What AI Model Powers Instagram Transcript Tools? We Name Ours
Ours runs on Cloudflare Workers AI, using the @cf/openai/whisper model. Audio goes to Cloudflare's edge for speech recognition, text comes back, and the file is deleted. Most competing tools describe their engine only as "advanced AI" or "advanced speech recognition technology" without naming it, so you cannot tell what is actually processing your audio.
Why we name the model
"Powered by AI" tells you nothing. Whisper is a specific, publicly documented speech recognition model with known behaviour: it handles multiple languages, it can return word-level timestamps, and its accuracy depends heavily on input audio quality. Naming it means you can look up what it does instead of taking a marketing claim on faith.
- You can check how the model behaves on your kind of audio before trusting it with real work.
- You can compare our output against any other Whisper-based tool on equal footing.
- If the model changes, we can tell you — because we know what we are running.
What happens to your audio, step by step
- Upload: the file leaves your browser and reaches our server. We do not pretend otherwise.
- Extract: if you upload video, audio is pulled out and converted to 16 kHz mono PCM WAV, which is the format Whisper expects.
- Transcribe: the audio is sent to Cloudflare Workers AI and the
@cf/openai/whispermodel returns the text. - Delete: the working file is removed after transcription.
How we verify it locally
Before shipping, we ran the same pipeline on this machine with a local Whisper build (faster-whisper, CTranslate2 backend, base model, CPU) to confirm end-to-end behaviour: upload, audio extraction, transcription, text output. The local run is our verification step. The production path uses Cloudflare's hosted model, because that is what can serve you without tying our capacity to one machine.
What this does not tell you
Knowing the model does not let us quote an accuracy percentage, and we do not publish one. Whisper's output quality varies with accent, background noise, music, and recording quality. A single number would have to be measured under stated conditions, and we have not published such a measurement.
Got a Reel you want as text? Upload the file and download the transcript as TXT, SRT, or VTT →
The model only ever sees audio. If you hand it a video file, Wondershare UniConverter converts it to MP3 or M4A first, which removes one moving part before the audio reaches the model.
FAQ
Is Whisper the same model other tools use?
Whisper is widely used, so many tools run some version of it. What differs is which size, which hosting, and whether they tell you. We tell you: Cloudflare Workers AI, @cf/openai/whisper.
Does naming the model matter for my results?
It matters for expectations. Whisper is strong on clear speech and weaker on noisy audio, overlapping speakers, and heavy music. Knowing the engine helps you judge whether a bad transcript is the tool or the source audio.
Do you use a local model in production?
No. The local Whisper run was a verification step on our own machine. Production runs on Cloudflare Workers AI.