Multi Talker Recurrent Neural Network Transducer for Intelligent Recognition Technology of Russian Speech

Nowadays, the rapid growth for voice-controlled interfaces and smart assistants has determined the need for accurate Speech Recognition Systems (SRS) based on Artificial Intelligence (AI). Hence, the Convolutional Neural Network (CNN) has developed as a Russian SRS, but failed in handling accents, languages, and dealing with background echo and noise. To overcome these issues, a Multi Talker Recurrent Neural Network Transducer (MT-RNNT) is proposed for intelligent recognition of Russian speech. Initially, the procedure starts with the collection of data from the Saint Petersburg State University (SPSU). Then, the data is further processed with Spectral Deduction (SD) in order to remove irrelevant frequency components and the noise from data. After that, the Praat Speech (PS) technique segments long speech sentences and detects the end-points of the signal and removes silence from the preprocessed signals. Next, the Capsule Network (CapsNet) is employed to extract features for capturing the variations in tone, and the spectral characteristics. Finally, the speech recognition is done by using proposed MT-RNNT to recognize the speech signal as Russian speech. From the results, the proposed MT-RNNT achieved outstanding results in accuracy of 99.2% when compared with existing Deep Convolutional Neural Network and Bidirectional Long Short-Term Memory (DCNN-BiLSTM) model.