Skip to main content

Whisper-Large-V3-Turbo

‍

The Speech-to-text guide‍ has runnable examples for basic transcription, multi-language usage, large-file chunking, and a transcription + summarisation pipeline.

For speaker separation, word timestamps, subtitles, or recordings longer than 30 minutes, see Cohere-Transcribe-03-2026‍.

Description​

“Whisper-Large-V3-Turbo” is a multilingual automatic speech recognition model (ASR) developed by OpenAI, optimized for speed and efficiency. It is based on the architecture of the well-known “Whisper-Large-V3” model, but uses a lighter decoder structure to significantly reduce latency with only a minimal loss in accuracy. The model supports over 99 languages and is ideal for transcribing speech inputs.

This model has the lifecycle status LTS‍. If we shut it down or replace it, we announce that at least 6 months in advance.

Token usage​

Whisper processes audio in 20 ms steps, which corresponds to 50 base tokens per second of audio. The billed token count is calculated by multiplying this value by a token factor of 0.5x, resulting in 25 billed tokens per second of audio.

The following limitations apply to this model on our platform:

  • Maximum file size: 25 MB per upload
  • Maximum audio duration: 30 minutes (1800 seconds) per request
  • Translation is currently not supported (to_language)
  • Supported output formats: json, verbose_json
    • response_format="text" is accepted but returns a JSON body regardless — use "json" instead
    • srt and vtt are not supported (HTTP 400)

Audio duration and file size​

Supported Input Formats​

mp3, ogg, wav, flac

Supported values for parameter language (ISO-639-1 language codes)​

af, ar, az, be, bg, bs, ca, cs, cy, da, de, el, en, es, et, fa, fi, fr, gl, he, hi, hr, hu, hy, id, is, it, ja, kk, kn, ko, lt, lv, mk, mi, mr, ms, ne, nl, no, pl, pt, ro, ru, sk, sl, sr, sv, sw, ta, th, tl, tr, uk, ur, vi, zh

Recommended Inference Parameters​

  • temperature=0, the default. Higher values make the model sample instead of picking the most likely word, and the transcript gets worse: on ten German and English broadcast and interview recordings, the word error rate rose from 18.4 % at 0 to 30.1 % at 1.0.
  • top_p=1.0
  • response_format="json"
  • language like language="de"should always be set explicitly to maximize accuracy. If no value is provided, German ("de") will be assumed by default, which may result in poorer outcomes for inputs in other languages.

Example output — response_format="json"​

{
  "text": "This is the transcribed text of a speech input.",
  "usage": {
    "type": "duration",
    "seconds": 8
  }
}

Example output — response_format="verbose_json"​

Returns additional metadata including detected language, duration, and per-segment timestamps:

{
  "text": "This is the transcribed text.",
  "language": "en",
  "duration": "8.0",
  "words": null,
  "segments": [
    {
      "id": 0,
      "avg_logprob": -0.45,
      "text": " This is the transcribed text.",
      "start": 0.0,
      "end": 2.4
    }
  ]
}

Terms of Use and Licensing​

The model is provided by OpenAI under the MIT License‍, and reuse of the generated content is subject to no additional restrictions.

On this page