Özet:
This article presents a new method for calculating
the signal energies of speech segments in voice activity detection
algorithms. In the study, the µ-law signal compression method is
adapted to calculate short-term signal energies. A simple voice
activity detection (VAD) algorithm is designed to demonstrate the
effectiveness of the proposed method. The same VAD algorithm was
also run with two different conventional energy calculation formulas
and the performance of each VAD was evaluated using time-domain
short-time energy features. The G729 standard VAD algorithm was
also used for performance comparison. During the test of the
analyzed detectors, many kinds of input speech signals with various
types of background environmental noise, such as restaurants,
vehicles, and streets, were tested. Using the new energy calculation
method, the VAD detector has improved detection accuracy
compared to VAD detectors based on the other two energy methods
and was able to effectively identify voice-active regions even in noisy
conditions at low SNR levels. The results revealed that the VAD
detector designed with the proposed new energy calculation formula
outperforms traditional energy-based voice activity detection
methods and provides noticeable increases in detection rate even
under adverse conditions.