IMPLEMENTATION OF THE SPEAKER IDENTIFICATION SYSTEM BASED ON THE PYANNOTE AUDIO PROCESSING LIBRARY

Abstract

In the field of machine learning, one of the main areas is speech processing and recognition. One of the important tasks of working with audio data is diarization. Diarization determines the time boundaries in the audio recording belonging to individual speakers, that is, figuratively speaking, solves the problem of «who speaks when?». However, known commercial and open source diarization tools use segment clustering, but do not answer the question «who exactly is speaking now?». There are systems that identify the speaker, but such systems are designed for the fact that there is only one speaker in the audio recording. Therefore, a relevant task is to create a diarization system that allows the identification of many speakers that arbitrarily change in audio recordings. In this study, we propose two architectures of speaker identification systems based on diarization, which work respectively on the per segment basis and per cluster analysis. To implement the system, we used the PyAnnote library, which is open source. The evaluation of the speaker identification system was carried out on the open audio database AMI Corpus, which contains 100 hours of annotated and transcribed audio and video data. Various metrics for assessing the accuracy of diarization are considered and, taking into account the specifics of the developed system, the expediency of using such an assessment as the F-Measure of identification is substantiated. The methodology of research is described, which included three experiments. The first experiment is aimed at studying the architecture of the identification system based on per segment analysis, and the second experiment is aimed at studying the architecture that uses per cluster analysis. The third experiment concerns the determination of the optimal training sample duration for the classifiers of the identification system. The experimental results showed that the cluster-based approach showed better identification results compared to the segment-based approach. It was also found that the optimal duration of audio data sampling for training the classifier for each specific speaker is 20 seconds.