• The Journal of the Acoustical Society of Korea
• /
• v.42 no.6
• /
• pp.544-551
• /
• 2023

Speech enhancement used to improve the perceptual quality and intelligibility of noise speech has been studied as a method using a complex-valued spectrum that can improve both magnitude and phase in a method using a magnitude spectrum. In this paper, a study was conducted on how to apply attention mechanism to complex-valued spectrum-based speech enhancement systems to further improve the intelligibility and quality of noise speech. The attention is performed based on additive attention and allows the attention weight to be calculated in consideration of the complex-valued spectrum. In addition, the global average pooling was used to consider the importance of the feature map. Complex-valued spectrum-based speech enhancement was performed based on the Deep Complex U-Net (DCUNET) model, and additive attention was conducted based on the proposed method in the Attention U-Net model. The results of the experiments on noise speech in a living room environment showed that the proposed method is improved performance over the baseline model according to evaluation metrics such as Source to Distortion Ratio (SDR), Perceptual Evaluation of Speech Quality (PESQ), and Short Time Object Intelligence (STOI), and consistently improved performance across various background noise environments and low Signal-to-Noise Ratio (SNR) conditions. Through this, the proposed speech enhancement system demonstrated its effectiveness in improving the intelligibility and quality of noisy speech.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.2
• /
• pp.94-101
• /
• 2023

A deep neural network composed of encoders and decoders, such as U-Net, used for speech enhancement, concatenates the encoder to the decoder through skip-connection. Skip-connection helps reconstruct the enhanced spectrum and complement the lost information. The features of the encoder and the decoder connected by the skip-connection are incompatible with each other. In this paper, for complex-valued spectrum based speech enhancement, Self-Attention (SA) method is applied to skip-connection to transform the feature of encoder to be compatible with the features of decoder. SA is a technique in which when generating an output sequence in a sequence-to-sequence tasks the weighted average of input is used to put attention on subsets of input, showing that noise can be effectively eliminated by being applied in speech enhancement. The three models using encoder and decoder features to apply SA to skip-connection are studied. As experimental results using TIMIT database, the proposed methods show improvements in all evaluation metrics compared to the Deep Complex U-Net (DCUNET) with skip-connection only.

• The Journal of the Acoustical Society of Korea
• /
• v.43 no.2
• /
• pp.253-259
• /
• 2024

This paper proposes a new speech enhancement model based on a complex nested U-Net with a dual-branch decoder. The proposed model consists of a complex nested U-Net to simultaneously estimate the magnitude and phase components of the speech signal, and the decoder has a dual-branch decoder structure that performs spectral mapping and time-frequency masking in each branch. At this time, compared to the single-branch decoder structure, the dual-branch decoder structure allows noise to be effectively removed while minimizing the loss of speech information. The experiment was conducted on the VoiceBank + DEMAND database, commonly used for speech enhancement model training, and was evaluated through various objective evaluation metrics. As a result of the experiment, the complex nested U-Net-based speech enhancement model using a dual-branch decoder increased the Perceptual Evaluation of Speech Quality (PESQ) score by about 0.13 compared to the baseline, and showed a higher objective evaluation score than recently proposed speech enhancement models.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.1
• /
• pp.38-44
• /
• 2022

Speech enhancement is performed to improve intelligibility and quality of the noise-corrupted speech. In this paper, speech enhancement performance was compared using different loss functions in time and frequency domains. This study proposes a combination of loss functions to utilize advantage of each domain by considering both the details of spectrum and the speech waveform. In our study, Scale Invariant-Source to Noise Ratio (SI-SNR) is used for the time domain loss function, and Mean Squared Error (MSE) is used for the frequency domain, which is calculated over the complex-valued spectrum and magnitude spectrum. The phase loss is obtained using the sin function. Speech enhancement result is evaluated using Source-to-Distortion Ratio (SDR), Perceptual Evaluation of Speech Quality (PESQ), and Short-Time Objective Intelligibility (STOI). In order to confirm the result of speech enhancement, resulting spectrograms are also compared. The experimental results over the TIMIT database show the highest performance when using combination of SI-SNR and magnitude loss functions.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.1
• /
• pp.30-37
• /
• 2022

This paper compares and evaluates model performance from two perspectives according to the learning target and network structure for training Deep Neural Network (DNN)-based speech enhancement models in the frequency domain. In this case, spectrum mapping and Time-Frequency (T-F) masking techniques were used as learning targets, and a real network and a complex network were used for the network structure. The performance of the speech enhancement model was evaluated through two objective evaluation metrics: Perceptual Evaluation of Speech Quality (PESQ) and Short-Time Objective Intelligibility (STOI) depending on the scale of the dataset. Test results show the appropriate size of the training data differs depending on the type of networks and the type of dataset. In addition, they show that, in some cases, using a real network may be a more realistic solution if the number of total parameters is considered because the real network shows relatively higher performance than the complex network depending on the size of the data and the learning target.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.18 no.6
• /
• pp.1302-1309
• /
• 2014

In this paper, we propose a method to improve the performance of the direction-of-arrival (DOA) estimation of a speech source using a multiple signal classification (MUSIC)-based algorithm. Basically, the proposed algorithm utilizes a complex coefficient band pass filter to generate the narrow band signals for signal analysis. Also, reverberation component reduction and quadratic function-based response approximation in MUSIC spatial spectrum are utilized to improve the accuracy of DOA estimation. Experimental results show that the proposed method outperforms the well-known generalized cross-correlation (GCC)-based DOA estimation algorithm in the aspect of the estimation error and success rate, respectively.Abstract should be placed here. These instructions give you guidelines for preparing papers for JICCE.