Resolve Free AI noise-cleanup example

Made September 18, 2026. DaVinci Resolve Free 21.1 on macOS.
Synthetic macOS Samantha narration; deterministic generated noise, not field recordings.
The same speech appears twice: 0-17s steady filtered noise and 120Hz hum,
17-34s intermittent Gaussian noise bursts over faint background noise.

Files:
noisy-source.mp4: reference picture and noisy audio.
noisy-input.wav: exact audio submitted to fal-ai/elevenlabs/audio-isolation.
isolated-original.mp3: unmodified returned model audio.
before-matched.wav: noisy input at -20 LUFS, constant gain.
cleaned-aligned.wav: returned audio resampled to 48kHz, level matched,
advanced 1312 samples (27.333ms) and trimmed/padded to exactly 34 seconds.
No speech time-stretch was used. This correction is specific to this recording.
clean-reference.wav: original speech before noise was added; useful for comparison.
measurements.json and export-verification.json: numerical checks, not quality scores.

Voyager-noise-cleanup.drp: editable Resolve project. Relink the media from this
folder after import if needed. Open Noise cleanup - AI audio. A1 is muted, A2
is the aligned cleaned WAV. Solo/mute carefully when comparing; do not play both.
noise-cleanup.fcpxml: original generated timeline import (absolute source paths
need updating on another computer). In our import, its audio-disable flag did
not mute A1; we clicked M on A1 in Resolve and verified the native render.

Scripts reproduce fixture generation and analysis with Python, NumPy, SciPy,
and FFmpeg. create_fixture.py expects narration.wav. create_video.py expects
Arial at the macOS system font location. analyze_audio.py uses the uncorrected
cleaned-48k.wav and generates before/after-matched.wav. Source normalization
and the exact correction are recorded in the JSON reports.

No keys or signed upload URLs are included. This one controlled synthetic test
cannot establish behavior on accents, clipping, wind, reverb or overlapping speakers.
