结合有监督联合一致性自编码器的跨音视频说话人标注

Efficient Audio-visual Cross-modal Speaker Tagging via Supervised Joint Correspondence Auto-encoder

跨模态说话人标注旨在利用说话人的不同生物特征进行相互匹配和互标注,可广泛应用于各种人机交互场合。针对人脸和语音两种不同模态生物特征之间存在明显的"语义鸿沟"问题,该文提出一种结合有监督联合一致性自编码器的跨音视频说话人标注方法。首先分别利用卷积神经网络和深度信念网络分别对人脸图像和语音数据进行判别性特征提取,接着在联合自编码器模型的基础上,提出一种新的有监督跨模态神经网络模型,同时嵌入softmax回归模型以保证模态间和模态内样本的相似性,进而扩展为3种有监督一致性自编码器神经网络模型来挖掘音视频异构特征之间的潜在关系,从而有效实现人脸和语音的跨模态相互标注。实验结果表明,该文提出的网络模型能够有效的对说话人进行跨模态标注,效果显著,取得了对姿态变化和样本多样性的鲁棒性。

Cross-modal speaker tagging aims to learn the latent relationship between different biometrics for mutual annotation, which can potentially be utilized in various human-computer interactions. In order to solve the ＂semantic gap＂ between the face and audio modalities, this paper presents an efficient supervised joint correspondence auto-encoder to link the face and audio counterpart, where by the speaker can be crosswise tagged. First, Convolutional Neural Network （CNN） and Deep Belief Network （DBN） are used to extract the discriminative features of the face and the audio samples respectively. Then, a supervised neural network model associated with softmax regression is embedded into a joint auto-encoder model, which can discriminatively preserving the inter-modal and intra-modal similarities. Accordingly, three different kinds of supervised joint correspondence auto-encoder models are presented to correlate the semantic relationships between the face and the audio counterparts, and the speaker can be crosswise annotated efficiently. The experimental results show that the proposed supervised joint auto-encoder is able to perform cross-modal speaker tagging with outstanding performance, and demonstrate the robustness to facial posture variations and sample diversities.