From border controls to personal devices, from online exam proctoring to human-robot interaction, biometric technologies are empowering individuals and organizations with convenient and secure authentication and identification services. However, most biometric systems leverage only a single modality, and may face challenges related to acquisition distance, environmental conditions, data quality, and computational resources. Combining evidence from multiple sources at a certain level (e.g., sensor, feature, score, or decision) of the recognition pipeline may mitigate some limitations of the common uni-biometric systems. Such a fusion has been rarely investigated at intermediate level, i.e., when uni-biometric model parameters are jointly optimized during training. In this chapter, we propose a multi-biometric model training strategy that digests face and voice traits in parallel, and we explore how it helps to improve recognition performance in re-identification and verification scenarios. To this end, we design a neural architecture for jointly embedding face and voice data, and we experiment with several training losses and audio-visual datasets. The idea is to exploit the relation between voice characteristics and facial morphology, so that face and voice uni-biometric models help each other to recognize people when trained jointly. Extensive experiments on four real-world datasets show that the biometric feature representation of a uni-biometric model jointly trained performs better than the one computed by the same uni-biometric model trained alone. Moreover, the recognition results are further improved by embedding face and voice data into a single shared representation of the two modalities. The proposed fusion strategy generalizes well on unseen and unheard users, and should be considered as a feasible solution that improves model performance. We expect that this chapter will support the biometric community to shape the research on deep audio-visual fusion in real-world contexts.