Bridging Gaps in Russian Language Processing: AI and Everyday Conversations

Contemporary advancements in NLP and neural network techniques are paving the way to enhance and harness traditional linguistic resources and corpora, as well as expand the methods of applying neural networks for complex language material. Thus, a weak point for both theoretical and applied linguistic tasks is the processing of spontaneous everyday speech. Two experiments described in this article are dedicated to the analysis of how successfully modern neural models cope with the recognition and generation of everyday Russian speech. The material for the experiments is the well-known ORD speech corpus, the largest collection of professional and mundane dialogues in Russian. The first experiment targets the pressing issue of increasing the volume of transcribed speech data through state-of-the-art automatic speech recognition techniques. Experimental recognition was conducted using two diverse methods – the NTR Acoustic Model and OpenAI’s Whisper system. The second experiment zeroes in on refining generative language models tailored for Russian using a conversational dataset. A prototype dialogue system, derived from the enhanced ruGPT-3 Small model, exemplifies the transformative potential of fine-tuning in dialogue generation tasks. The acquired results are utilized to enrich datasets for recognizing everyday Russian speech and for constructing chatbots that emulate spontaneous Russian conversations.