Lombard effect compensation for speech and speaker recognition systems using deep learning neural networks

Abstract

The speech in a noisy environment, called the Lombard speech newline(LS), is more intelligible than speech in a laboratory environment, called the newlinenormal speech (NS). The involuntary tendency of a person speaking with more newlinevocal effort in a noisy environment to improve the intelligibility of voice is newlinetermed as Lombard effect (LE). The speech production changes due to noise newlinemanifests itself in the form of acoustic-phonetic changes in Lombard speech. newlineSpeech systems trained to recognize the normal speech features, when tested newlinewith features extracted from Lombard speech, undergo loss in performance newlinedue to the mismatch in the train-test conditions. It also becomes essential newlineto look for complimentary speech cues from other modalities under certain newlineadverse condition where standard normal microphone cannot be used. The main newlineobjective of this thesis is to reduce the spectral dissimilarities between NS and newlineLS, and to explore the possibilities of combining speech cues from different newlinemodalities so as to improve the recognition performance of speech-based newlinerecognition systems. newlineThe acoustic-phonetic and articulatory differences between NS and newlineLS on different sound units are observed for vocal-tract, excitation source as newlinewell as prosodic features of Lombard speech. The variations observed for newlinedifferent parameters are dependent on many factors like gender of the speaker, newlinelevel and type of noise, mode of passing the masking noise to induce Lombard newlineeffect in a speaker, to name a few. In order to overcome the performance loss of newlinespeech systems due to Lombard speech, Lombard effect compensation methods newlineare carried out either at the feature level or at model level. newline