Extracting speech from video recordings of a workshop with strong noise and preparing clean tracks with a QC report

Initial situation - there are 14 videos (MP4 H.264, 1080p, 25 fps) with a total duration of about 3 hours. Filmed in a production workshop with the equipment running - constant broadband noise, periodic pulses and whistles, and clipping in places. In the frame, 1-2 people give instructions; speech is needed to be as intelligible as possible for subsequent editing of training videos. The original sound is 1 stereo track from the camera, no external microphones. Sometimes they say not to the camera, the level floats.

Technical environment and subject area - post-production for an industrial training video, the task is specifically about extracting useful sound (speech) from a video with an aggressive noise environment using professional tools (iZotope RX Advanced / analogue, spectral editing, multiband dynamics, de-reverb if necessary). The work must be carried out with an understanding of artifacts (musical noise, lisp, phase distortion, pumping), and with control of speech intelligibility.

The specific result is to provide 1-2 final audio tracks and intermediate versions for comparison for each video:
1) Clean Dialogue - selected and leveled speech, minimal possible noise, ready for installation.
2) Natural Dialogue - a more gentle treatment for cases where aggressive cleaning produces artifacts.
Output format - WAV 48 kHz 24-bit, mono or stereo according to rationale, synchronization with the source video without drift. Additionally - EDL/marker sheet (TXT/CSV) with time codes of problem areas (clipping, unintelligible speech, overlapping voices, sharp hits) and a brief commentary on what was done.

Restrictions - You cannot use generative voice recording, dubbing or TTS. You cannot “cut” phrases for the sake of purity - the meaning of the instruction must be preserved, pauses are allowed only if this does not break perception. Local edits along the timeline are acceptable, but without changing the lip synchronization in the frame. You cannot introduce audible timbre distortions as a “radio effect”.

Measurable acceptance criteria - in areas where speech is audible in the original, after processing it should be consistently intelligible without noticeable digital artifacts when listening to headphones and monitors. Target parameters for finals:
- Loudness - integrally about -16 LUFS for Clean Dialogue and about -18 LUFS for Natural Dialogue, True Peak is not higher than -1.0 dBTP.
- The level of background noise in pauses is a reduction of at least 8 dB relative to the source in most scenes, without pronounced musical noise.
- No audible desync relative to video, drift 0, up to a maximum of 1 frame at the end of each file.
- For each video, attach A/B excerpts of 30-60 seconds (WAV) “before/after” from the most difficult scenes.

Stages of implementation - primary analysis and selection of processing strategy by type of noise, rough cleaning of 1 test clip for 3-5 minutes to coordinate the approach, then batch processing of all clips with manual spectral editing of problem areas, final QC with checking levels, peaks, synchronization and preparation of a marker sheet.

I will provide the sources via the cloud; if necessary, I will provide 2-3 supporting fragments where maximum legibility is important (terms, numbers, commands).