Single-microphone, speaker-independent speech separation is normally performed through two steps: (i) separating the specific speech sources, and (ii) determining the best output-label assignment to find the separation error. The second step is the main obstacle in training neural networks for speech separation. Recently proposed Permutation Invariant Training (PIT) addresses this problem by determining the output-label assignment which minimizes the separation error. In this study, we show that a major drawback of this technique is the overconfident choice of the output-label assignment, especially in the initial steps of training when the network generates unreliable outputs. To solve this problem, we propose Probabilistic PIT (Prob-PIT) which considers the output-label permutation as a discrete latent random variable with a uniform prior distribution. Prob-PIT defines a log-likelihood function based on the prior distributions and the separation errors of all permutations; it trains the speech separation networks by maximizing the loglikelihood function. Prob-PIT can be easily implemented by replacing the minimum function of PIT with a soft-minimum function. We evaluate our approach for speech separation on both TIMIT and CHiME datasets. The results show that the proposed method significantly outperforms PIT in terms of Signal to Distortion Ratio and Signal to Interference Ratio.
Co-channel speech recordings typically contain significant amounts of overlap in which the intelligibility and quality of the desired speech is degraded by interference from a competing talker. Convolutive Non-negative Matrix Factorization (CNMF) has been shown to be a successful approach in detecting overlap by extracting specific acoustic basis dimensions for each speaker from an audio stream. While the results of CNMF have been successful, it requires isolated single speech recordings for each speaker to derive their corresponding bases functions/dimensions. In our previous work, The Teager-Kaiser Energy Operator (TEO)-based Pyknogram has been introduced which does not require prior information concerning the speakers. In this study, Pyknogram and CNMF based solutions for overlap detection within audio streams have been examined using the GRID dataset. TEO-based Pyknogram is shown to achieve a relative 8-10% lower Equal Error Rate (EER) compared to CNMF features. Another drawback of CNMF is that its performance drops considerably when dealing with spontaneous speech that has not been considered for extracting bases in the training step. In addition to the experiments on GRID corpus, a secondary evaluation is also performed based on naturalistic audio streams with overlap. Specifically, we collected a real-world audio database of US Presidential debates stemming from the last 12 years that are challenging due to overlap, changing Signal to Interference Ratio (SIR), and environmental noise, etc. Our experiments indicate that TEO-based Pyknogram is well suited for detecting overlap in challenging real world scenarios such as the US presidential debates.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.