Improving deep speech denoising by Noisy2Noisy signal mapping

Alamdari, Nasim; Azarang, Arian; Kehtarnavaz, Nasser

doi:10.1016/j.apacoust.2020.107631

Cited by 35 publications

(19 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…x m < l a t e x i t s h a _ b a s e = " X z To overcome such limitation, some research has attempted to train a DNN without the clean target [16][17][18]. An interesting strategy aiming at this goal is Noise-target Training (NeTT) as shown in Fig.…”

Section: Dnnmentioning

confidence: 99%

“…An interesting strategy aiming at this goal is Noise-target Training (NeTT) as shown in Fig. 1 training target is mixture of speech and noise, and the input signal is simulated by using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some using the same clean speech and some [19], and was later applied to speech enhancement [18]. In Noise2Noise, pairs of noisy signals that consist of different noises and exactly the same clean target are utilized for training.…”

Section: Dnnmentioning

confidence: 99%

“…However, it is not applicable to audio applications because it is impossible to observe multiple noisy signals with exactly the same speech signal. Therefore, the previous research has utilized clean speech to simulate a pair of noisy signals [18] as in Fig. 1 (b), which inherits the limitation of Clean-target Training.…”

Section: Dnnmentioning

confidence: 99%

See 2 more Smart Citations

Noisy-target Training: A Training Strategy for DNN-based Speech Enhancement without Clean Speech

Fujimura

Yatabe

Miyazaki

2021

2021 29th European Signal Processing Conference (EUSIPCO)

View full text Add to dashboard Cite

Deep neural network (DNN)-based speech enhancement ordinarily requires clean speech signals as the training target. However, collecting clean signals is very costly because they must be recorded in a studio. This requirement currently restricts the amount of training data for speech enhancement less than 1/1000 of that of speech recognition which does not need clean signals. Increasing the amount of training data is important for improving the performance, and hence the requirement of clean signals should be relaxed. In this paper, we propose a training strategy that does not require clean signals. The proposed method only utilizes noisy signals for training, which enables us to use a variety of speech signals in the wild. Our experimental results showed that the proposed method can achieve the performance similar to that of a DNN trained with clean signals.

show abstract

Section: Dnnmentioning

confidence: 99%

Section: Dnnmentioning

confidence: 99%

See 1 more Smart Citation

Noisy-target Training: A Training Strategy for DNN-based Speech Enhancement without Clean Speech

Fujimura

Yatabe

Miyazaki

2021

2021 29th European Signal Processing Conference (EUSIPCO)

View full text Add to dashboard Cite

show abstract

“…The main innovations and contributions of this research are as follows: (1) The EEMD was introduced into MSPCA denoising to achieve improved signal decomposition adaptability compared to wavelet decomposition. (2) The Hankel matrix was used to transform a single EEMD component into high-dimensional data that are suitable for conducting PCA denoising on only one signal. (3) Soft thresholding was selected for further denoising of the PCA-denoised component.…”

Section: Introductionmentioning

confidence: 99%

“…Consequently, to improve signal detection, classification, and phase arrival picking, it is necessary to remove unwanted noise. Currently, signal denoising is widely used in the fields of medical signals [ 1 ], speech signals [ 2 ], mechanical diagnosis [ 3 ], structural health monitoring [ 4 , 5 , 6 ], and seismic monitoring networks [ 7 , 8 , 9 ]; however, this study focused specifically on microseismic (MS) signal denoising.…”

Section: Introductionmentioning

confidence: 99%

EEMD and Multiscale PCA-Based Signal Denoising Method and Its Application to Seismic P-Phase Arrival Picking

Peng

Guo

Shang

2021

Sensors

View full text Add to dashboard Cite

Signal denoising is one of the most important issues in signal processing, and various techniques have been proposed to address this issue. A combined method involving wavelet decomposition and multiscale principal component analysis (MSPCA) has been proposed and exhibits a strong signal denoising performance. This technique takes advantage of several signals that have similar noises to conduct denoising; however, noises are usually quite different between signals, and wavelet decomposition has limited adaptive decomposition abilities for complex signals. To address this issue, we propose a signal denoising method based on ensemble empirical mode decomposition (EEMD) and MSPCA. The proposed method can conduct MSPCA-based denoising for a single signal compared with the former MSPCA-based denoising methods. The main steps of the proposed denoising method are as follows: First, EEMD is used for adaptive decomposition of a signal, and the variance contribution rate is selected to remove components with high-frequency noises. Subsequently, the Hankel matrix is constructed on each component to obtain a higher order matrix, and the main score and load vectors of the PCA are adopted to denoise the Hankel matrix. Next, the PCA-denoised component is denoised using soft thresholding. Finally, the stacking of PCA- and soft thresholding-denoised components is treated as the final denoised signal. Synthetic tests demonstrate that the EEMD-MSPCA-based method can provide good signal denoising results and is superior to the low-pass filter, wavelet reconstruction, EEMD reconstruction, Hankel–SVD, EEMD-Hankel–SVD, and wavelet-MSPCA-based denoising methods. Moreover, the proposed method in combination with the AIC picking method shows good prospects for processing microseismic waves.

show abstract

Audio Denoising Using Deep Neural Networks

Mohammed

Radhika²

2022

Intelligent Data Communication Technologies and Internet of Things

View full text Add to dashboard Cite

Improving deep speech denoising by Noisy2Noisy signal mapping

Cited by 35 publications

References 31 publications

Noisy-target Training: A Training Strategy for DNN-based Speech Enhancement without Clean Speech

Noisy-target Training: A Training Strategy for DNN-based Speech Enhancement without Clean Speech

EEMD and Multiscale PCA-Based Signal Denoising Method and Its Application to Seismic P-Phase Arrival Picking

Audio Denoising Using Deep Neural Networks

Contact Info

Product

Resources

About