Audio Source Separation by Shoji Makino

Author:Shoji Makino
Language: eng
Format: epub
Publisher: Springer International Publishing, Cham

8.3 Binary Features for Audio Signals

Before delving into the details about BNNs, we would like to discuss the meanings of binary features for audio signals. Since BNNs will eventually replace the real-valued variables and logic with bitwise arithmetic, they can be seen as a way to approximate a Boolean mapping function whose input and output are bit strings. Hence, in order to make this argument complete, we need to ensure that the input and output of the network have their corresponding binary representations as well. Since most of the time the original signals are real-valued, we need a way to convert them into bit patterns without affecting the performance of the source separation system.

There are two important criteria we can use when choosing a binary feature extraction method: whether it is efficient and invertible. First, since the conversion from the raw input (e.g. magnitudes of Fourier spectra) to any bit string is an additional procedure, it should not take up too much resources for the system-wide efficiency. Second, once the BNN produces bit strings as its output, the conversion from the output to the original signal domain must be efficient and not as lossy as possible, too. There are such binary feature extraction methods that also defines its corresponding decoding procedure, which automatically converts the bit string back to the raw signal domain. Otherwise, we need to construct a scheme that finds the most similar bit string from the database (e.g. in terms of Hamming distance), and then take the signal that has generated the matching bit pattern as the output of the network.3 In any ways, the inversion procedure should be able to recover the original signal with least error, and as efficiently as possible.

Note that IBMs are naturally binary variables, so they can serve as the binary target variables as they are. If the system needs to use IRM or any other real-valued target variables, we need to binarize the target variables as well, which are supposed to be converted back to the raw signal using an appropriate conversion process.

In the following sections we review some existing hashing techniques to discuss their pros and cons as our candidate binary feature extraction methods.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.