NAS-VAD: Neural Architecture Search for Voice Activity Detection

1. Introduction

• Deep learning improved VAD detection, especially in noisy environments, while contextual information was explored to improve performance at low SNRs.Manual architecture design requires substantial engineering effort and becomes more challenging as networks grow.Source: §1 · p. 1
• NAS had not been applied to VAD, although it had succeeded in image classification and automatic speech recognition.The paper identifies manual architecture design as time-consuming and increasingly challenging for larger, more complicated networks.Source: §1 · p. 1
• The proposed VAD NAS framework expands the search space with attention operations and modifies the macro-architecture for VAD models’ shallow depths.The authors report outperforming manually designed models across audio datasets and improved generalization.Source: §1 · p. 1

2. Background and Related Work

• Cell-based NAS searches operations and connections within repeating cells, reducing search-space size while retaining performance.The paper distinguishes this from searching a whole architecture or optimizing only the inner structure within a fixed macro architecture.Source: §2 · p. 1
• NAS strategies include evolutionary methods, Bayesian optimization, reinforcement learning, gradient-based methods, and random search.Search spaces may cover a whole architecture or only operations and connections within a fixed macro architecture.Source: §2 · p. 1
• Earlier speech-related NAS work adjusted search spaces or macro structures but omitted sequential operations reported to improve VAD.The paper states that no previous study had explored NAS for VAD.Source: §2 · pp. 1–2

3. NAS-VAD

• NAS-VAD combines a VAD-oriented search space with a modified macro-architecture to search for models with fewer layers and improved performance.The approach uses a cell-based search and includes diverse operations intended to benefit VAD.Source: §3 · p. 2

3.1. Search Strategy

• NAS-VAD applies cell-based Bayesian optimization to automatically search VAD architectures, using NAS-BOWL and a Weisfeiler-Lehman subtree graph kernel for sample-efficient search.A Gaussian process serves as the surrogate model to reduce the cost of training and evaluating candidate models.Source: §3.1 · p. 2, §3 · p. 2

3.2. Search Space

• The VAD search space contains 18 operations, including convolutional, attention, feed-forward, squeeze-and-excitation, gated-linear, skip-connection, and zero operations.Attention spans time, feature, or both dimensions, with two or four heads; the space also includes residual blocks and feed-forward variants.Source: §3.2 · p. 2

3.3. Macro Architecture

• The macro-architecture uses four cells to reflect the shallower depth of VAD models and reduce latency.At the reduction stage it halves feature width, not time height; two convolutional layers precede the first cell, and a linear layer follows the last.Source: §3.3 · p. 2

4. Experiments

• The comparison used manually designed state-of-the-art neural VAD baselines within the same frame-wise framework; stochastic and algorithmic VAD approaches were excluded.Those approaches could not work on the non-consecutive audio frames used in the comparison.Source: §4 · p. 2

4.2. Results

• NAS-VAD outperformed every baseline on noise-added datasets across training and test datasets, using 151K parameters—fewer than any other network.Table 1 compares the model with four baselines.Source: Table 1 · p. 3, §4.2 · p. 4
• On real-world datasets, NAS-VAD achieved the best performance on most test sets; CV-trained models led on every set, while TIMIT-trained models missed room and AVA.The authors suggest that training on a small number of artificial noise-added samples may lead to overfitting to those datasets.Source: Table 2 · p. 3, §4.2 · p. 4

4.1.1. Datasets

• TIMIT and Common Voice 7.0 supplied clean speech; SoundIdeas and selected Audioset noises were added at randomly sampled SNRs from -10 to 10 dB.AVA and ACAM served as real-world test datasets; CV ground-truth labels were extracted using an algorithmic VAD method.Source: §4.1.1 · p. 3

4.1.2. Architecture search

• Architecture search used the TIMIT-based dataset, AUC as its metric, and 16 default channels to target a model size similar to the smallest state-of-the-art VAD model.Source: §4.1.2 · p. 4

4.1.3. Model Evaluation

• Evaluation used noise-added speech datasets for training and testing, followed by real-world recordings to assess generalization to unseen audio.Models were trained at least three times, and average test performance and standard deviation were reported.Source: §4.1.3 · p. 4

4.3. Ablation Study

• The proposed search space and macro-architecture improved performance, especially on CV, compared with the original DARTS search space.The authors report that the compared models used the same search algorithm and had similar parameter counts.Source: Table 3 · p. 4, §4.3 · p. 4

• NAS-VAD combines a search space and macro-structure optimized for automatic VAD model design.Source: §5 · p. 4
• The authors report that models searched by NAS-VAD outperform existing VAD models.Source: §5 · p. 4
• The authors expect the framework could also be used to improve other audio-processing tasks.Source: §5 · p. 4