Speech to text API: Whisper transcription, processed in the EU

POST /v1/audio/transcriptions takes an audio file as multipart/form-data and returns what was said. Send the file in the file field and the model id in model, for example whisper-large-v3-turbo. Common formats such as mp3, mp4, m4a, wav, webm, ogg and flac are accepted.

Whisper models recognise speech in close to a hundred languages, including every major EU language, and detect the language automatically if you do not specify one. The default response is {"text": "..."}, which is often all you need to feed the transcript into a summary, a search index or a ticket.

language takes an ISO 639-1 code such as de, fr or nl. Setting it improves accuracy and speed on short or noisy recordings, where automatic detection can guess wrong. prompt gives the model context: names, product terms or the previous segment's text, so that it spells "Kubernetes" or a colleague's surname correctly.

response_format can be json, text, srt, vtt or verbose_json. Use srt or vtt for subtitles. verbose_json returns the detected language, duration and a segments array with start and end times, and with timestamp_granularities[] set to word you get word-level timings as well. temperature is best left at 0 for transcription.

A common workflow records a support call or meeting, transcribes it with verbose_json to keep timestamps, then sends the transcript to a chat model with a prompt that asks for a summary, action items and any commitments made. Store the segments with their times so that people can jump to the moment in the recording.

For large archives of recordings, upload them and process them in the background instead of holding a connection open for each one. Keep the language setting per customer or per team, and keep a short glossary of names and terms to pass in prompt for every file.

Files above the maximum upload size return HTTP 413 or a 400 error. Split long recordings into parts, ideally at silences, and keep the last sentence of one part as the prompt for the next so that context carries over. The size limit depends on your plan, and we confirm it with you.

An unsupported or corrupted file returns a 400 error that names the problem. Long stretches of silence or music can make Whisper models repeat a phrase or invent text, so trim silence first or use voice activity detection. Speaker labels are not part of this response; add a diarisation step if you need to know who said what.