Spectro-Temporal Attention-Based Voice Activity Detection

NAS-VAD: Neural Architecture Search for Voice Activity Detection

“…To further improve the performance at low signal-to-noise ratios (SNRs), several studies have proposed the use of context information. For instance, Zhang et al [2] proposed aggregating predictions from different temporal contexts, and other approaches employed diverse forms of attention mechanisms [3], [4], [5].…”

“…As opposed to the previous cell-based NAS methods that use eight cells [21], [19], we use only four cells because general VAD models [32], [3], [4], [5] have a shallower depth for lower latency. In our four-cell structure, the first two cells are convolution cells, and the next two cells are attention cells.…”

“…As an input feature, we used logmelspectrogram with 80 mel filterbanks, viewing each spectrogram as a single channel image. As in earlier VAD work [2], [3], [4], [5], we trained each model to detect voice activity in seven neighboring frames [t-19, t-10, t-1, t, t+1, t+10, t+19] for a given frame at time t.…”

“…3) Model Evaluation: For performance comparison, we used the following manually-designed models: bDNN [32], ACAM [3], spectro-temporal attention VAD [4], and selfattentive VAD [5]. For self-attentive VAD, we used the official model implementation, and for other models, we implemented them as described in the original papers.…”

“…V OICE detection performance has been improved by adopting deep neural networks [1], [2], specifically with the use of context information and the attention mechanism [3], [4], [5]. In the general deep learning research domain, the need for a reduction of human effort in network design has led to the emergence of the neural architecture search (NAS) technique that automatically optimizes model structures and operations.…”