Unsupervised voice activity detection based on the envelope's fractal dimension

Abstract

Currently, voice activity detection (VAD) is utilized in many fields, including forensics, healthcare, and medicine, to detect vocal anomalies, as well as in telecommunications and mobile telephony. Due to its importance and the difficulty of distinguishing between speech and nonspeech segments, especially in noisy environments (low signal-to-noise ratio (SNR)), this area remains under continuous development. Most existing VAD algorithms require predefined thresholds or training data, which reduces their compatibility. This study proposes an unsupervised VAD system that utilizes the Katz algorithm to calculate the fractal dimension of envelopes obtained through a single frequency filtering (SFF) approach. This method allows for high temporal and frequency resolution. The proposed VAD algorithm does not require any training data and is suitable for various types of noise and SNRs. To evaluate the effectiveness of the proposed method, two different databases are used: the Texas Instruments Massachusetts Institute of Technology (TIMIT) database and the King Saud University (KSU) Arabic speech database. The experimental results reveal an average detection accuracy of 96.86%, demonstrating its considerable value in various applications.