Improvement of the Quality of Automatically Recognised Speech from Financial Services Advertising

The objective of the graduate qualification thesis is to reduce the number of errors in recognized speech by defining error classes in existing speech-to-text models. The result of the final qualification work is the development of an algorithm for detecting and classifying speech recognition errors to improve the quality of automatic recognition, fragmentary testing of the algorithm in practice, as well as putting forward a proposal for possible error correction for existing speech-to-text models. In this graduate qualification work, the following stages were implemented: an overview of the specifics of advertising texts, including the specifics of advertising of financial activities; classification of speech-to-text systems, as well as review and analysis of existing speech recognition tools; review and analysis of existing classes of errors in speech recognition, methods for their detection and correction, as well as methods for evaluating the results; compilation of a database of advertising audio files; conducting tests on the basis of existing models and analyzing the results; development of an algorithm for detecting and classifying speech recognition errors; putting forward a proposal for possible error correction for existing speech-to-text models. Key words: automatic speech recognition; speech recognition errors; machine learning; advertising of financial services. The work consists of 100 pages, made with the assistance of 64 sources, contains 8 tables, 20 figures, 5 appendixes.