Synthetic Audio Data Generation Algorithm for the Diarization Problem

In this paper an approach to building a dataset for the diarization problem is considered. A pipeline is proposed for fine-tuning a pretrained diarization model on synthetic data in order to increase diarization accuracy for Russian language. The suggested algorithm for processing audio data allows to add random portions of silence between adjacent audio segments, make several segments partially overlap and add specific type of noise to the audio track. Diarization error rate metric is calculated on the generated audio files by utilizaing Python’s pyannote diarization model. The results show that the artificially generated synthetic audio data is adequate and closely resembles real audio tracks. The proposed method involves speech extraction, concatenation, normalization, audio track formation, and noise overlaying. Our experiments have shown that generating longer and more diverse conversations leads to better quality data.