Traditional voice activity detection (VAD) methods work well in clean and controlled scenarios, with performance severely degrading in real-world applications. One possible bottleneck for such supervised VAD training is its requirement for clean training data and frame-level labels. In contrast, we propose the GPVAD framework, which can be easily trained from noisy data in a weakly supervised fashion, requiring only clip-level labels. We proposed two GPVAD models, one full (GPV-F), which outputs all possible sound events and one binary (GPV-B), only splitting speech and noise. We evaluate the two GPVAD models and a CRNN based standard VAD model (VAD-C) on three different evaluation protocols (clean, synthetic noise, real). Results show that the GPV-F demonstrates competitive performance in clean and noisy scenarios compared to traditional VAD-C. Interestingly, in real-world evaluation, GPV-F largely outperforms VAD-C in terms of frame-level evaluation metrics as well as segment-level ones. With a much lower request for data, the naive binary clip-level GPV-B model can still achieve a comparable performance to VAD-C in real-world scenarios.
View on arXiv