Home / Reel Transcript Model Loading

Why Does an Instagram Reel Transcript Load a Model First?

XAPN recognizes Reel speech in your browser, so it must initialize its English speech model before showing text. A loading message after opening the page does not prove the model files were downloaded again; the browser may reuse cached files. The site fetches Reel audio first, then decodes it, initializes the recognizer if needed, and transcribes. Wait for Done. and visible words before using the result.

What we observed and inspected

On October 11, 2026 (Los Angeles time), we reloaded XAPN's live homepage in the Codex in-app browser. The result returned to No transcript yet. and the status to Ready. We used the public example Reel already in the input and selected Get transcript. The immediate status was Fetching the Reel audio locally.... The run later reached Done. with nonempty text. We did not republish the creator's words. This confirms one completed run after a reload, not the duration of each stage.

We also inspected the JavaScript served by that homepage. Its click handler clears the old result, posts the link to /audio, decodes the returned audio, and converts it to one 16 kHz channel. A page-scoped recognizer variable starts empty. Only when that variable is empty does the page call pipeline('automatic-speech-recognition', 'Xenova/whisper-tiny.en'). The model then receives the audio samples; recognized text is placed in the result area. Another run in the same loaded page can reuse that initialized variable. Reloading creates a new page and a new variable.

Before this check, the stage message said “Downloading free Whisper tiny model (first run only).” The code did not measure transferred bytes or inspect cache use before showing it. We changed the message to “Loading the Whisper tiny model in this browser” so it describes initialization without asserting a fresh download. That change affects the status wording, not the model or recognition algorithm. Browser caching can affect later runs, but this particular inspection does not establish whether the model was fetched from network during our run.

Evidence boundary: we did not clear the browser cache, record model transfer sizes, time a cold start, or compare devices. We observed the initial audio-fetch status and the final result; intermediate messages may have passed too quickly to capture. No download-size or speed claim follows from this test.

While the tool is running

1. Keep the tab open after selecting Get transcript. Audio retrieval and model work happen in separate stages.
2. If it stays at audio fetching, check the Reel link and availability first. A model-cache change will not fix inaccessible audio.
3. After Done., confirm words appear and review them against the Reel before quoting. Save the result before refreshing the page.

The failure-stage guide separates audio and recognition problems. The refresh guide explains why on-page text should be saved. Neither guide turns the model-loading status into proof of a new download.

Questions

Does “loading” mean the model downloads every time? No such conclusion follows from the status. The page initializes the model, but does not report whether files came from cache or network.

Does Done. guarantee accuracy? No. It reports completion; review the text against the audio.

Model loading is only one stage of the conversion. To convert a Reel to text, the page must also retrieve and decode that Reel's audio. Keep the tab open until the recognition status finishes.