Automatic Speech Recognition (ASR) System using convolutional and Recurrent neural Network Approach

Abstract

Nowadays, speech recognition is an active research field, where various deep neural architectures are explored. The published successful models are optimized on massive, transcribed datasets, most of which are closed. A deep neural network solves two closely related tasks. It learns to recognize phonemes and formulate grammar rules at the same time. A model can parallel and accurately build both of them when a training corpus is large enough. However, inflected languages such as Polish contain much more grammar rules to define than in the case of English. Therefore, to achieve comparable results in the Polish language, the corpus must be substantially larger than the one presented for the English language. In contrast, to build more massive datasets, we present the Synthetic Boosted Model, which is an attempt to use synthetic data to enrich more profound the implicit language model. In the presented work, we propose the new model architecture, the new objective function, and the new training policy.