Real-time neural network voice activity detection for speech recognition

The paper investigates the task of real-time speech recognition in a noisy environment. We propose an original approach of adapting modern neural network algorithms of voice activity detection RealVADR to solve the problem of real-time speech recognition using sound interval processing. The influence of the parameters of this algorithm on the quality of speech recognition is considered, as well as methods of optimising its parameters. Experiments have been conducted both on the existing open dataset CommonVoice and on several custom datasets collected in a noisy robotic environment. They showed that the application of the proposed approach allows obtaining in real time a recognition quality comparable to offline recognition.