Purpose: To evaluate and compare the performance of 2 methods of voice activity detection (neck-attached accelerometer vs. binaural recordings) in field studies in environments where voice activity normally occurs.
Method: A group of 11 healthy adults wore recording equipment during their lunch break. We used binary classification to analyze the results from the 2 methods. The output was compared to a gold standard, obtained through listening tests, and the probability for sensitivity (Ps) and false positive (Pf) was rated. The binary classifiers were set for consistent sensitivity of 99%; thus, the lower false positive rate would indicate the method with the better performance.
Results: The neck-attached accelerometer (Pf = 0.5%) performed significantly (p < .001) better than the binaural method (Pf = 7%).
Conclusion: The neck-attached accelerometer is more suitable than the binaural method for voice assessments in environments where people are speaking in close proximity to each other and where the signal-to-noise ratio is moderate to low.