A novel approach for time-scale modification (TSM) of speech based on temporal continuous nonnegative matrix factorisation (TCNMF) is presented. First, the magnitude spectrum of the speech is factorised to the nonnegative space and the time-varying gains, and then the TSM problem is transformed into an interpolation problem of the timevarying gains, which leads to a better performance over the traditional methods based on waveform overlap-add. The superiority of the proposed approach is confirmed by the comparative tests against the traditional methods, including OLA, SOLA, WSOLA, and PSOLA.Introduction: The technology of time-scale modification (TSM) of speech can adjust the speed of a speech while keeping its perceptual features, including the pitch period, the formant structure, and so on. So it sounds like the speaker changes the speed of the speech initiatively.Early in 1984, Griffin and Lim proposed a method called OLA [1], which divides the speech into a series of overlap-added segments by a window function and through adjusting the length of the overlap parts, the time-scale of the speech can be compressed or expanded. But the defect of this method is that the phases of the processed speech are discontinuous. To overcome this defect, Roucos and Wilgus proposed a method called SOLA [2], and Verhelst and Roelands proposed a method called WSOLA [3]. These two methods introduce an offset to correct the discontinuous phase. However, the voiced speech exhibits periodical character, and the former methods will destroy the pitch structure of the speech during their processing. This will introduce metalling sounds into the processed speech. Then, Moulines et al. proposed a method called TDPSOLA [4]. This method operates the speech according to the unit of the pitch periods, so it can avoid destroying the pitch structure of the speech. So it depends on accurate pitch marks, and detecting the accurate pitch marks is a challenging task.
A deep learning scheme based on compressive sensing to detect community structure of large-scale social network is presented. Our contributions in this work are as follows: First, we reduced the high-dimensional feature of social media data via compressive sensing by using random measurement matrix; Second, deep belief network is employed to learn unsupervised from the low-dimensional samples; Finally the model is fine-tuned by supervised learning from a small scale sample sets with class labels. The effectiveness of the proposed scheme is confirmed by the experiment results.
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.