No abstract

Singer identification is to automatically identify the singer in a music recording, such as a polyphonic song. A song has two major acoustic components that are singing vocals and background accompaniment. Although identifying singers is similar to speaker identification, it is challenging due to the interference of background accompaniment on the singer-specific information in singing vocals. We believe that separating the background accompaniment from the singing vocal will help us to overcome the interference. In this work, we extract the singing vocals from polyphonic songs using Wave-U-Net based audio-source separation approach. The extracted singing vocals are then used in i-vector based singer identification system. Further, we explore different state-of-the-art audio-source separation methods to establish the role of considered method in application to singer identification. The proposed singer identification framework achieves an absolute accuracy improvement of 5.66% over the baseline without audio-source separation.

No abstract

Speech activity detection (SAD) is a part of many speech processing applications. The traditional SAD approaches use signal energy as the evidence to identify the speech regions. However, such methods perform poorly under uncontrolled environments. In this work, we propose a novel SAD approach using a multi-level decision with signal knowledge in an adaptive manner. The multi-level evidence considered are modulation spectrum and smoothed Hilbert envelope of linear prediction (LP) residual. Modulation spectrum has compelling parallels to the dynamics of speech production and captures information only for the speech component. Contrarily, Hilbert envelope of LP residual captures excitation source aspect of speech. Under uncontrolled scenario, these evidence are found to be robust towards the signal distortions and thus expected to work well. In view of different levels of interference present in the signal, we propose to use a quality factor to control the speech/nonspeech decision in an adaptive manner. We refer this method as multi-level adaptive SAD and evaluate on Fearless Steps corpus that is collected during Apollo-11 Mission in naturalistic environments. We achieve a detection cost function of 7.35% with the proposed multi-level adaptive SAD on the evaluation set of Fearless Steps 2019 challenge corpus.

Sonorant sounds are characterized by regions with prominent formant structure, high energy and high degree of periodicity. In this work, the vocal-tract system, excitation source and suprasegmental features derived from the speech signal are analyzed to measure the sonority information present in each of them. Vocal-tract system information is extracted from the Hilbert envelope of numerator of group delay function. It is derived from zero time windowed speech signal that provides better resolution of the formants. A five-dimensional feature set is computed from the estimated formants to measure the prominence of the spectral peaks. A feature representing strength of excitation is derived from the Hilbert envelope of linear prediction residual, which represents the source information. Correlation of speech over ten consecutive pitch periods is used as the suprasegmental feature representing periodicity information. The combination of evidences from the three different aspects of speech provides better discrimination among different sonorant classes, compared to the baseline MFCC features. The usefulness of the proposed sonority feature is demonstrated in the tasks of phoneme recognition and sonorant classification.