Analysis of methods and approaches to audio data processing

Within the research, it is necessary to assess the automatic speech recognition systems available today, adapting them to the Ukrainian language and considering the trade-offs between accuracy, performance, and various other factors. The aim of the research was to compare the existing methods and technologies for speech recognition, with the goal of creating a system for assessing the informational impact of the analysed audio files. The study utilised traditional hidden Markov processes and Gaussian mixtures models, hybrid neural network models, such as DeepSpeech, Wav2Vec 2.0, and Whisper, cloud-based solutions, such as those by Google, Amazon, Microsoft, and IBM. Audio processing was carried out using speech detection and mel-frequency coefficients; word recognition accuracy, processing time on the CPU and GPU, response latency and instability, and the impact of recognition errors on the natural language processing pipeline with tokenisation, topic classification and sentiment analysis was assessed. As a result, it became possible to establish the dependence of the accuracy and performance of Automatic Speech Recognition systems upon the architecture of the recognition systems and the hardware upon which they are installed. Classic Hidden Markov Models were found to have the lowest performance in recognising the words spoken within audio files, with word error rates between 18 and 30%, as well as processing times of 51 to 72 seconds, indicating their limited suitability for Ukrainian language recognition. End-to-end neural network models, however, had significantly better recognition performance, with DeepSpeech models having an error rate of 22%, Wav2Vec models having an error rate of 12%, and whisper models having an error rate of only 7%. Models based upon deep learning and transformers were also found to be robust to phonetic variations of the Ukrainian language. Furthermore, the local end-to-end neural network models had significantly better processing speeds than either cloud-based solutions or classic Hidden Markov Models, with whisper models taking only around 10 seconds to process an audio file. Cloud-based recognition systems had similar accuracy to local models (between 7 and 10% error rate), but required dependence upon the network, indicating potential privacy issues for those using such systems. Thus, results of this research indicate that whisper and Wav2Vec models are the best methods to utilise for audio content analysis and information effect detection systems. The findings of this research can be applied to the development of automatic transcription systems, audio monitoring systems, voice analytics services, or the development of information influence detection modules