Recently sequence-to-sequence models have started to achieve state-of-the-art\nperformance on standard speech recognition tasks when processing audio data in\nbatch mode, i.e., the complete audio data is available when starting\nprocessing. However, when it comes to performing run-on recognition on an input\nstream of audio data while producing recognition results in real-time and with\nlow word-based latency, these models face several challenges. For many\ntechniques, the whole audio sequence to be decoded needs to be available at the\nstart of the processing, e.g., for the attention mechanism or the bidirectional\nLSTM (BLSTM). In this paper, we propose several techniques to mitigate these\nproblems. We introduce an additional loss function controlling the uncertainty\nof the attention mechanism, a modified beam search identifying partial, stable\nhypotheses, ways of working with BLSTM in the encoder, and the use of chunked\nBLSTM. Our experiments show that with the right combination of these\ntechniques, it is possible to perform run-on speech recognition with low\nword-based latency without sacrificing in word error rate performance.\n