Introduction

In typical day-to-day scenarios such as classroom lectures and auditorium speeches, we encounter various kinds of noises from appliances like fan and air conditioner and surrounding areas in these speech signals. These noises makes listeners difficult to understand and interpret the information in given audio signals.

Problem Statement

Most of the noises in indoor speech signals can be typically categorised into two kinds - additive noises and unwanted frequency noises. Therefore, we design a framework to postprocess these speech signals to suppress the forementioned types of noise without too much degradation of the required information.

Proposed Approach

We propose a two-block framework to suppress additive noises and unwanted frequencies.

• Band-pass filtering
  • In this block, a desired frequency band is determined using domain knowledge depending on the speech signal.
  • A band pass filter will be designed and original audio signal will be filtered using this filter in the fourier domain.
  • Finally, the signal will be constructed back into time domain.
• Spectral Subtraction
  • Assuming additive noise model, the original signal can be expressed in given format.
  • The speech signal is segmented into multiple parts with N samples in each part.
  • Each speech segment is windowed using Hamming window before N-point Discrete fourier transformed is applied (Window application reduces the effect of discontinuties at segment’s endpoints.)
  • The time-averaged noise spectrum is derived from the periods where only the noise is present.
  • The noise suppressed spectrum estimate is calculated by subtracting the noise spectrum from the actual spectrum through sliding window.
  • Finally, the processed spectrum is transformed into time domain using overlap-add method.

  

  Original Signal	Processed Signal

  Original Spectrum	Processed Spectrum

Results

sample results can be found in samples directory.

Conclusion

To conclude, these techniques are limited to speech signals with only additive noises and unwanted frequencies with certain valid assumptions. In the future, this project can be extented to real-time processing techniques for noise cancellation.