In my last post, I mentioned how I wanted to work on both fine-tuning my product and looking into different methods of managing background noise. Since last week's post, I have made slow progress in the fine-tuning of my product and intend to continue down this path going forward. Along with this, I have continued my research into the most effective ways of handling background noise. In my research, I have found the two key issues to consider when eliminating background noise. These issues are (i) isolating the frequencies of the background noise and (ii) eliminating this noise without damaging any necessary sound data. Since discovering this I have focused my research mainly on the first of these two issues being that I believe it to be the more complicated one. As far as the fine-tuning adjustments to my product go I have isolated one particular section of my code to be a potential source of error for the failing trials. This section contains the weights I assigned to certain pieces of information that are collected from any given phoneme in the spectrogram for the voice recording. For example, in vowel phonemes, I have weighted or increased the importance of, the frequencies of each formant in the middle of the isolated vowel phoneme. It is information weights like this one that make it possible for one phoneme to be confused for another and if this process were to continue, one name for another. For this reason, I will be focusing on specifically fine-tuning these weights to try and achieve a higher rate of success from now on.
Comments