Filtering is the art of removing parts of a signal that are not required and retaining those parts that are required. The four main kinds of filter are shown in Figure 34. For example, a notch filter will remove the noise at the 1200 Hz peak.
You should still have Interactive 5 open in a separate tab, containing your recording of the word ‘Yes’ with an added noise tone of 1200 Hz. Select the ‘Notch’ button and use the ‘Right cutoff’ and ‘Left cutoff’ sliders to create a notch filter, as shown in Figure 34(a). Then click on the ‘Apply filters’ button next to the ‘Notch’ button.
Your result should be similar to that shown in Figure 35. As you can see, the noise has been removed without doing too much damage to the original signal.
Click on the play button (the triangle) at the bottom left of the interactive to play the filtered signal.
- a.Was the noise removed as you listened to your signal?
- b.Was the signal damaged, giving low-fidelity reproduction?
- a.You should have found that the noise added to the signal was removed completely.
- b.The filtered version of your signal should sound exactly the same as the original, or very close to it.
If this experiment did not work well for you, please try it again. It should work robustly and the results can be remarkable.
Click on the ‘Bandpass’ button and then click on the ‘Apply filters’ button. What do you see and hear now?
In the time-domain representation of the filtered sound, you are likely to see only a solid band of noise. When you click on the play button, you should hear only the noise – your ‘yes’ should have been completely filtered out.
Click on the ‘Low-pass’ button and then click on the ‘Apply filters’ button. What do you see and hear now?
In the time-domain representation of the filtered sound, you should see everything below the right cutoff filtered in, including the noise. This time the ‘yes’ will be muffled because the higher frequencies were lost.
Click on the ‘High-pass’ button and then click on the ‘Apply filters’ button. What do you see and hear now?
In the time-domain representation of the filtered sound, you should see everything above the left cutoff filtered in, including the noise. This time the ‘yes’ will be muffled because the lower frequencies were lost.
Select ‘High-pass’ and change the left cutoff to about 3 kHz. Click on ‘Apply filters’. Does this filter in or out any discernible part of the ‘yes’ sound?
The high-pass filter should remove the noise and most of the ‘yes’ sound, but you may find that the ‘s’ part can still be heard clearly. This supports the earlier conjecture that the tail of the ‘yes’ sound is the relatively high-frequency ‘sss’ as the word ends.
When you first looked at the recordings of ‘yes’ and ‘no’ on this course, it was noted that they have different shapes in the time domain and that this might be useful for speech recognition. As can be seen in Figure 36, ‘yes’ and ‘no’ also have different patterns in the frequency domain and this too is useful for speech recognition.
You can now close the interactive.