Robust recognition of strongly distorted speech

Abstract

The automatic speech recognition systems have become a part of our daily lives. People often rely on virtual personal assistants in smartphones, use their voice to control intelligent devices in cars and smart homes or communicate with automatic dialogue systems in call-centres. Since these systems often suffer from a performance drop in realistic acoustic conditions which are characterized by strong distortions, a large portion of research still must be focused on robust front-end algorithms and acoustic modelling methods for distorted speech recognition. This thesis is focused on these compensation methods working at the level of front-end processing and acoustic modelling, whose aim is to compensate the degradation introduced by a distant microphone, noisy environments and a lossy compression. The techniques for noisy and distant speech recognition studied in this thesis were focused on front-end noise suppression techniques, feature normalization techniques, acoustic model adaptations and discriminative training. Said techniques were evaluated in three different car conditions and two different public environments. The experiments have proved, that extended spectral subtraction can bring significant improvement even for the state-of-the-art systems in public environments with a strong noise and for a far-distance microphone recordings. The evaluation of compressed speech recognition examined the degrading effects of lossy compression on fundamental frequency, formants and smoothed LPC spectrum and for standard MFCC and PLP features used for ASR. The low-pass filtering and the areas of very low energy in a spectrogram were identified as the two main reasons of degradation. The practical experiments evaluated the contributions of specific feature extraction setups, combinations of normalization and compensation techniques, supervised and unsupervised adaptation and discriminative training methods and finally the matched training. The largest contributions were gained from the application of adaptation techniques, subspace GMM and discriminative training. A novel algorithm named Spectrally selective dithering (SSD) was proposed within this thesis, which compensated the effect of spectral valleys. The contribution of said algorithm was verified for both GMM-HMM and DNN-HMM speech recognition systems for Czech and English and for a GMM-HMM system for German. The practical experiments proved that the proposed algorithm can lower WER for all languages with GMM-HMM systems. Concerning DNN-HMM system, a significant contribution was achieved only for Czech.