IMPORTANCE OF EARLY AND LATE REFLECTIONS FOR AUTOMATIC SPEECH RECOGNITION IN REVERBERANT ENVIRONMENTS
Michael Kleinschmidt · 2003
In the European norm DIN EN ISO 3382 (1) about the measurement of reverberation time of rooms with hints to other acoustical parameters an early-to- late energy ratio is defined as a parameter that represents a ratio of early reflections energy to energies of reflections arriving after a certain critical delay time. The background understanding consists of the idea that in reverberant room situations early reflections improve the auditory perception while late portions, arriving with more than the critical delay time, have a detrimental effect. According to (1), the usual critical delay times for speech and music perception are 50 ms and 80 ms, respectively. This paper evaluates the importance of early and late reflections for the accuracy of automatic speech recognition (ASR). The effective time cutoff be- tween conducive and detrimental portions of the impulse response is determined. This is done by successively deleting certain components of a room impulse re- sponse by setting them to zero. These modified impulse responses serve for rever- beration of speech material that is recognized with a recognizer previously trained on non-reverberated speech. Canceling out portions with a detrimental effect for the recognition should cause increasing recognition accuracies, while eliminating conducive portions should result in lower accuracy rates. In this sense, recogni- tion performance is observed dependent on manipulations of the underlying room impulse response. The results are substantially comparable to human perception: Early reflections up to a certain critical delay time can carry useful information and contribute to the recognition accuracy when detrimental late portions are present. This cutoff time is found in the range of 25 ms to 50 ms for different room impulse responses and two different front ends.