Experimental Results for Baseline Speech Recognition Performance using Input Acquired from a Linear Microphone Array

Experimental Results for Baseline Speech Recognition Performance using Input Acquired from a Linear Microphone Array Harvey F. Silverman, Smart E. Kirtman, John E. Adcock and Paul C. Meuse Laboratory for Engineering Man/Machine Systems (LEMS) ABSTRACT In this paper, baseline speech recognition performance is determined both for a single remote microphone and for a signal derived from a delay-and-sum beamformer using an eight-microphone linear array. An HMM-based, connected-speech, 38-word vocabulary (alphabet, digits, 'space', 'period'), talker-independent speech recognition system is used for testing performance. Normal performance, with no language model, i.e., raw word-level performance, is currently about 81% for a set of talkers not in the training set and about 91% for training set data. The system has been trained and tested using a close-talking bead-mounted microphone. Since a meaningful comparison requires using the same speech, the existing speech database was appropriately pre-filtered, played out through a transducer (speaker) in the room environment, picked-up by the microphone array, and re-stored as a digital file. The resulting file was post-processed and used as input to the recognizer; the recognition performance indicates the effect of the input device. The baseline experiment showed that both a single remote microphone and the beamformed signal reduced performance by 12% in a room with no other talkers. For the array tested, the error is generally attributable to reverberation off the floor and ceiling. 1. Introduction It is widely accepted that appropriate data-acquisition tech- nology must be available in order to make speech-recognition a viable computer input mode [1, 2, 3]. While work has been done in the area of signal conditioning [4], for the last three years, research at Brown University has been in progress to develop hardware, software and algorithms as a means to make non-intrusive speech acquisition a practical reality [5, 6] Principal focus to date has been to use the phase relationships among a group of microphones spaced in a line - hence a linear array - for the remote, real-time acquisition of a talker's data. Various beamforming and talker location/tracking algorithms have been studied, reported, and evaluated relative to listening quality [7, 8, 9, 10, 11, 12] The quality of a speech data acquisition system may be assessed in several ways. For many applications, evaluation is usually given, quantitatively, in terms of some signal- to-noise measure or human-listening experiment score, or qualitatively in terms of human evaluation. However, for a system whose output is fed to a speech recognizer, the *Supported principally by NSF/DARPA Grant No. IRI-8901882 Division of Engineering Brown University Providence, R102912 recognition performance is an excellent, quantifiable measure; this approach and its results make up the body of this paper. A key problem for such systems to overcome is that of rever- beration. Acoustic reflections in a normal room environment make the output of a remote microphone quite different from that taken from the normal, close-talking, recognizer micro- phone. Several ways have been suggested to alleviate this problem: • A more focused array system will attenuate reflections coming from a wider …

Content truncated to 500 words