This paper addresses the task of phonetic segment classification, a fundamental challenge in speech therapy, and proposes a method that utilizes joint embeddings learned from ultrasound tongue imaging (UTI), audio, and mmWave radar data. To create embeddings, we compiled a collection of raw mmWave radar signals synchronized with ultrasound images and audio, focusing on specific consonants. The embeddings are derived from artificial neural network models trained on this dataset. Additionally, recent advances have introduced jointly trained models that, while leveraging multiple data sources during training, are designed to operate with only a single modality at inference, enhancing practicality without sacrificing performance. During inference, our model, USRadioAI, solely utilizes radar data, excluding UTI and audio, hence improving practicality for potential speech therapy applications. Experimental results show that USRadioAI performs as well as approaches that use both audio and UTI, while outperforming those that rely only on a single modality, such as radar, audio, or UTI, in real-time phonetic classification.