Date of Award
Doctor of Philosophy (PhD)
Electrical and Computer Engineering (Holcomb Dept. of)
Melissa C. Smith
Speech enhancement is the task that aims to improve the quality and the intelligibility of a speech signal that is degraded by ambient noise and room reverberation. Speech enhancement algorithms are used extensively in many audio- and communication systems, including mobile handsets, speech recognition, speaker verification systems and hearing aids. Recently, deep learning has achieved great success in many applications, such as computer vision, nature language processing and speech recognition. Speech enhancement methods have been introduced that use deep-learning techniques, as these techniques are capable of learning complex hierarchical functions using large-scale training data. This dissertation investigates the deep learning based speech enhancement and its application to robust Automatic Speech Recognition (ASR).
We start our work by exploring generative adversarial network (GAN) based speech enhancement. We explore the techniques to extract information about the noise to aid in the reconstruction of the speech signals. The proposed framework, referred to as ForkGAN, is a novel general adversarial learning-based framework that combines deep-learning with conventional noise reduction techniques. We further extend ForkGAN to M-ForkGAN, which integrates feature mapping and mask learning into a unified framework using ForkGAN. Another variant of ForkGAN, named S-ForkGAN, operates on spectral-domain features, which could directly apply to ASR. Systematic evaluations demonstrate the effectiveness of the proposed approaches.
Then, we propose a novel multi-stage learning speech enhancement system. Each stage comprises a self-attention (SA) block followed by stacks of temporal convolutional network (TCN) blocks with doubling dilation factors. Each stage generates a prediction that is refined in a subsequent stage. A fusion block is inserted at the input of later stages to re-inject original information. Moreover, we design several multi-scale architectures with perceptual loss. Experiments show that our proposed architectures can achieve the state of the art performance on several public datasets.
Recently, modeling to learn the acoustic noisy-clean speech mapping has been enhanced by including auxiliary information such as visual cues, phonetic and linguistic information, and speaker information. We propose a novel speaker-aware speech enhancement (SASE) method that extracts speaker information from a clean reference using long short-term memory (LSTM) layers, and then uses a convolutional recurrent neural network (CRN) to embed the extracted speaker information. The SASE framework is extended with a self-attention mechanism. It is shown that a few seconds of clean reference speech is sufficient, and that the proposed SASE method performs well for a wide range of scenarios.
Even though speech enhancement methods that are based on deep learning have demonstrated state-of-the-art performance when compared with conventional methodologies, current deep learning approaches heavily rely on supervised learning, which requires a large number of noisy- and clean-speech sample pairs for training. This is generally not practical in a realistic environment. One cannot simultaneously obtain both noisy and clean speech samples. Thus, most speech enhancement approaches are trained with simulated speech and clean targets. In addition, it would be hard to collect large-scale dataset for the low-resource languages. We propose a novel noise-to-noise speech enhancement (N2N-SE) method that addresses the parallel noisy-clean training data issue, we leverage signal reconstruction techniques by only using corrupted speech. The proposed N2N-SE framework includes a noise conversion module that is an auto-encoder that learns to mix noise with speech, and a speech enhancement module, that learns to reconstruct corrupted speech signals.
In addition to additive noise, speech is also affected by reverberation, which is caused by the attenuated and delayed reflections of sound waves. These distortions, particularly when combined, can severely degrade speech intelligibility for human listeners and impact applications, e.g., automatic speech recognition (ASR) and speaker recognition. Thus, effective speech denoising and dereverberation will benefit both speech processing applications and human listeners. We investigate the deep-learning based approaches for both speech dereverberation and speech denoising using the cascade Conformer architecture. The experimental results show that the proposed cascade Conformer can be effective to suppress the noise and reverberation.
Lin, Ju, "Deep Learning Based Speech Enhancement and Its Application to Speech Recognition" (2021). All Dissertations. 2939.