Triple-0: Zero-shot denoising and dereverberation on an end-to-end frozen anechoic speech separation network

TL;DR

This paper introduces a new method for speech enhancement that uses a pre-trained network without any fine-tuning to clean up noisy and reverberated speech.

Contribution

The novel contribution is a zero-shot speech denoising and dereverberation method using a frozen pre-trained network without architectural changes or fine-tuning.

Findings

The frozen network performs well in zero-shot testing under noisy and reverberant conditions without prior exposure.

The model outperforms the WPE algorithm in terms of speech quality and intelligibility metrics on the same training dataset.

The proposed model achieves comparable performance to deep learning SD&D algorithms under varying noise and reverberation conditions.

Abstract

Speech enhancement is crucial both for human and machine listening applications. Over the last decade, the use of deep learning for speech enhancement has resulted in tremendous improvement over the classical signal processing and machine learning methods. However, training a deep neural network is not only time-consuming; it also requires extensive computational resources and a large training dataset. Transfer learning, i.e. using a pretrained network for a new task, comes to the rescue by reducing the amount of training time, computational resources, and the required dataset, but the network still needs to be fine-tuned for the new task. This paper presents a novel method of speech denoising and dereverberation (SD&D) on an end-to-end frozen binaural anechoic speech separation network. The frozen network requires neither any architectural change nor any fine-tuning for the new task, as is usually required for transfer learning. The interaural cues of a source placed inside noisy and echoic surroundings are given as input to this pretrained network to extract the target speech from noise and reverberation. Although the pretrained model used in this paper has never seen noisy reverberant conditions during its training, it performs satisfactorily for zero-shot testing (ZST) under these conditions. It is because the pretrained model used here has been trained on the direct-path interaural cues of an active source and so it can recognize them even in the presence of echoes and noise. ZST on the same dataset on which the pretrained network was trained (homo-corpus) for the unseen class of interference, has shown considerable improvement over the weighted prediction error (WPE) algorithm in terms of four objective speech quality and intelligibility metrics. Also, the proposed model offers similar performance provided by a deep learning SD&D algorithm for this dataset under varying conditions of noise and reverberations. Similarly, ZST on a different dataset has provided an improvement in intelligibility and almost equivalent quality as provided by the WPE algorithm.

Figures