A modular system for improving speech intelligibility under extreme acoustic conditions: subjective evaluation of parameter influence

We present a modular system for improving speech intelligibility in extreme acoustic conditions. These conditions may be high reverberation, very low Signal-to-Noise Ratios, or heavy babble noise. Our aim is to improve a noise-free speech signal by means of signal processing algorithms with respect to its intelligibility for the broadcast in such environments. In this paper we particularly conduct a subjective evaluation of the modules' and their respective parameters' eectiveness.We asked a group of subjects to transcribe sense-free sentences from a set of speech signals, which has been processed by the presented system and convoluted with impulse responses of medium and high reverberation. By calculating the corresponding word error rate we demonstrate that reproduction speed, vowel- to-consonant ratio and between-utterance pauses are all very eective in improving intelligibility in this task. On the other hand, pitch shifting and pauses inserted at vowel positions even degrade the intelligibility of speech signals.