<p>The performance of the Automatic Speech Recognition systems is impressive in the clean and closed room environment, and it significantly deteriorates in the presence of uncontrolled conditions. The traditional features like Mel Frequency Cepstral Coefficients and Perceptual Linear Prediction Features work well for the clean data. The use of these techniques on data contaminated by environmental noise, data having transient noise and sudden bursts in spoken sounds, speech data that is nonstationary in nature results in degraded ASR performance. To address this nonstationary nature of spoken data and environmental noises, Wavelet Transform driven Wavelet Packet Thresholding and Perceptual Wavelet Packet Cepstral features are proposed. This technique handles the non-stationary speech signal very well as the Wavelet transform is localized both in time and frequency. The proposed features are integrated with hybrid Deep Neural Network-Hidden Markov Model(DNN-HMM) acoustic model, and n-gram language model for recognizing spoken continuous Kannada sentences collected from uncontrolled conditions. The performance of this work, utilizing the proposed features, was evaluated using a standard Kannada speech corpus and the benchmarking TIMIT dataset. To assess robustness, experiments were further conducted on versions of both datasets contaminated with various types of noise at different Signal-to-Noise Ratio (SNR) levels, derived from the standard NOIZEUS database. A comparative analysis between the proposed features and conventional baseline features was performed. The results demonstrate improved performance over the baseline Mel Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP) features, respectively. These findings confirm that the proposed Automatic Speech Recognition (ASR) system achieves superior performance compared to MFCC and PLP features.</p>