I am trying a simple tutorial as mentioned here in tensor flow tutorials. I ran streaming accuracy and the final output look like this.
53.0% matched, 51.0% correctly, 2.0% wrongly, 0.0% false positives.
Its mentioned that it outputs information about the number of words correctly matched, how many were given the wrong labels, and how many times the model triggered when there was no real word spoken.
If i add 51%+2%+0% doesn't add up to 100%. What is the accuracy ? It is 51/53 *100 or is it just 53 %.
After physically checking the detected words. I figured that accuracy is just 51%. It seems like lot of words were not detected may be because of the suppression period followed after the detection of the word or noise.