Towards Lightweight Speaker Verification via Adaptive Neural Network Quantization

TL;DR

This work tackles the challenge of deploying speaker verification systems on mobile devices by introducing adaptive neural network quantization methods. It presents three main strategies: adaptive uniform precision quantization (KMQAT) that uses per-layer k-means centroids, mixed-precision quantization guided by Hessian-based layer sensitivity with a multi-stage fine-tuning (MSFT) schedule, and specialized 1-bit binary quantizers (static and adaptive). Empirical results on VoxCeleb show lossless 4-bit quantization for ResNet and DF-ResNet families with substantial compression, while mixed-precision quantization offers further performance gains and flexible bit-width targets; adaptive binary quantization substantially closes the gap in 1-bit regimes. The comprehensive comparisons indicate the proposed methods outperform prior lightweight SV approaches across model sizes, enabling efficient mobile deployment without sacrificing accuracy. Overall, the work delivers practical, high-performance, compact SV systems suitable for real-world applications.

Abstract

Modern speaker verification (SV) systems typically demand expensive storage and computing resources, thereby hindering their deployment on mobile devices. In this paper, we explore adaptive neural network quantization for lightweight speaker verification. Firstly, we propose a novel adaptive uniform precision quantization method which enables the dynamic generation of quantization centroids customized for each network layer based on k-means clustering. By applying it to the pre-trained SV systems, we obtain a series of quantized variants with different bit widths. To enhance the performance of low-bit quantized models, a mixed precision quantization algorithm along with a multi-stage fine-tuning (MSFT) strategy is further introduced. Unlike uniform precision quantization, mixed precision approach allows for the assignment of varying bit widths to different network layers. When bit combination is determined, MSFT is employed to progressively quantize and fine-tune network in a specific order. Finally, we design two distinct binary quantization schemes to mitigate performance degradation of 1-bit quantized models: the static and adaptive quantizers. Experiments on VoxCeleb demonstrate that lossless 4-bit uniform precision quantization is achieved on both ResNets and DF-ResNets, yielding a promising compression ratio of around 8. Moreover, compared to uniform precision approach, mixed precision quantization not only obtains additional performance improvements with a similar model size but also offers the flexibility to generate bit combination for any desirable model size. In addition, our suggested 1-bit quantization schemes remarkably boost the performance of binarized models. Finally, a thorough comparison with existing lightweight SV systems reveals that our proposed models outperform all previous methods by a large margin across various model size ranges.

Figures (9)

• Figure 1: The pipeline of k-means clustering based quantization levels. Using 3-bit quantization as an illustration, this process involves three steps: partition, clustering and rescaling to yield eight quantization centroids.
• Figure 2: The comparison of different 4-bit quantization levels, including Uniform, PoT, APoT and k-means clustering (ours).
• Figure 3: The pipeline of mixed-precision quantization. It consists of three steps. Firstly, KMQAT is utilized to produce uniform precision quantized models. Then, mixed-precision search is performed to generate the optimal bit combination for each network layer. Finally, we introduce multi-stage fine-tuning strategy to quantize and fine-tune network in a progressive manner.
• Figure 4: The overview of both static and adaptive binary quantization. Static quantization (left) projects real-valued weights into a fixed binary set $q \in\left\{-1, +1\right\}$ across all layers. In contrast, adaptive quantization (right) is capable of flexibly selecting a distinct binary set $Q \in\left\{b_1, b_2 \right\}$ for each layer to better align with the distribution of real-valued weights.
• Figure 5: Pre-trained weight distributions for the first lower and last deeper convolutional layers in ResNet34.