Search

Because of their simple design structure, end-to-end deep learning (E2E-DL) models have gained a lot of attention for speech enhancement. A number of DL models have achieved excellent results in eliminating the background noise and enhancing the quality as well as the intelligibility of noisy speech. Designing resource-efficient and compact models during real-time processing is still a key challenge. In order to enhance the accomplishment of E2E models, the sequential and local characteristics of speech signal should be efficiently taken into consideration while modeling. In this paper, we present resource-efficient and compact neural models for end-to-end noise-robust waveform-based speech enhancement. Combining the Convolutional Encode-Decoder (CED) and Recurrent Neural Networks (RNNs) in the Convolutional Recurrent Network (CRN) framework, we have aimed at different speech enhancement systems. Different noise types and speakers are used to train and test the proposed models. With LibriSpeech and the DEMAND dataset, the experiments show that the proposed models lead to improved quality and intelligibility with fewer trainable parameters, notably reduced model complexity, and inference time than existing recurrent and convolutional models. The quality and intelligibility are improved by 31.61% and 17.18% over the noisy speech. We further performed cross corpus analysis to demonstrate the generalization of the proposed E2E SE models across different speech datasets. [ABSTRACT FROM AUTHOR]

Speech enhancement (SE) by transient noise suppression refers to the process of estimating the desired signal from a transient noise (TN)-corrupted speech signal. It has important applications in speaker verification and identification, voice-based biometric systems, hearing aids, video conferences, and many others. Enhancement of TN-corrupted signals is both important and challenging due to their high randomness, high short-time energy, and wide frequency domain distribution. This paper presents a novel approach to SE that leverages the power of sequential sparse nonnegative matrix factorization (SNMF), enhanced dictionary learning (DL), and Gini index (GI)-based fusion. We use a multi-step approach to enhance noisy speech. Our approach uses semi-supervised learning since the TN used for dictionary training is derived from the noisy signal using optimally-modified log-spectral amplitude (OMLSA) and is then purified. We involve both the dictionaries of the noisy signal and external clean speech in the DL process to merge the strengths of both dictionaries. Firstly, the dictionary of the noisy signal is obtained through SNMF and is then decomposed into speech-dominant and noise-dominant submatrices using semi-supervised learning. Then this speech-dominant submatrix is combined with the clean speech dictionary to construct an improved speech dictionary. This enhanced dictionary and the external clean speech dictionary are then used for the SE in the testing phase, resulting in two different initial estimated signals. To improve the accuracy of the initial estimates, we apply the GI to obtain the final estimate. Furthermore, the noisy signal's phase is also enhanced. A significant increase is observed in the results through different evaluation measures. [ABSTRACT FROM AUTHOR]

Transient interferences such as keystrokes, mouse clicks and hammering pose a significant challenge in the single channel speech enhancement due to their abrupt and non-continuous nature. Traditional noise suppression algorithms and even many non-stationary noise reduction algorithms do not adequately suppress transient interference. Therefore, in this work, we propose a semi-supervised single channel transient noise suppression method to effectively suppress the transient interference without significant audible distortion. The proposed algorithm consists of training and testing stages. In the training stage, the proposed technique first uses the optimally modified-log spectral amplitude (OMLSA) estimator to estimate the transient noise from the noisy speech signal. After that, we eliminate the residual speech components from the estimated noise obtained from OMLSA based on the correlation coefficient, by taking correlation between the estimated noise with the available clean speech data from the dataset passed through the voice activity detector for silence zones removal. Afterwards, we use this noise for training the noise dictionary in sparse non-negative matrix factorization. Clean speech data is used for speech dictionary training. In the enhancement stage, the dictionaries are fixed and concatenated, to obtain the corresponding activation matrices. The clean speech dictionary and the corresponding weight matrix are used to reconstruct the estimated speech. The experimental results reveal that the proposed algorithm provided better performance compared to other existing algorithms in the speech quality evaluation metrics. [ABSTRACT FROM AUTHOR]

In real-time applications, the aim of speech enhancement (SE) is to achieve optimal performance while ensuring computational efficiency and near-instant outputs. Many deep neural models have achieved optimal performance in terms of speech quality and intelligibility. However, formulating efficient and compact deep neural models for real-time processing on resource-limited devices remains a challenge. This study presents a compact neural model designed in a complex frequency domain for speech enhancement, optimized for resource-limited devices. The proposed model combines convolutional encoder–decoder and recurrent architectures to effectively learn complex mappings from noisy speech for real-time speech enhancement, enabling low-latency causal processing. Recurrent architectures such as Long-Short Term Memory (LSTM), Gated Recurrent Unit (GRU), and Simple Recurrent Unit (SRU), are incorporated as bottlenecks to capture temporal dependencies and improve the performance of SE. By representing the speech in the complex frequency domain, the proposed model processes both magnitude and phase information. Further, this study extends the proposed models and incorporates attention-gate-based skip connections, enabling the models to focus on relevant information and dynamically weigh the important features. The results show that the proposed models outperform the recent benchmark models and obtain better speech quality and intelligibility. The proposed models show less computational load and deliver better results. This study uses the WSJ0 database where clean sentences from WSJ0 are mixed with different background noises to create noisy mixtures. The results show that STOI and PESQ are improved by 21.1% and 1.25 (41.5%) on the WSJ0 database whereas, on the VoiceBank+DEMAND database, STOI and PESQ are improved by 4.1% and 1.24 (38.6%) respectively. The extension of the models shows further improvement in STOI and PESQ in seen and unseen noisy conditions. • Proposed Deep Models with LSTM, GRU, and SRU Model for Speech Enhancement. • To Capture Temporal Sequences, LSTM, GRU, and SRU are applied to extracted Spectral Features. • Attention Gates are added to the Skip Connections to Focus on Important Spectral Features. • The Proposed SE Models are Evaluated on two datasets (WSJ0 and VoiceBank+DEMAND). [ABSTRACT FROM AUTHOR]