There are several ways you could approach this, and you will likely need to use a mixture of different techniques to get the best result.
This sounds like an exercise in forensics rather than a quest for clean audio so some of the solutions offered here may not be of any use
If you can't hear the voices because the background noise is too loud or because the background noise is in the same frequency band as the voices, there is a strong probability that you will not be able to recover anything more intelligible than you already have.
Low-pass filters are designed to only let audio below a set frequency through to the listener - these are great for removing high-frequency hissing noises.
Similarly, high-pass filters only let audio above a set frequency through to the listener - these are great for removing low-frequency rumbles.
Notch filters are useful for removing a specific band of frequencies. These can be very useful for lessening things like electrical interference, wind noise, rumbling or other band-specific sounds.
A dynamic EQ will react to both frequency and volume to boost or cut specific frequency ranges. They can be fine-tuned to a very precise degree and, as they're dynamic, they're great for trying to zero in on moving targets like human voices.
A noise gate measures the volume of incoming audio and only lets it through if it's loud enough. This should help to get rid of some of the background noise when the voices are not sounding.
Spectrograms are a way of visualising changes within a whole range of frequencies over time. You get a 'heat-map' style image of the frequencies in your audio - this can be extremely useful in identifying which frequencies should be cut and when.
A note about Frequency Ranges
Human voices sit between around 0.5kHz to 9kHz with intelligibility working in the 2kHz to 4kHz range. The actual ranges involved will be smaller but their bounds depend on who is talking - male or female, adult or child.
You will need to pay special attention to any audio in these bands - cutting too much will reduce intelligibility.
Most, if not all of these tools, are available built into free audio editing software such as Audacity or can be found as free plugins.
I would start by removing anything outside those frequencies - use a low-pass filter to cut everything above 9kHz and a high-pass filter to cut everything below 0.5kHz - This removes everything that is definitely not a voice. Then have a look at a spectrogram of what you have left and figure out if you can use a gate and where you might be able to use notch filters and/or EQ.
Getting something intelligible back is going to take a lot of work. Getting something clean back is likely to be impossible.