With those splits, F2L is where the time is — 9 seconds is over half your solve, so that's the place to work first, and I'd leave last layer alone for now.
A drill that helped me more than just doing slow solves: do a block of untimed solves where you force yourself to find the next pair before you finish the current one — even if it means turning really slowly, or pausing your hands while your eyes keep moving. If you catch yourself staring at the pair you're solving, that's the pause to work on, not your turning speed.
For cross, try planning the cross and spotting your first pair during inspection, even if you only manage it on easy scrambles at first. That cross-to-first-pair link is usually worth more than shaving a move off the cross itself.
A simple session that ties it together: 10 slow, look-ahead-only solves, then 10 timed solves where you only judge yourself on whether the pauses got shorter, not the time. Once that F2L time starts coming down, the average follows — it's the biggest chunk by a long way.