[Python] How to automatically and reliably find level above signal only in region of low...

I have samples of wav files, for which I can plot sum of librosa.feature.spectral_contrast (red line on picture). There are regions of voice and silence/low noise in wavefile (sound amplitude is blue on picture). I added green line, which should be maximum of this red line in regions of no voice. How can I do it automatically and reliably (I have thousands of files, cannot do it manually). Amplitude ratios of silence/voice regions may vary in different files.

Continue reading...