State-of-the-art speech enhancement techniques use deep neural networks to improve distorted speech signals. These networks employ an encoder-decoder framework, with the encoder extracting features from the input signal. Our research suggests using discrete wavelet transform (DWT) features as an alternative to existing methods. DWT features work well with time-domain features and improve performance in the adaptive FullSubNet+ framework. This study proposes using wavelet packet decomposition (WPD) to extract features and discarding sub-band WPD features that harm performance. Our method outperforms the original A-FSN in objective speech metrics, making it a promising speech enhancement framework.

The Study of Improving the Adaptive FullSubNet+ Speech Enhancement Framework with Selective Wavelet Packet Decomposition Sub-Band Features

Speech separation has recently made significant progress thanks to the fine-grained vision used in time-domain methods. However, several studies have shown that adopting Short-Time Fourier Transform (STFT) for feature extraction could be beneficial when encountering harsher conditions, such as noise or reverberation. Therefore, we propose a magnitude-conditioned time-domain framework, ConSep, to inherit the beneficial characteristics. The experiment shows that ConSep promotes performance in anechoic, noisy, and reverberant settings compared to two celebrated methods, SepFormer and BiSep. Furthermore, we visualize the components of ConSep to strengthen the advantages and cohere with the actualities we have found in preliminary studies.

ConSep: a Noise- and Reverberation-Robust Speech Separation Framework by Magnitude Conditioning

An important more recent thread of speech enhancement work is to utilize fine-grinded local spectral patterns with sub-band processing that complement full-band features nicely. To extend the efficacy of sub-band spectral information, we propose Adaptive-FSN, a fully convolutional real-time speech enhancement framework, to dynamically acquire a sub-band embedding within a wide range of sub-band frequencies. We exploit an adaptive subband encoder to portray sub-band processing that encapsulates a wide range of sub-band units. Then we build this effective sub-band embedding with a Conformer-based structure and multi-view attention. As for the full-band features, we make use of the FullSubNet+ architecture with its full-band extractor to get global spectral information. Finally, a Conformer-based fusion model combines the above information sources to predict the complex ideal ratio mask (cIRM). Experimental results on the VoiceBank-DEMAND benchmark task reveal that this novel framework outperforms FullSubNet+ by promoting the quality of processed utterances and reducing the implementation complexity for faster real-time computation.

Adaptive-FSN: Integrating Full-Band Extraction and Adaptive Sub-Band Encoding for Monaural Speech Enhancement

In recent years, deep neural network (DNN)-based time-domain methods for monaural speech separation have substantially improved under an anechoic condition. However, the performance of these methods degrades when facing harsher conditions, such as noise or reverberation. Although adopting Short-Time Fourier Transform (STFT) for feature extraction of these neural methods helps stabilize the performance in non-anechoic situations, it inherently loses the fine-grained vision, which is one of the particularities of time-domain methods. Therefore, this study explores incorporating time and STFT-domain features to retain their beneficial characteristics. Furthermore, we leverage a Bi-Projection Fusion (BPF) mechanism to merge the information between two domains. To evaluate the effectiveness of our proposed method, we conduct experiments in an anechoic setting on the WSJ0-2mix dataset and noisy/reverberant settings on WHAM!/WHAMR! dataset. The experiment shows that with a cost of ignorable degradation on anechoic dataset, the proposed method manages to promote the performance of existing neural models when facing more complicated environments.

Bi-Sep: A Multi-Resolution Cross-Domain Monaural Speech Separation Framework

Osteoporosis is a progressive metabolic bone disorder characterised by a decline in bone mineral density (BMD). Some previous studies have reported a controversial relationship between uric acid and osteoporosis. This cross-sectional study aimed to assess the association between serum uric acid levels and BMD in older adults from Taiwan. Data from participants aged ≥ 60 years were collected from 2008 to 2018. Furthermore, the participants were classified based on uric acid level quartiles. Regression models were used to assess the association between uric acid levels and bone health, including BMD values and risk of at least osteopenia. Crude and adjusted models of potential confounders, including age, sex and body mass index (BMI), were used. Compared with the first quartile of uric acid levels, the odds ratios for osteoporosis decreased in the higher uric acid level groups after adjustment for age, sex and BMI. The boxplot analysis showed that BMD values were higher in the groups with higher uric acid levels; moreover, the results of the multivariable linear regression model showed a consistent pattern. Notably, uric acid levels were positively correlated with BMD values. Higher uric acid levels in the elderly population might decrease the risk of at least osteopenia. As opposed to the anti-hyperuricemic policy for younger adults with a relatively lower risk of osteoporosis, BMD evaluation and urate-lowering therapy, goal adjustments should be considered for older adults with lower uric acid levels.

https://www.mdpi.com/1660-4601/20/4/3448/pdf?version=1676514565

Association between Serum Uric Acid Levels and Bone Mineral Density in Taiwanese Elderly Population

This study investigates constructing low-distortion utterances to benefit downstream automatic speech recognition (ASR) systems at the front-end stage based on a speech enhancement (SE) network. With the dual-path Transformer network (DPTNet) as the SE archetype, we make effective use of short-time discrete cosine transform (STDCT) features to infer the respective mask-estimation network. Furthermore, we seek to jointly optimize the spectral-distance loss and the perceptual loss for the training of the model components of our proposed SE model so as to enhance the input utterances without introducing significant distortion. Extensive evaluation experiments are conducted on the VoiceBank-DEMAND and VoiceBank-QUT tasks, containing stationary and non-stationary noises, respectively. The corresponding results show that the proposed SE method yields competitive perceptual metric scores on SE but significantly lower word error rates (WER) on ASR in relation to several top-of-the-line methods. Notably, the proposed SE method works remarkably well on the VoiceBank-QUT ASR task, thereby confirming its excellent generalization capability to unseen scenarios.

Kuan-Hsun Ho

Papers

Adaptive-FSN: Integrating Full-Band Extraction and Adaptive Sub-Band Encoding for Monaural Speech Enhancement

Bi-Sep: A Multi-Resolution Cross-Domain Monaural Speech Separation Framework

Association between Serum Uric Acid Levels and Bone Mineral Density in Taiwanese Elderly Population

ConSep: a Noise- and Reverberation-Robust Speech Separation Framework by Magnitude Conditioning

Investigating Low-Distortion Speech Enhancement with Discrete Cosine Transform Features for Robust Speech Recognition