Call Transcription Methodology for Contact Center Systems

Abstract

Nowadays, one of the key areas of research on contact centre systems is their automation. The main element that influences the possibility of automation of contact centre processes is the call transcription methods implemented by automatic speech recognition (ASR) systems. Such systems enable developing intention recognition methods and, consequently voice bots. The current solutions used in ASR systems for many less popular languages do not guarantee a fully satisfactory transcription quality for hotline voice calls. This is due to the unique characteristics of the sound signal generated there, whose quality parameters differ significantly from those of studio recordings. The paper presents a comparative study of selected speech recognition systems that were additionally supplemented with elements of preprocessing of sound recordings and postprocessing of originally produced transcriptions. As for preprocessing, the following methods were tested: separation of the client and agent channels into two independent signals, training of ASR systems, and audio signal correction. With regards to postprocessing, on the other hand, tests were performed for inarticulate sounds, normalization of standard phrases (e.g. numbers, dates, times, etc.), and identification of close-sounding phrases and foreign language phrases, and lemmatization. Based on the research conducted and the analyses performed, a new method of call transcription intended specifically for contact center systems was proposed. The research conducted for this paper was based on the Polish language model, for which major problems are observed with the quality of automatic contact center call transcriptions.

Keywords

• Automatic speech recognition
• call centre
• contact centre
• transcription
• word error rate