Transcribe with your own OpenAI-compatible key

Rooster transcribes on your Mac by default. If you would rather use a cloud service - OpenAI's, or another that speaks the same API - you can, with your own key. Nothing about it goes through Rooster.

Set it up

1. Open Settings and choose Transcription.
2. Set Transcription provider to Cloud API.
3. Under OpenAI-compatible transcription:
   • Base URL is OpenAI's address to begin with; change it for another provider.
   • Model is the speech-to-text model to ask for - whisper-1 to begin with.
   • Paste your key into API key.
4. Press Save key in Keychain. Rooster says the key was saved.

Your key is kept in macOS Keychain, not in Rooster's files. To remove it, clear the field and press Save key in Keychain again.

What is sent

When you transcribe, Rooster compresses the recording's sound and sends that, with your key and the chosen language, to the address you set. The picture, your project and anything else stay on your Mac. Your provider bills you for what you use.

OpenAI's own service accepts up to 25 MB of compressed sound per recording; Rooster checks before uploading and says so if a recording is too long. Other services set their own limits.

What you give up

Cloud transcripts come back a sentence at a time, without the time of each word. So, for those projects:

• Deleting words in the script cuts nothing - not even a whole sentence - because a cut needs to know when each word was said. You can still cut on the timeline.
• Captions are timed per sentence, and word-by-word highlighting has nothing to light up.
• Removing filler words and shortening pauses need word timings, and say they cannot work on the transcript.
• A typed script cannot be aligned: aligning needs the time of each word.
• The transcription glossary changes nothing.

To get all of these back, switch the provider to Local Parakeet and transcribe again. See Transcribe a video on your Mac.

If it fails

Rooster says why - a rejected key, no credit left, too many requests, a timeout, or a reply that was not a transcript (often a model that is not a speech-to-text one). See Transcription failed or found no speech.