Transcriptions

Last updated November 13, 2026

POST /v1/audio/transcriptions serves OpenAI's speech-to-text API, with your own OpenAI or Groq key. Point any OpenAI SDK at Tempr and it transcribes on the same virtual key, limits and logs as your chat traffic.

Request

POST https://api.temprhq.io/v1/audio/transcriptions
Authorization: Bearer tvk_...
Content-Type: multipart/form-data

Form fields

Example

curl https://api.temprhq.io/v1/audio/transcriptions \
  -H "Authorization: Bearer $TEMPR_KEY" \
  -F file=@meeting.wav \
  -F model=openai/gpt-transcribe \
  -F language=en

With the OpenAI Python SDK:

from openai import OpenAI

client = OpenAI(base_url="https://api.temprhq.io/v1", api_key="tvk_...")

with open("meeting.wav", "rb") as audio:
    result = client.audio.transcriptions.create(
        model="groq/whisper-large-v3-turbo",
        file=audio,
    )
print(result.text)

Response

With response_format json:

{ "text": "Let's move the release to Thursday." }

With text, the transcript alone as text/plain. Every response carries x-tempr-request-id.

Models & prices

The model picks the provider, and the request runs on your own key for it (the organization's key, for an organization's virtual key). Add one in the Portal's provider keys; see Providers & BYOK. A custom endpoint you set for the provider, such as a regional host, is used for transcriptions too.

OpenAI shuts down whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe on February 26, 2027. From that day, a recording sent to one of them goes to openai/gpt-transcribe instead, at its price.

The provider's own model name works too, as the OpenAI SDKs send it: gpt-transcribe is OpenAI's, and a bare whisper-large-v3 or whisper-large-v3-turbo is Groq's, the only provider serving them. Any other model gets 400 model_not_found, with the list of the models above.

Prices are the providers' list prices per minute of audio. Tempr works out the length from the audio itself, not from what the provider reports, and your provider bills you; Tempr charges nothing per transcription.

Audio

The Gateway takes one format, so it can measure each recording's length, and with it the cost, from the audio itself:

• WAV with 16-bit PCM samples, one channel (mono), recorded at 8 to 48 kHz. Anything else gets 415 unsupported_audio.
• Up to about two and a half minutes. A longer recording gets 413 audio_too_long.
• Up to 10 MB. A larger file gets 413 audio_too_large.

To convert another file, ffmpeg does it in one line, for example ffmpeg -i input.m4a -ac 1 -ar 16000 -c:a pcm_s16le output.wav. For longer audio, split it and send the parts.

What applies

Transcriptions go through the same pipeline as chat, with these differences:

Errors

In OpenAI's error shape, like every Gateway error; see Errors for the ones every endpoint shares, such as authentication and limits.