A short documentary podcast has clear narration, environmental sound from a recording studio, and a licensed musical bed. The host’s voice should remain intelligible on a phone while the room never becomes unnaturally silent. DaVinci Resolve Studio has AI Audio Assistant for preparing an automated Fairlight mix, but its output is a starting point, not an editorial signoff. The workflow here starts with three separate tracks, creates a reference mix, and requires a full human listen for music ducking, speech artifacts, and changes in the noise floor. Our mixer drawings are conceptual teaching illustrations rather than screenshots or measured plugin results.
Separate the three kinds of sound before automation
Set up dialogue, music and ambience on different labeled tracks. A microphone clip that contains both the host and background music is not an isolated dialogue stem, so automatic classification may not provide meaningful separation. Label the original files, preserve an unprocessed project duplicate, and mark the moments when the host stops speaking. Aim for a concrete deliverable: a twenty-second opening with dialogue, a breath-length pause, and a music transition. That small passage exposes whether ducking sounds natural. Keep licenses for the musical cue and the original field audio near the project; AI mixing changes neither rights nor the truthfulness of a recording.
Start the Studio-only Audio Assistant in the documented order
Blackmagic Design’s DaVinci Resolve 20 New Features Guide identifies AI Audio Assistant as a Studio-only feature. In the timeline, separate different audio elements by track, then use Timeline > AI Tools > Audio Assistant, select the intended delivery standard, and start Auto Mix. Blackmagic describes classification and track organization, dialogue leveling, optional voice isolation and de-essing, music ducking, sound-effect handling and mastering. Those are capabilities of the tool, not evidence that a particular podcast has passed an audio test. Record your settings and the chosen delivery format before comparing exports.
Compare the same passage with assistance on and off
Duplicate the timeline or save two labeled mixes: untouched baseline and assistant-treated preview. Match playback loudness sufficiently that a louder version does not automatically seem more professional. Listen to three intervals: host speech over music, host speech without music, and ambience-only pause. Ask whether consonants stay crisp or whether the voice acquires a metallic edge from aggressive isolation. Notice if the music pumps upward between words; that can be more distracting than a constant quieter bed. A simple three-column listening log makes failures visible without pretending to report objective results from a test we did not run.
Check music ducking as a storytelling decision
Ducking should support speech rather than behave like a rigid volume switch. Place a marker at the start and end of the host’s longest sentence, at a breath pause, and at the first moment when music should lead the scene. Inspect the assistant-created automation around those markers. If the swell after a sentence is too sudden, change fade timing manually rather than allowing repeated surges. A podcast containing an emotional pause may require music to remain restrained even though there is no speech. Automatic speech detection cannot infer that editorial intention. Keep a note explaining each intentional override for the final reviewer.
Inspect voice cleanup and room-tone continuity
Compare the host’s breath, sibilants and low-frequency background against the unprocessed microphone file. Voice Isolation and De-Essing, when applied, can help with artifacts but may also exaggerate differences between adjacent clips. Listen to a few seconds of pure studio ambience, because an abrupt collapse into silence sounds like a damaged edit. If cleaning removes a word ending, reduce treatment or revert that segment. Never replace indistinct speech by guessing its wording. In interviews, retain the source recording as the authority for quotes. The conceptual waveforms in this guide illustrate what to inspect; they do not claim a measured reduction in decibels.
Choose a delivery standard without inventing loudness results
The assistant asks for a delivery standard, and that choice should come from the publisher’s actual audio specification. Do not promise that a file meets a specific LUFS or true-peak number merely because the preset was selected. Use meters and a compliant loudness measurement in the final export process, and save the report when applicable. Listen on both neutral headphones and an ordinary phone speaker, where low music energy and vocal intelligibility can behave differently. If a platform has its own normalization, treat the platform preview as a separate check. The release decision should document the target, measurement method and any correction made afterward.
Build a review sheet another editor can follow
Create a three-row sheet for dialogue, music and room ambience. For each row record original file, assistant action, one audible concern and decision. The host row may identify a breath that became sharp; the music row may flag a swell over a proper noun; the ambience row may reveal an unwanted gap. Add timecodes and avoid subjective labels such as “perfect” without evidence. If the mix contains sponsor text or safety information, verify that the complete line remains unmasked. Make the revision reversible by keeping source stems and an export of the baseline. This creates an auditable handoff rather than one unexplained AI button press.
Approve the complete episode, not just the demo clip
Listen from start to end with the original script or transcript. Check that transitions between guests or locations do not make room tone jump. Spot-check the loudest sound effect, the quietest interview phrase and the first seconds after the outro begins. Export a brief version history describing the Audio Assistant preset and manual edits. If a listener reports missing syllables, return to the untouched audio instead of reprocessing blindly. The deliverable is an approved mix that keeps words intelligible and preserves intentional ambience, with known limitations recorded. Automation can save time, but editorial responsibility stays with the person publishing the podcast.
How We Research
This DaVinci Resolve AI Audio Assistant workflow is an editorial demonstration built from the following primary sources: Blackmagic Design DaVinci Resolve 20 New Features Guide; Blackmagic Design DaVinci Resolve. The described scenario is Three-track documentary podcast: host narration, quiet room ambience, and licensed instrumental music. Our visual diagrams are independently drawn conceptual teaching aids, not product screenshots, vendor outputs, controlled benchmarks or evidence of hands-on software testing. We cross-checked described menu behavior and important limitations against the linked guidance, then defined a human review plan for factual accuracy, consent, accessibility and release quality. We did not upload private production media, purchase a plan or generate a claimed performance score. Readers should verify their software version and account eligibility and compare their own outputs against untouched originals before releasing anything.
Frequently Asked Questions
Blackmagic lists AI Audio Assistant as a Studio-only feature in its Resolve 20 guide.
Yes. The official workflow advises keeping differing audio elements off the same track before automatic mixing.
Blackmagic says music ducking and automation may be added; an editor should still listen and adjust.
No. Select the correct delivery target and measure the exported file against the publisher specification.
No. Keep original stems and a baseline mix for reversibility and accurate quotation.