Real life demonstration of noise suppression technology is quite challenging since you need to find a way to simulate naturally-sounding noise in different physical setups – conference room, video conference, in front of a big audience, etc.
There is one particular test, well known in automotive industry, to simulate high-wind noise. One person talks to the microphone (mic) while another person blows directly into the microphone simply using their mouth or maybe a pen/straw – to create even more air pressure.
In this article we will explain why this test in most cases doesn’t make sense for MEMS microphones.
Internal structure of MEMS mic
So again, let’s imagine a person is speaking loudly while another person is blowing into the mic. You record the audio coming from the mic and play it back. What you see is that when blowing happens the voice cannot be heard.
Now the question is: does the mic capture the voice at all? And this is a really interesting question. How can a mic not capture the voice?
Apparently it might not capture it due to its physical limitations.
MEMS mics have a thin membrane that receives the vibration from sound waves. The motion created by the vibrating membrane carries over to the voice coil and as that voice coil moves in its magnetic field, it creates a unique electric signal depending on the types of vibrations picked up by the membrane.
Now, membrane has a pressure limit. If there is too much air pressure the signal created by it will fluctuate too much and will start clipping – thus resulting in distorted output.
Example of “blowing into mic” audio
Below is an example waveform of an audio showing the use case we have described: one person blows into the mic while another person is speaking.
You can easily distinguish the parts where the person blows into the mic.
When we look more closely into the part where blowing happens we see that there is a clipping between -1 to 1. When clipping happens some information of the input signal can be lost.
This is why we can’t hear human voice in the blowing segments.
The takeaway here is that this test makes sense only when the microphone’s physical characteristics allow for handling the air pressure created by “blowing into it”. From what we see the mics integrated into mobile devices or laptops are not designed for this and doing such tests on them isn’t accurate.
Can such high wind noise be cleaned by Noise Suppression?
Now the question remaining is this: if the mic is able to handle the pressure and allows the other person’s voice go through it – can the wind noise be suppressed at such low SNR (speech to noise ratio) condition?
The answer is yes and the spectrograms below demonstrate that.
We used software to simulated various low SNR conditions with wind and speech and then used 2Hz Technology to suppress noise on these audio streams. As you see from the spectrograms below our Deep Neural Network is still able to recognize human speech and extract it from the audio. The results worsen as the SNR gets lower. In -20db SNR the output is less intelligible.
Top spectrogram is the noisy speech.
Bottom one is the original non-noisy speech.
2Hz processed one is in the middle.
- “Blowing into the mic” test doesn’t always make sense and you need to know the physical characteristics of the mic before doing such test. Typically it’s not accurate to do them on phones and laptops.
- If the mic is good enough and the noise suppression algorithm is robust it’s possible to suppress very high wind noise and extract the human speech out of it