I'm working on a site which will take live audio input and perform AI analysis. The model requires an fft data plot as input - which I can create using the AnalyserNode from the JS audio API (get frequency method).
Having plotted this data - whilst the technique picks up a little on the dominant frequencies of the audio sample - it tends to register large decibel values for many frequencies across the board (0-20k Hz) for sounds which should really produce an uneven distribution weighted towards the "pitch" of the sample (when I say 'should' I refer to the result from a FFT analysis of wav samples from the same microphone ran through a python FFT module); I know there are background frequencies picked up from these samples so perhaps it is simply being oversensitive but the decibel readings are surely too great at extremely high frequencies for that to be the sole cause.
Any ideas how I can improve the accuracy of these FFT results? The AI needs clear data with obvious dominant frequencies.
Here is the sort of decibel distro I would like for a sample of someone talking: 
And here is what I get: 