An Investigation on Wireless Speech Recognition by Data Contamination and Robust Training Techniques

This paper is concerned with the robust endpoint detection and noisy speech recognition over wireless network. Firstly, the MLP-based and GMM-based endpoint detection incorporated with data contamination and continuous spectral subtraction techniques were investigated. Then, for noisy wireless speech recognition, a combined technique of data contamination and robust training was proposed to separately model the environmental characteristics and phonetic information. According to the results from an abbreviated stock name recognition task, we observe that the proposed techniques has the potential to improve robustness not only on diverse data contaminated training data, but also on the unmatched noisetype condition between training and testing environments.