In my old blog I have shown the progress and process performed to train a CNN model able to understand specific voice commands.
One of the audio processing methods that have been employing is that of extracting the MEL Cepstral coefficients from the spectrogram of all audio samples. However, these coefficients contain the actual pattern of my voice, therefore, it would be important to check if the device can be used by someone else, that has a different voice.
I have appointed a friend of mine as the test subject, with a completely different pitch, tone and accept compared to me, to record different samples and test the accuracy of the best model found in the last blog.
I have run 2 tests.
First test:
My friend recorded 1 command for air quality, 1 command for health, 1 for air quality, and 2 for the precipitation forecast and 2 for weahter classification.
Here we can see the actual outcomes:
From here we can see that the actual accuracy is 57% in understanding someone else's voice.
Second test:
For the second run, I have recorded 7 commands myself, idetifiable with the starting "c" in the next table, and 7 new by my friend identifiable by "v".
And here we can see the accuracy and Feature Explorer:
We can notice that the model has 100% recognition for my commands, and only 4 out of 7, once again, were recognised from my friend. Resulting again 57% chance of someone else's voice being recognised.
It is important to note that this test was only to gauge an understanding of the system, and a more in depth test with 100-200 samples from a different person will be used to check the actual accuracy.
Kommentare