Ghost Hunting w/ Audacity

I am using the latest edition of Audacity to process wav files downloaded from my Olympus VN4100-PC digital voice recorder. Specifically, I use the voice recorder for ghost hunting supposedly haunted houses. Yes, I DO get responses to the questions I ask in otherwise empty houses, and many are clear enough to require no enhancements, but many are whispered responses only slightly louder than the noise floor. My attempts to “clean up” these recordings by variously amplifying and using the noise removal function are not giving me good results, so I welcome any guidance the Audacity community can offer. Forget, for a minute, that the voices are those of ghosts, and approach the problem from a forensic point of view of trying to selectively isolate and clarify recordings of whispered voices in general. What advice can you give me?

Even if the recordings were of ghosts, further processing, (e.g. amplification, filtering, noise-reduction), may make the noises sound more like words but actually make them even less significant from an evidentiary perspective …

e.g. here is a processed recording of an inanimate thing … [ no actual voices , just audio pareidolia ]

e.g. filters applied to plain white-noise can produce vowel sounds …

Quite the opposite, actually. The more I edit these samples, the less like words many of them sound. That’s the problem I want to solve - being able to isolate and enhance the samples w/o screwing them up.

Quiet voices that are mostly masked by noise are extremely difficult to isolate. You can easily cut out low bass and high treble using the “telephone” setting in the Equalization effect. This can be effective in improving intelligibility if the voice is reasonably clear before you start (but the high frequency cut-off should be pushed up to around 7 to 8 kHz otherwise it is not possible to distinguish an “s” from an “f” other than by context). If that does not work well enough, then “trying harder” to make out voices will either fail, or will increase the “aural pareidolia” effect, which then casts doubts on the entire exercise. As an example, resonant filters can really pull voices out of a bed of noise, but they can also produce very convincing “voices” from random noise, so even though they can work, they also invalidate themselves if you are seeking evidence. Try using a steep high pass filter at about 200 Hz and a steep low pass filter at around 7000 to 8000 Hz. If that does not work well enough, then you need to get a better recording because further processing reduces the credibility of the evidence.