Whisper Large v3 API

Transcribe or translate speech audio into text across 99 languages, with segment/word timestamps and optional speaker diarization.

0.01

OpenAPI

Input

Audio URL

URL of the audio file to transcribe. Supported formats: mp3, mp4, mpeg, mpga, m4a, wav or webm.

Additional Settings

Customize your input with more control.

Task

Whether to transcribe the audio in its spoken language or translate it into English text.

Language

ISO code of the spoken language. Leave unset to auto-detect.

Diarize

Annotate which speaker said each chunk. Adds processing time (and therefore cost).

Chunk Level

Timestamp granularity: none, segment (default), or word.

Batch Size

Min: 1 - Max: 64

Internal batch size for inference.

Prompt

Optional text hint to bias transcription toward specific terms or spelling.

Num Speakers

Expected number of speakers. Only used when diarize is true; leave unset to auto-detect.

You need to be logged in to run this model and view results.

Output

Breaking news this evening, scientists at the Coastal Research Station have recorded the highest ocean temperatures in more than 40 years, raising fresh concerns about the health of nearby coral reefs.

{
  "error": "",
  "inferenceTime": 1675,
  "output": " Breaking news this evening, scientists at the Coastal Research Station have recorded the highest ocean temperatures in more than 40 years, raising fresh concerns about the health of nearby coral reefs.",
  "input": {
    "task": "transcribe",
    "prompt": "",
    "diarize": false,
    "language": null,
    "audio_url": "https://media.modelrunner.ai/SNOFBbxV7AMdxGHPq0Th9.mp3",
    "batch_size": 64,
    "chunk_level": "segment",
    "num_speakers": null
  },
  "logs": "Generated 1 output(s)"
}

Generated in 1.675 seconds

Logs (1 lines)

Examples

Good afternoon and welcome aboard flight 217 to Tokyo. Our estimated flight time today is 11 hours and 20 minutes. Please fasten your seatbelts and stow your tray tables for departure.

In today's episode, we explore how small daily habits repeated with patience and consistency compound into remarkable results over the course of a single year.

Whisper Large v3 API

Whisper Large v3 is a speech-to-text AI model by openai. On ModelRunner it runs through a REST API or via MCP from any AI assistant, at $0.01 per response.

POST https://queue.modelrunner.run/openai/whisper

cURL

# Submit a request to the queue. Input fields go at the top level of the
# body. The optional reserved "metadata" object holds your own flat string
# tags — stored on the request, never sent to the model; filter later with
# GET https://queue.modelrunner.run/requests?metadata=<url-encoded JSON>.
curl -X POST https://queue.modelrunner.run/openai/whisper \
  -H "Authorization: Key $MRUN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "task": "transcribe",
    "prompt": "",
    "diarize": false,
    "language": null,
    "audio_url": "https://media.modelrunner.ai/SNOFBbxV7AMdxGHPq0Th9.mp3",
    "batch_size": 64,
    "chunk_level": "segment",
    "num_speakers": null,
    "metadata": {
      "project": "my-project"
    }
  }'
# → { "request_id": "...", "status_url": "...", "response_url": "..." }

# Poll status_url until "COMPLETED", then fetch the result
curl "https://queue.modelrunner.run/openai/whisper/requests/$REQUEST_ID/status" \
  -H "Authorization: Key $MRUN_API_KEY"
curl "https://queue.modelrunner.run/openai/whisper/requests/$REQUEST_ID" \
  -H "Authorization: Key $MRUN_API_KEY"

JavaScript

import { modelrunner } from "@modelrunner/client";

const result = await modelrunner.subscribe("openai/whisper", {
  input: {
    "task": "transcribe",
    "prompt": "",
    "diarize": false,
    "language": null,
    "audio_url": "https://media.modelrunner.ai/SNOFBbxV7AMdxGHPq0Th9.mp3",
    "batch_size": 64,
    "chunk_level": "segment",
    "num_speakers": null
  },
});
console.log(result);

Python

import os
import requests

headers = {"Authorization": f"Key {os.environ['MRUN_API_KEY']}"}

submitted = requests.post(
    "https://queue.modelrunner.run/openai/whisper",
    headers=headers,
    json={
      "task": "transcribe",
      "prompt": "",
      "diarize": false,
      "language": null,
      "audio_url": "https://media.modelrunner.ai/SNOFBbxV7AMdxGHPq0Th9.mp3",
      "batch_size": 64,
      "chunk_level": "segment",
      "num_speakers": null
    },
).json()

# Poll submitted["status_url"] until "COMPLETED", then:
result = requests.get(submitted["response_url"], headers=headers).json()

