Using Transformer Models to Separate and Reduce Noise in Speaker Voice in Noisy Conference Scenarios

Nowadays, online meetings have become an important way for people to communicate at work. With the application popularization of this meeting form and the emergence of various new functions, people's requirements for voice quality are also continuously rising. Especially in the case of noisy backgrounds and multiple speakers, obtaining clear speech signals and accurately separated speech is still a problem that needs to be solved today. To solve this problem, this study proposes a multi-speaker speech separation and noise reduction system based on the pre-trained model of the SpeechBrain library. This system combines the SpectralMaskEnhancement model and the SepformerSeparation model, respectively used for speech denoising and multi-speaker speech separation in complex noise environments. By combining the advantages of the two models, the separation and denoising effects are effectively improved. The experiment used the LibriMix dataset, and the results showed that this method achieved 18.0 dB in signal-to-noise ratio improvement (SI-SNRi) and 18.3 dB in source-to-distortion ratio improvement (SDRi), significantly improved compared to the traditional Wavesplit and the separate Sepformer model. Through experimental comparison, the fusion model outperformed conventional methods in denoising and speech separation performance, especially in complex scenarios with multi-speaker overlap and strong background noise. The method in this study can effectively improve the voice quality of the conference in noisy backgrounds and multi-speaker scenarios, and play a role in other audio fields.