An experimental study on fundamental frequency detection in reverberated speech with pre-trained recurrent neural networks

Abstract

The detection of the fundamental frequency (f0) in speech signals is relevant in areas such as automatic speech recognition and identification, with multiple potential applications. For example, in virtual assistants, assistive technology devices and biomedical applications. It has been acknowledged that the extraction of this parameter is affected in adverse conditions, for example, when reverberation or background noise is present. In this paper, we present a new method to improve the detection of the f0 in speech signals with reverberation, based on initialized Long Short-term Memory (LSTM) neural networks. In previous works, LSTM has used weights initialized with random numbers. We propose an initialization in the form of an auto-associative memory, which learns the identity function from non-reverberated data. The advantages of our proposal are shown using different objective quality measures, in particular, in the detection of segments with and without f0.