Transcription

Turn an audio file into text. OpenAI-compatible request and response, Whisper-class models underneath.

Endpoint:

POST /v1/audio/transcriptions

Auth: Bearer token or x-api-key. Requires scope inference.audio. Content-Type: multipart/form-data. Max file size: 25 MB.

Example

from openai import OpenAI

client = OpenAI(
    api_key="sk-mel-<YOUR_API_KEY>",
    base_url="https://api.melious.ai/v1",
)

with open("meeting.mp3", "rb") as f:
    result = client.audio.transcriptions.create(
        model="<STT_MODEL_ID>",   # a transcription model from the hub
        file=f,
        language="de",
    )
print(result.text)

curl https://api.melious.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer sk-mel-<YOUR_API_KEY>" \
  -F model="<STT_MODEL_ID>" \
  -F language="de" \
  -F file=@meeting.mp3

Pick a current transcription model ID from melious.ai/hub/models (filter by audio), or call GET /v1/models?include_meta=true and look for _meta.type == "audio".

Request (multipart fields)

Field	Type	Default	Description
`file`	file	—	Audio file. Formats: mp3, mp4, mpeg, mpga, m4a, wav, webm.
`model`	string	`whisper-large-v3-turbo`	Transcription model ID.
`language`	string	auto-detect	ISO-639-1 code (`"de"`, `"fr"`, `"en"`, …).
`response_format`	string	`"json"`	`"json"`, `"text"`, `"srt"`, `"vtt"`, `"verbose_json"`.
`temperature`	number	`0`	Sampling temperature, `[0, 1]`.

Whisper-class models handle 50+ languages and auto-detect by default. If you know the language, set it — detection adds a few hundred milliseconds and occasionally picks wrong for short clips.

Response

response_format: "json" (default):

{
  "text": "Hamburg, Lübeck, Bremen.",
  "task": "transcribe",
  "language": "de",
  "duration": 2.1,
  "seconds": 2.1,
  "segments": null,
  "words": null
}

duration and seconds carry the same value; read seconds going forward.

verbose_json adds a segments array with per-segment timestamps. text returns a plain-text body. srt and vtt are accepted but currently return the same JSON object as json — build subtitles from verbose_json segments until we fix this.

What about translations?

OpenAI's /v1/audio/translations endpoint (transcribe and translate to English in one step) isn't implemented. Workaround: transcribe in the source language, then pipe the text through Chat completions with a translation prompt.

What about text-to-speech?

POST /v1/audio/speech (OpenAI text-to-speech shape, scope inference.audio) exists but has no documented model yet.

Audio as chat input

A different path from transcription: some chat models accept audio directly as message content, then reason about it rather than just transcribing. That's not this endpoint — it's Chat completions with audio content blocks. Look for "audio" in _meta.input_modalities on GET /v1/models?include_meta=true to find those models.

Errors

Status	`type`	`code`	When
`413`	—	—	The request body is over 25 MB. The gateway rejects it before the route runs.
`404`	`invalid_request_error`	`model_not_found`	Unknown model.
`403`	`authentication_error`	`auth_insufficient_scope`	The key lacks the `inference.audio` scope.

Models for transcription-model discovery • Routing for cost savings on bulk transcription.

On this page