In this paper, the bottleneck feature extraction technique with MLP is used on Punjabi adult speech recognition. Nowadays, neural networks are most widely used approaches for training and testing the system. It helps to recognize the back probabilities among various phoneme set. This input info includes at some point get wrapped, and it becomes difficult to prepare them on Hidden Markov Model (HMM) based state-of-the-art synthesis. Here, context-based model is trained on Deep Neural Network (DNN) and after that on Bottleneck-Neural Network (BN-NN) system with the use of Multi-layer Perceptron (MLP). The baseline ASR is performed with different environment conditions on different modelling system. To improve the performance of a system, MLP-based supervised learning method utilizing for adjoining voice outlines related data to change the design of profound neural system DNN by extracting the bottleneck features. Finally, the MLP are used as input for the DNN-HMM and BN-NN state-of-the-art system. This paper presents the larger improvement obtained by applying the MLP feature vector with the relative improvements of 4.03% which is achieved on the Punjabi ASR with varying the several attributes associated with BN-NN and DNN-HMM modelling approaches.
CITATION STYLE
Bala, S., Kadyan, V., & Bhardwaj, V. (2021). Bottleneck feature extraction in punjabi adult speech recognition system. In Lecture Notes in Networks and Systems (Vol. 171, pp. 493–501). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-981-33-4543-0_53
Mendeley helps you to discover research relevant for your work.