Home/Integrations/Transcriptions/OpenAI Whisper via Replicate
Whisper on every recording, without running a GPU.
AddPipe extracts an audio track from each recording. A webhook hands it to Whisper on Replicate, which bills by the second and calls you back with the transcript when it is done.
Update: Update: The AddPipe Platform now has built-in support for audio transcription, which allows you to automatically detect the language of the audio and transcribe it. You can find more information about this feature in the documentation.
No idle GPU
Whisper needs a GPU to be fast, but recordings arrive in bursts. Replicate runs the model on demand and bills per second, so nothing is paid for between recordings.
Strong multilingual accuracy
Whisper handles accented and multilingual speech well, which matters for international panels and support recordings.
Pin the model version
Replicate addresses models by exact version, so a transcript produced today can be reproduced next year. That is worth having when the output is evidence.
What this is good for.
Multilingual panels
Whisper handles accented and mixed-language speech better than most alternatives, without per-language configuration.
Bulk archives
Transcribe a backlog of recordings in one pass, paying only for the compute the batch actually uses.
Reproducible output
A pinned model version means the same audio produces the same transcript later, which matters when it is part of a record.
Details.
Why Replicate rather than running Whisper myself?
Whisper is an open model but it needs a GPU to run at a sensible speed. Replicate hosts it and bills per second of compute, so you get Whisper without keeping a GPU running between recordings.
Which file gets sent?
The extracted audio track, name_audio.mp4, produced by the transcoding engine alongside the video. Enable it in the Transcoding Engine settings.
How does the transcript come back?
Replicate predictions are asynchronous. Pass a webhook URL when you create the prediction and Replicate calls you when it finishes, rather than you polling for it.
Does the audio have to be publicly reachable?
Replicate needs to fetch the file, so give it a URL it can read. A signed URL with a short expiry from your own bucket is the usual approach.
Which Whisper variant should I use?
large-v3 for accuracy, or one of the faster distilled variants when throughput matters more. Replicate hosts several; pin a specific model version so results stay reproducible.
Can I get word-level timestamps?
Yes, depending on the model version you pick. Check the model's schema on Replicate for the exact input parameters it accepts.
Related integrations.
Run Whisper on your recordings.
Webhooks and the audio track output are open during the 14-day trial. No card required.