Dynamic Audio Creative Built for Insertion, Not Reuse

Twelve to twenty finished variants is where a dynamic audio campaign stops being a spreadsheet exercise and starts doing something. Below that you are running static creative with a city name dropped into it. Above it you are splitting a budget into cells too small to read, and the reads get worse, because by take forty the voice talent is guessing at context. What decides the variant count: Not your segment list. Your impression volume. At a 0.4 percent promo-code redemption rate, which is ordinary for a 30-second mid-roll on a considered purchase, separating two variants that differ by a quarter takes roughly 250,000 impressions per arm. A twelve-cell test on a 500,000-impression flight measures nothing, and it will still hand you a winner, which is the dangerous part. We size the variant set against the media plan first and write to that number, rather than writing to the segmentation deck and hoping the flight grows. How the audio gets made: One session, one performer, every variant. Variable lines are recorded in the same room on the same day, with matched room tone bridging each splice, so the insert does not announce itself as an insert. Masters go out as 48 kHz 24-bit WAV with 192 kbps stereo MP3 alongside, normalised to minus 16 LUFS integrated with a minus 1 dBTP ceiling. That level matters for a reason beyond spec compliance: an ad arriving four decibels louder than the host is the fastest way to teach a listener where the skip button is. Trafficking is VAST 4.2 with macro-driven variant selection, built and tested inside Megaphone's targeted-impressions setup and Acast's dynamic creative tooling before a euro is committed, because a variant that fails to resolve serves your default and you will not notice for a fortnight. What we will not do: We do not clone a host's voice. Synthetic replication of a named person's read is refused whether or not the contract technically permits it, because it damages the single asset that makes host-read audio work at all. Synthesis has one use here: factual variable fields such as branch address, opening hours or a date, where the recorded set would otherwise run past a hundred takes, and only when it sits inside a recorded bed and passes a blind listen against the human takes. We do not build weather triggers or sports-score triggers. If a trigger cannot be tied to a measurable difference in response, it is decoration, and it is decoration you pay production hours for. We do not put more than three variable slots in a 30-second read. Four is audible. Who should not buy this: Brands with one offer, one price and one audience. You need a better single 30, and it will beat a dynamic set on the same budget. Teams who cannot get legal to clear variant copy inside two weeks, because approval, not production, is where these projects die. Anyone measuring this on last-click, since the majority of podcast response arrives as direct or branded search inside 72 hours and your attribution model will attribute it elsewhere. What arrives: The variant matrix mapped to the targeting keys your platforms can actually address. Masters and encoded files. Tested VAST tags with a documented fallback. A one-paragraph note per variant stating what it tests and what result would change the buy. A readback call where we go through the first week of delivery and tell you which cells to kill.