Related Experiment Video
Updated: May 12, 2025
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
462
CROSS-DOMAIN SPEECH ENHANCEMENT WITH A NEURAL CASCADE ARCHITECTURE
Heming Wang1, DeLiang Wang1,2
1Department of Computer Science and Engineering, The Ohio State University, USA.
Summary
This study introduces a novel cascade neural network for monaural speech enhancement. The approach uses multiple speech representations to significantly improve speech quality and intelligibility by reducing background noise.
Area of Science:
- Signal Processing
- Artificial Intelligence
- Speech Technology
Background:
- Monaural speech enhancement is challenging due to the presence of background noise in single-channel audio.
- Existing methods often struggle to effectively separate speech from noise, impacting communication clarity.
Purpose of the Study:
- To propose a novel cascade neural network architecture for monaural speech enhancement.
- To leverage multiple speech representations for improved noise suppression.
Main Methods:
- A three-module cascade architecture processing speech in spectral magnitude, waveform, and complex spectrogram domains.
- Simultaneous optimization of all modules using a triple-domain loss function during training.
- Utilizing the WSJ0 SI-84 corpus for experimental evaluation.
Main Results:
- The proposed cascade architecture achieves superior speech enhancement results.
- Demonstrated substantial outperformance compared to previous baseline methods.
- Significant improvements in both objective and subjective measures of speech quality and intelligibility.
Conclusions:
- The novel cascade architecture effectively suppresses background noise in monaural speech.
- The multi-domain representation approach enhances speech quality and intelligibility.
- This method offers a promising advancement in single-channel speech enhancement technology.
