Who Said That? a Comparative Study of Non-negative Matrix Factorization Techniques

Abstract

In noisy environments it is difficult for a computer to understand what a person is saying, especially when there are multiple speakers. In this paper we concentrate on separating overlapping speech. Non-negative matrix factorisation (NMF) is a method of doing source separation without needing a lot of data. The choice of cost function can have a significant impact on the performance of NMF. We evaluate NMF using three different cost functions (Euclidean, Itakura-Saito and Kullback-Leibler), including modifications using sparsity, convolution or additional information in the form of the direction of arrival. We conduct this evaluation on three different speech corpora. Adding directional information to NMF in the form of nonnegative tensor factorisation (NTF) gives us the best result on the map task and vocalization corpora and the Itakura-Saito cost function performs best on the acoustic-camera corpus. In this paper, we show that the Itakura-Saito cost function is the most robust cost function when the recording contains noise. We do this by applying acoustic evaluation measurements.

Non-negative Matrix Factorization on a Multi-lingual Overlapped Speech Signal: A Signal and Perception Level Analysis

“…The above result may be explained by the fact that all http://journals.uob.edu.bh [25], (b) NMF-EUC [4], (c) NMF-EUC, current investigation on an overlapped speech signal comprising same language speeches by English female and male speakers and (d) NMF-EUC, current investigation on an overlapped speech signal comprising different language by English male and Marathi female speakers. The y-axis scale for all the figures is same Indo-Aryan languages like Marathi and Bengali have more aspirated consonants than English, which are produced with an audible expulsion of breath, whereas the unaspirated are pronounced with minimal breath.…”

“…They also point out that there is immense scope for improving audio source separation in overlapping speech scenarios. DNN, though it shows promising results in separation performance, is characterized by high computational complexity and suffers degraded performance on problems with limited training data or small data sets [25]. NMF, on the other hand, is still prevalent for separation with limited training datasets.…”

“…Figure5. SDR, SIR, and SAR comparison of (a) NMF-EUC(sparse)[25], (b) NMF-EUC[4], (c) NMF-EUC, current investigation on an overlapped speech signal comprising same language speeches by English female and male speakers and (d) NMF-EUC, current investigation on an overlapped speech signal comprising different language by English male and Marathi female speakers. The y-axis scale for all the figures is same…”