Editing BLIND NORMALIZATION OF SPEECH FROM DIFFERENT CHANNELS AND SPEAKERS

Title: BLIND NORMALIZATION OF SPEECH FROM DIFFERENT CHANNELS AND SPEAKERS

Abstract: This research explores a technique to create a universal representation of speech signals that remains consistent despite variations in channel and speaker transformations. This technique, known as blind normalization, involves rescaling the speech signal based on its instantaneous level at each time, with the form of the scale function determined by recently encountered signal levels. This method allows the speech signal to be represented in a way that remains invariant under any invertible time-independent transformation, making it possible to normalize signals related by channel-dependent and speaker-dependent transformations without having to characterize or compensate for these transformations.

Research Question: Can we develop a technique to normalize speech signals that are related by channel-dependent and speaker-dependent transformations, without having to characterize or compensate for these transformations?

Methodology: The study employs a technique called blind normalization, which involves rescaling the speech signal based on its instantaneous level at each time. The form of the scale function is determined by recently encountered signal levels, ensuring that the rescaled speech representation is invariant under any invertible time-independent transformation. This technique is illustrated by applying it to time-dependent spectra of speech that have been filtered to simulate the effects of different channels.

Results: The experimental results show that the rescaled speech representations are largely normalized (i.e., channel-independent), despite the channel-dependence of the raw (unrescaled) speech. This suggests that the technique is successful in creating a universal representation of speech signals that remains consistent despite variations in channel and speaker transformations.

Implications: The blind normalization technique has significant implications for the field of speech processing. It may make it possible to develop more efficient and effective speech recognition systems, as well as improve the quality of communication in noisy environments. Additionally, the technique could be applied to other time-dependent signals, such as music or seismic data, to create universal representations that remain consistent despite variations in signal transformations.

Link to Article: https://arxiv.org/abs/0204003v1
Authors: 
arXiv ID: 0204003v1

[[Category:Computer Science]]
[[Category:Speech]]
[[Category:Technique]]
[[Category:Transformations]]
[[Category:Channel]]
[[Category:Signal]]