group-B-audio-analyzer 0.0.3

A Python package for audio analysis and .wav file classification.

Audio analysis The customer request is to create a tool to denoise audio tracks and classify them into three classes: human voices, musical instruments, and others.

To denoise the audio you should not use any ML algorithm but you are suggested to adopt a specific Wiener filter..

To classify the audio tracks you can perform the audio analysis in the Fourier Domain (applying FFT to the original signal).

Use visualisation to show the differences among the three classes of audio tracks.

Once finished with this task, you can compare this classification with the one obtainable with a Convolution Neural Network (CNN) applied to the images obtained from padding the audio tracks. (For this step you can take advantage of PyTorch or Keras).

The data To structure and test the first class of your audio data analysis pipeline, the denoiser, a possibility is to use the clean subset of audio tracks in the Freesound mono audio track dataset, DBR-dataset, first adding white random noise to each track, and then, trying to remove the white noise from the signal.

Using the same dataset you can also test your classifier.

Once you have tested the audio analysis pipeline on this dataset, create a small data set yourself, recording similar audio tracks and paying attention to the standardization of the input data.

Useful tools To convert the audio tracks into signals treatable with Scipy you can use the PyAudio library.

To construct the Wiener filter you can use the Scipy and Numpy libraries.