Audio Detection testing

Two different tests were ran to verify that audio detection is working as intended.

The first suite of tests were ran by intermittently speaking at various distances from the microphone, at a maximum range of about 4 feet. There would be pauses between speaking in order to test the deletion functionality of extraneous files as well. Minimal background noise was allowed in this test.

Without background noise, the application ran almost perfectly. Unfortunately, Google Speech does not register screams as speech.

The following is the results of the testing done with the microphone close to speakers playing a sound clip of a house fire, provided by this video. We also had voices go somewhat lower, attempting to mimic that of a person who has been breathing in smoke.

Unfortunately, there were many false negatives, with zero speech detection whatsoever. The file audio6.mp3 that can be seen in the screenshot is the result of closing the detection application early. It contains what would have been sent to the Google Speech client for detection. Listening to it, the fire sound was incredibly overwhelming. This led us to believe that the fire sound we have may be too loud, and our microphone may not be good enough for proper testing.

To test this hypothesis, we ran the detection test again with the same background sound, but with file deletion disabled.

Out of four files that were created, only one of them had the audio properly detected. The first three files had speech attempting to mimic that of a person who has been in fire and breathing smoke again. The last file, which was transcribed nearly accurately, had speech yelling "Help, help me!". Google Speech was able to transcribe the second "help me!" but not the first "Help".

In the future, to attempt to counteract this, we will try to apply audio filters, which would hopefully clean the audio enough to the point the speech can be detected again.