Applause-Noise-Reduction

Goal is to clean concert audio record from applauses.

Datasets

• MUSDB18-HQ, for high quality music audios
• Free Sound, for random concert noises

Defining Ideal Masks for clean audio estimtaion

Pre-proceesing

• downsample all music audio from 44.kHz to 16kHz
• make mono from stereo by averaging left and right audio signals

• splitting audio into 5 sec intervals with 2.5 sec overlap

• get STFT features

Creating a noisy dataset from audio and noises datasets

• adding random noises to audio with different SNR's

Estimating the mask of clean audio signal

• using Bi-GRU model

Post-processing

• from STFT going back to audio signal
• adding phase info from noisy audio to estimated audio