Search

Automatic Speech Recognition (ASR) systems are notorious for their poor performance in adverse conditions, leading to high sensitivity and low robustness. Due to the costly and time-consuming nature of creating extensive speech databases, addressing the issue of low robustness has become a prominent area of research, focusing on the synthetic generation of speech data using pre-existing natural speech. This paper examines the impact of standard data augmentation techniques, including pitch shift, time stretch, volume control, and their combination, on the accuracy of isolated-word ASR systems. The performance of three machine learning models, namely Hidden Markov Models (HMM), Support Vector Machines (SVM), and Convolutional Neural Networks (CNN), is analyzed on two Serbian corpora of isolated words. The Whi-Spe speech database in neutral phonation is utilized for augmentation and training, and a specifically developed Python-based software tool is employed for the augmentation process in this research study. The conducted experiments demonstrate a statistically significant reduction in the Word Error Rate (WER) for the CNN-based recognizer on both testing datasets, achieved through a single augmentation technique based on pitch-shifting.

This article presents an analysis of different cepstral normalization techniques in automatic recognition of whispered and bimodal speech (speech+whisper). In these experiments, conventional GMM-HMM speech recognizer was used as speaker-dependant automatic speech recognition system with special Whi-Spe corpus containing utterance recordings in normally phonated speech and whisper. The following normalization techniques were tested and compared: CMN (Cepstral Mean Normalization), CVN (Cepstral Variance Normalization), MVN (Cepstral Mean and Variance Normalization), CGN (Cepstral Gain Normalization) and quantile-based dynamic normalization techniques such as QCN and QCN-RASTA. The experimental results show to what extent each of these cepstral normalization techniques can improve whisper recognition accuracy in mismatched train/test scenario. The best result is obtained using CMN in combination with inverse filtering which provides an average 39.9 percent improvement in whisper recognition accuracy for all tested speakers.