Energy redistribution speech intelligibility enhancement, vocalic and transitional cues

Abstract

Two novel speech enhancement algorithms are presented that automatically increase intelligibility in noisy environments while maintaining the signal power and naturalness of the original speech. These energy redistribution (ER) algorithms move signal energy to targeted regions of relatively high information content that are crucial for intelligibility. The boosted regions are originally of low energy and therefore usually the first segments lost with the addition of environmental noise. The ER voiced/unvoiced (ERVU) method transfers energy from voiced speech to regions of unvoiced speech, while the ER spectral transition (ERST) method moves energy from spectrally stationary regions to spectrally transitional regions. Hand-held cell phones and public address systems are expected to be the dominant applications for these techniques. Standard noise reduction methods such as spectral subtraction are assumed to have already been applied to the voice signal before broadcast. Using human listening tests, it was found that both algorithms boost the intelligibility of speech in noisy environments by nearly 7% over the original unprocessed signals, without degrading naturalness or increasing signal power. Furthermore, both algorithms allow for controlling the trade-off between boost gain and speech naturalness.