Input parameters

Name	Type	Required	Description
audio_url	string (uri)	yes	URL of the audio file to transcribe. Supported formats: mp3, mp4, mpeg, mpga, m4a, wav or webm.
task	enum	no	Whether to transcribe the audio in its spoken language or translate it into English text. Default: "transcribe".
language	enum	no	ISO code of the spoken language. Leave unset to auto-detect. Default: null.
diarize	boolean	no	Annotate which speaker said each chunk. Adds processing time (and therefore cost). Default: false.
chunk_level	enum	no	Timestamp granularity: none, segment (default), or word. Default: "segment".
batch_size	integer	no	Internal batch size for inference. Default: 64.
prompt	string	no	Optional text hint to bias transcription toward specific terms or spelling. Default: "".
num_speakers	integer	no	Expected number of speakers. Only used when diarize is true; leave unset to auto-detect. Default: null.

Machine-readable: OpenAPI schema · llms.txt

Use Whisper Large v3 from Claude & Cursor (MCP)

Point Claude Code, Claude Desktop, Cursor, or any MCP client at the ModelRunner MCP server and Whisper Large v3 becomes a tool your assistant can call directly — it authorizes via OAuth (no API key in config) and runs this model with the run_model tool using the endpoint openai/whisper.

MCP client config (Claude Desktop, Cursor)

{
  "mcpServers": {
    "modelrunner": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.modelrunner.run/mcp"]
    }
  }
}

Claude Code

claude mcp add --transport http modelrunner https://mcp.modelrunner.run/mcp

Then ask your assistant, for example: “Run openai/whisper on ModelRunner to generate text”. MCP setup guide.

Model Details

Whisper Large v3 is OpenAI's open-source, multilingual speech recognition model. Pass a URL to an audio file (mp3, mp4, mpeg, mpga, m4a, wav, or webm) and it returns the full transcript as plain text. It handles 99 languages with automatic language detection, can translate any spoken language into English text, and returns segment- or word-level timestamps plus optional speaker diarization. It is a robust, general-purpose default for turning recordings into searchable, usable text.

## Best for - Transcribing meetings, interviews, podcasts, lectures, and voice notes into text - Multilingual transcription where the spoken language is unknown or mixed (99 languages, auto-detected) - Translating non-English audio directly into English text in one call (`task: "translate"`) - Generating timestamped segments or word-level chunks for captioning and subtitling - Speaker-attributed transcripts of multi-person recordings via diarization

## Choose another model when - You want to generate speech from text rather than transcribe it — use a text-to-speech model - You need live, streaming transcription of an in-progress call — this processes a complete uploaded file and returns a finished transcript - You want maximum-accuracy English transcription with built-in audio-event tagging (laughter, applause) — consider `elevenlabs/scribe-v1`

## Tips - Leave `language` unset to auto-detect the spoken language; set an ISO code (e.g. `en`, `es`, `fr`, `de`, `ja`) only when the language is known, to skip detection. - Set `task` to `translate` to force English output regardless of the source language; keep `transcribe` (default) to output in the spoken language. - Use `chunk_level` to control timestamp granularity: `segment` (default) returns sentence-level chunks, `word` returns per-word timestamps, `none` skips timestamp tokens for a small speed-up. - Enable `diarize` only for multi-speaker audio; it labels who spoke each chunk but adds processing time (and therefore cost, since billing tracks compute time).

## Advanced Configuration - `task` (default `transcribe`): `transcribe` keeps the spoken language; `translate` outputs English. - `chunk_level` (default `segment`): timestamp granularity — `none`, `segment`, or `word`. - `diarize` (boolean, default `false`): annotate which speaker said each chunk. Requires more compute time. - `num_speakers` (default auto): hint the expected speaker count; only used when `diarize` is `true`. - `prompt` (default empty): a text hint to bias transcription toward specific terms or spelling. - `batch_size` (default 64): internal batching; leave at the default unless tuning throughput.

Note on cost: this model bills per transcription request. Pricing is flat per output regardless of audio length.

To run via the ModelRunner JavaScript client: ```js import { modelrunner } from "@modelrunner/client";

const result = await modelrunner.subscribe("openai/whisper", { input: { audio_url: "https://media.modelrunner.ai/iuneUX0YY4AtcsceV9HHp.mp3", task: "transcribe", chunk_level: "segment", }, }); ```

Whisper Large v3 API

Model Input

Input

Additional Settings

Model Output

Output

Model Example Requests

Examples

Whisper Large v3 API

cURL

JavaScript

Python

Input parameters

Use Whisper Large v3 from Claude & Cursor (MCP)

MCP client config (Claude Desktop, Cursor)

Claude Code

Model Details

Model Details