FullSubNet+: Channel Attention Fullsubnet with Complex Spectrograms for Speech Enhancement
FullSubNet+: полносвязная FullSubNet с механизмом канального внимания и комплексными спектрограммами для улучшения речи
2022-04-27
SCID: 54.1/e4nvdxz7
Discuss with AI
DNS ChallengeFullSubNet+channel attentioncomplex spectrogramstemporal convolutional network
Figures from the paper
Abstract (AI)
Previously proposed FullSubNet has achieved outstanding performance in Deep Noise Suppression (DNS) Challenge and attracted much attention. However, it still encounters issues such as input-output mismatch and coarse processing for frequency bands. In this paper, we propose an extended single-channel real-time speech enhancement framework called FullSubNet+ with following significant improvements. First, we design a lightweight multi-scale time sensitive channel attention (MulCA) module which adopts multi-scale convolution and channel attention mechanism to help the network focus on more discriminative frequency bands for noise reduction. Then, to make full use of the phase information in noisy speech, our model takes all the magnitude, real and imaginary spectrograms as inputs. Moreover, by replacing the long short-term memory (LSTM) layers in original full-band model with stacked temporal convolutional network (TCN) blocks, we design a more efficient full-band module called full-band extractor. The experimental results in DNS Challenge dataset show the superior performance of our FullSubNet+, which reaches the state-of-the-art (SOTA) performance and outperforms other existing speech enhancement approaches.
Key Findings
1
FullSubNet+ extends FullSubNet into a single-channel, real-time speech enhancement framework addressing input-output mismatch and coarse frequency-band processing.
2
FullSubNet+ uses noisy-speech magnitude, real, and imaginary spectrograms as inputs, enabling fuller exploitation of phase information.
3
On the DNS Challenge dataset, FullSubNet+ achieves state-of-the-art performance and outperforms existing speech enhancement approaches.
4
Replacing LSTM layers with stacked temporal convolutional network blocks produces a more efficient full-band extractor.
5
The lightweight multi-scale time-sensitive channel attention (MulCA) module combines multi-scale convolution and channel attention to emphasize discriminative frequency bands for noise reduction.
Research Object
single-channel noisy speech enhancement system (FullSubNet+)
Research Subject
real-time noise reduction performance through discriminative frequency-band processing and magnitude–phase spectrogram modeling
Publication Details
Publication Date
2022-04-27
Journal
Publisher
ISSN
Cited by
127
Access Type
Author Information
Download PDF
Subscribe to digest