Conv-TasNet is a deep learning framework for end-to-end time-domain speech separation that outperforms previous time-frequency masking methods. It uses a fully-convolutional architecture with a linear encoder and decoder, and a temporal convolutional network (TCN) for mask estimation. The encoder generates a representation of the speech waveform optimized for separating individual speakers, while the TCN models long-term dependencies of the speech signal. The masks are applied to the encoder output to separate the speakers, and the decoder reconstructs the waveforms. Conv-TasNet achieves superior performance in both objective and subjective quality measures, with a smaller model size and shorter latency compared to previous methods. It is suitable for both offline and real-time speech separation applications. The system is robust to variations in the starting point of the mixture and has a smaller model size and shorter latency than previous methods. Conv-TasNet also shows better performance in separating speech from noise and reverberation. The framework is suitable for applications such as embedded systems and wearable hearing devices. Conv-TasNet can also serve as a front-end module for tandem systems in other audio processing tasks. However, it has limitations in long-term tracking of speakers and generalization to noisy and reverberant environments. Further research is needed to address these limitations.Conv-TasNet is a deep learning framework for end-to-end time-domain speech separation that outperforms previous time-frequency masking methods. It uses a fully-convolutional architecture with a linear encoder and decoder, and a temporal convolutional network (TCN) for mask estimation. The encoder generates a representation of the speech waveform optimized for separating individual speakers, while the TCN models long-term dependencies of the speech signal. The masks are applied to the encoder output to separate the speakers, and the decoder reconstructs the waveforms. Conv-TasNet achieves superior performance in both objective and subjective quality measures, with a smaller model size and shorter latency compared to previous methods. It is suitable for both offline and real-time speech separation applications. The system is robust to variations in the starting point of the mixture and has a smaller model size and shorter latency than previous methods. Conv-TasNet also shows better performance in separating speech from noise and reverberation. The framework is suitable for applications such as embedded systems and wearable hearing devices. Conv-TasNet can also serve as a front-end module for tandem systems in other audio processing tasks. However, it has limitations in long-term tracking of speakers and generalization to noisy and reverberant environments. Further research is needed to address these limitations.