Toward the pre-cocktail party problem with TasTas+
Deep neural network with dual-path bi-directional long short-term memory (BiLSTM) block has been proved to be very effective in sequence modeling, especially in speech separation, e.g. DPRNN-TasNet <cit.>, TasTas <cit.>. In this paper, we propose two improvements of TasTas <cit.> for end-to-end approach to monaural speech separation in pre-cocktail party problems, which consists of 1) generate new training data through the original training batch in real time, and 2) train each module in TasTas separately. The new approach is called TasTas+, which takes the mixed utterance of five speakers and map it to five separated utterances, where each utterance contains only one speaker's voice. For the objective, we train the network by directly optimizing the utterance level scale-invariant signal-to-distortion ratio (SI-SDR) in a permutation invariant training (PIT) style. Our experiments on the public WSJ0-5mix data corpus results in 11.14dB SDR improvement, which shows our proposed networks can lead to performance improvement on the speaker separation task. We have open-sourced our re-implementation of the DPRNN-TasNet in https://github.com/ShiZiqiang/dual-path-RNNs-DPRNNs-based-speech-separation, and our TasTas+ is realized based on this implementation of DPRNN-TasNet, it is believed that the results in this paper can be reproduced with ease.
READ FULL TEXT