22sGems on YouTube

22sGems is Jon Singer's fine-mineral brand, and every specimen appears first as an Instagram Reel: a hand turning a crystal in the light, a licensed music track underneath, and a caption written like a voice-over. This project takes each of those Reels to YouTube as a Short that speaks. The music is gone, and in its place is a narration in a clone of the owner's own voice, reading his own caption. It runs unattended, and the channel now holds more than 400 videos.

The music has to go because Instagram's licence does not travel. The Instagram Graph API withholds the video file for Reels with copyright-flagged audio - on this account that was every Reel measured, 53 of 53 - so the download falls back to the public reel page, and a Reel uploaded to YouTube with its Instagram track would be claimed by Content ID. The audio is therefore stripped at transcode, and the narration is a replacement track, never a mix over the original.

The voice is an ElevenLabs Instant Voice Clone of the owner, made from recordings he supplied and with his consent. A stock voice was the first plan; his own won the comparison, because a dealer's voice is the one thing on these videos a competitor cannot copy. He sent a 56-second phone recording and then three cleaner takes, and the production voice is built from those three, 2 minutes 57 seconds of speech. The first take was left out on purpose: its noise floor sat 12 to 15 dB higher, and a clone learns the room along with the voice. It is a clone and not a recording, and it is labelled as one: every narrated upload carries YouTube's altered or synthetic content disclosure.

There is no language model in the narration path. The owner's captions are already scripts - a hook, two or three beats of prose, then a block of specifications - so the code keeps the hook and the prose, adds a spoken identity line and a call to action that depends on whether the piece is for sale, and trims the result to a character budget derived from the real length of the video. Then it makes the text speakable: dashes become pauses, hashtags and emoji go, abbreviations are spelled out, shouted headlines are title-cased, a leading weight becomes its own sentence, and a small pronunciation dictionary fixes the words the engine gets wrong, each entry chosen by ear from numbered A/B clips.

The rest is audio engineering. The ElevenLabs v3 model ignores the speed setting, so pace is set afterwards: ffmpeg's atempo filter, with a ratio computed for every render from its measured words per minute toward a target of 115 and clamped between 0.8 and 1.35. Loudness is normalised in two passes to -16 LUFS, the voice gets 0.6 seconds of silence before the first word and 1.6 after the last, and the picture is never cut to fit the voice: if the narration runs longer the last frame holds, and if the footage runs longer it plays on in silence. A quality gate rejects a render whose pace misses the target, whose stretch ratio hit its clamp or whose read looks truncated, and a rejected file is renamed so nothing downstream can pick it up.

Publishing goes through the YouTube Data API as plain REST calls: OAuth, a resumable upload, a title built as mineral, locality and weight, a description with a Buy link to the Shopify product and a link to the specimen's gallery page, and subtitles as a real caption track. Afterwards the video id is written back to the Shopify product and to a map in R2 that the gallery reads at runtime. Shorts that went up silent before the voice existed were replaced one by one, most-viewed first, because YouTube offers no way to change the audio of a published video. The whole thing runs in a Cloudflare Container started by a Worker on cron triggers: two full passes a day and an hourly pass that only works the YouTube queue.