Run Whisper Large V3 Turbo in the EU through an OpenAI-compatible API

Whisper is OpenAI's family of open models for automatic speech recognition and speech translation. Large V3 Turbo, released in October 2024, is a pruned and fine-tuned version of Whisper Large V3: the same model, except that the number of decoder layers has been cut from 32 to 4. That makes it much faster, with about 809 million parameters in total.

It transcribes speech in the language spoken and can translate speech into English text. The model card lists 99 languages. Whisper processes audio in 30-second windows; longer recordings are transcribed by sliding over the audio in chunks, which the serving layer handles for you. The weights are released under the MIT licence.

Whisper is the standard choice for turning recordings into text: meetings, interviews, voicemails, podcasts and support calls. It copes well with accents and background noise in the major European languages, adds punctuation, and can return timestamps for subtitles. The Turbo version keeps much of Large V3's accuracy while running far faster.

It only transcribes or translates; it does not summarise or answer questions about the audio. For that, use Voxtral Small or pass the transcript to a language model. Translation goes into English only. Accuracy is lower for languages with little training data, and like all speech models it can occasionally invent words in long silences or very noisy passages.

inference.eu supports the OpenAI transcription request format, so code that already uses OpenAI's speech-to-text API keeps working. Change the base URL to the inference.eu API, use your inference.eu key and set the model to whisper-large-v3-turbo.

Upload the audio file with an optional language hint, which improves accuracy and speed when you know the spoken language in advance. Ask for timestamps if you need subtitles or want to link text to positions in the recording. For translation into English, use the translation request instead of transcription. The OpenAI SDKs for Python and JavaScript support both out of the box.

Whisper was trained by OpenAI, but on inference.eu the open weights run on infrastructure in the European Union. Your audio files and transcripts are processed in EU data centres, are not sent to OpenAI, and are never used to train any model.

Every customer gets a data processing agreement under Article 28 of the GDPR and our subprocessor list, which matters for recordings that contain names, health details or customer information. Billing is in euros. The MIT licence allows commercial use without restriction. You remain responsible for telling people they are being recorded and for having a legal basis to do so.