计算机科学
卷积神经网络
判别式
变压器
编码器
嵌入
说话人识别
语音识别
特征提取
字错误率
模式识别(心理学)
人工智能
操作系统
物理
量子力学
电压
作者
Qiuyu Zheng,Zengzhao Chen,Zhifeng Wang,Hai Liu,Mingxing Lin
标识
DOI:10.1016/j.eswa.2023.123004
摘要
Transformer models have demonstrated superior performance across various domains, including computer vision, natural language processing, and speech recognition. The success of these models can be attributed to their robust parallel capacity and high computation speed, primarily reliant on the attention layer. In the domain of speaker recognition, state-of-the-art results have been achieved using convolutional neural network (CNN) architectures, particularly with speaker embeddings represented by x-vectors and r-vectors. However, existing CNN-based methods tend to focus on local features while overlooking the global dependence of voiceprint features, resulting in the loss of crucial information. Moreover, the presence of noise in audio data is an influential factor that cannot be disregarded, as it significantly impacts the extraction of discriminative speaker embeddings. To address these challenges, we propose a novel model called the Multi-Scale Expand Convolution Transformer (MEConformer). This model aims to convert variable-length audio into a fixed low-dimensional representation. The MEConformer leverages a CNN framework with expanded receptive fields to capture frame-level features effectively. Additionally, we introduce a transformer encoder that incorporates contextual dependencies, enabling the extraction of both frame-level and discourse-level feature representations. Furthermore, we present a multi-scale residual aggregation strategy, which facilitates the efficient transmission of voiceprint information across the model. By combining these innovative components, the MEConformer achieves a state-of-the-art Equal Error Rate (EER) of 3.72% on the VoxCeleb1 test set. Furthermore, it demonstrates EERs of 5.94% and 3.72% on the VoxCeleb1-H and VoxCeleb-E datasets, respectively. The code for the proposed MEConformer model will be made publicly available at https://codeocean.com/capsule/4563012/tree.
科研通智能强力驱动
Strongly Powered by AbleSci AI