforming and the design fac- tors governing the ability of steered arrays to combat noise and room reverberation. It discusses the as-yet rudimentary algorithms for sound source location and speech/non-speech detection. It then describes an ini- tial application of an autodirective array and a limited-vocabulary connected-word speech recog- nizer for voice control of a video/audio teleconferenc- ing system. It concludes by indicating the directions for research needed to refine further the capabilities of hands-free natural sound pickup. • Acoustic Beamforming The signal output H from an arbitrary array of N discrete omnidirectional acoustic sensors due to a time-harmonic plane wave with wavevector k is N-1 H(k, r) = ~ a n e -jk'r" , (1) n=O where an is the amplitude weighting of sensor n, r,t is the position vector of sensor n with respect to some defined origin, and the bold case indicates a vector quantity. The time-harmonic term is omitted for compactness. The array can be steered to wave arrivals from different directions by …

Content truncated to 500 words