ESPnet is an open-source end-to-end speech processing toolkit designed for automatic speech recognition (ASR) and other speech processing tasks. It utilizes widely-used deep learning frameworks, Chainer and PyTorch, as its main deep learning engine. ESPnet follows the Kaldi ASR toolkit style for data processing, feature extraction, and recipe setup, providing a complete environment for speech recognition experiments. The toolkit supports various end-to-end ASR techniques, including hybrid CTC/attention architectures, RNNLM integration, and fast CTC computation using the warp CTC library. ESPnet provides recipes for major ASR benchmarks such as WSJ, LibriSpeech, TED-LIUM, CSJ, AMI, HKUST Mandarin CTS, VoxForge, and CHiME-4/5. It also supports multilingual ASR and noise-robust/far-field speech recognition. ESPnet's architecture includes a Kaldi-style data preprocessing module, an attention-based encoder-decoder, and hybrid CTC/attention training and decoding. The toolkit simplifies the ASR pipeline by using Python and offers efficient training and recognition with a reduced number of code lines compared to other systems. Experimental results show that ESPnet achieves competitive performance with state-of-the-art hybrid HMM/DNN systems, particularly in tasks like CSJ and HKUST. ESPnet is also an official baseline for the CHiME-5 challenge. The toolkit is actively developed with features such as multi-GPU support, data augmentation, and multilingual ASR experiments.ESPnet is an open-source end-to-end speech processing toolkit designed for automatic speech recognition (ASR) and other speech processing tasks. It utilizes widely-used deep learning frameworks, Chainer and PyTorch, as its main deep learning engine. ESPnet follows the Kaldi ASR toolkit style for data processing, feature extraction, and recipe setup, providing a complete environment for speech recognition experiments. The toolkit supports various end-to-end ASR techniques, including hybrid CTC/attention architectures, RNNLM integration, and fast CTC computation using the warp CTC library. ESPnet provides recipes for major ASR benchmarks such as WSJ, LibriSpeech, TED-LIUM, CSJ, AMI, HKUST Mandarin CTS, VoxForge, and CHiME-4/5. It also supports multilingual ASR and noise-robust/far-field speech recognition. ESPnet's architecture includes a Kaldi-style data preprocessing module, an attention-based encoder-decoder, and hybrid CTC/attention training and decoding. The toolkit simplifies the ASR pipeline by using Python and offers efficient training and recognition with a reduced number of code lines compared to other systems. Experimental results show that ESPnet achieves competitive performance with state-of-the-art hybrid HMM/DNN systems, particularly in tasks like CSJ and HKUST. ESPnet is also an official baseline for the CHiME-5 challenge. The toolkit is actively developed with features such as multi-GPU support, data augmentation, and multilingual ASR experiments.