This paper presents a method for determining subword units for end-to-end prosody generation. End-to-end prosody generation using deep neural networks (DNNs) is expected to directly generate a prosody sequence from text without any professional knowledge in the target language. In natural language processing, language model-based language-independent subword tokenization was previously proposed for determining subwords suitable for end-to-end language processing. However, the subwords determined by the language models are not appropriate for end-to-end speech processing. In this paper, we propose a language-independent algorithm for determining subwords that maximize acoustic model likelihoods. The proposed algorithm iterates expectation-maximization (EM)-based training of DNN acoustic models and likelihood-based construction of the subword vocabulary. In the experimental evaluation, we discuss the stability of the EM-based training and analyze subword vocabularies determined by the conventional language model-based and proposed acoustic model-based methods.