Audio Transcriptions

The AI Gateway proxies audio transcription requests the same way it proxies chat: point an OpenAI-compatible client at a gateway route and call its /audio/transcriptions endpoint. The request is the standard OpenAI multipart/form-data upload (the audio file rides along with the form fields), so client.audio.transcriptions.create works unchanged. Usage is recorded for every transcription call, and estimated cost is tracked when pricing data is available for the model.

Transcriptions currently require a bring-your-own-key (BYOK) provider — the upstream provider bills your own credential directly, and requests routed to a Logfire built-in provider are refused for now (see Providers).

Which providers can serve transcriptions

Transcriptions are available for provider types whose /audio/transcriptions endpoint the gateway proxies as an OpenAI-compatible request: OpenAI, Azure Foundry, Groq (whisper models), Mistral (voxtral models), Ollama, and custom OpenAI-compatible providers.

Models and response formats

Two kinds of transcription models are metered differently:

Token-priced models (gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-transcribe): the JSON response carries a usage object with audio and text token counts, and the gateway prices them from its model catalog.
Duration-priced models (whisper-1, Mistral’s voxtral-mini-latest, Groq’s whisper-large-v3 and whisper-large-v3-turbo): the response reports the audio duration, which the gateway records as usage. Groq’s responses carry no usage object at all, so request response_format: verbose_json there — the gateway meters the duration that format reports.

The response_format field passes through to the provider, so the model must support the format you request: whisper-1 accepts all the OpenAI formats (json, verbose_json, text, srt, vtt), while the token-priced models above accept only json (gpt-4o-transcribe-diarize also documents text) — check your provider’s own documentation for other models. If the provider has Require cost data enabled, formats that produce no usage (text/srt/vtt) are refused up front.

Streaming transcription and audio/translations are not supported through the gateway yet.

Sending a transcription request

Address the request to <gateway-base-url>/<route>/audio/transcriptions as multipart/form-data with a file and a model field (plus the optional OpenAI fields: language, prompt, temperature, response_format). The gateway caps the request body at 30MB, slightly above the 25MB audio file limit OpenAI and Groq document, so a maximal upload plus its form fields still fits.

The examples below use an openai route in the US region pointing at your own OpenAI provider credential; see the gateway base URLs for other regions.

Terminal

curl https://gateway-us.pydantic.dev/proxy/openai/audio/transcriptions \
  -H "Authorization: Bearer <YOUR_GATEWAY_API_KEY>" \
  -F file=@recording.wav \
  -F model=gpt-4o-mini-transcribe

from openai import OpenAI

client = OpenAI(
    api_key='<YOUR_GATEWAY_API_KEY>',
    base_url='https://gateway-us.pydantic.dev/proxy/openai',
)

with open('recording.wav', 'rb') as audio_file:
    transcription = client.audio.transcriptions.create(
        file=audio_file,
        model='gpt-4o-mini-transcribe',
    )
print(transcription.text)

import fs from 'node:fs'

import OpenAI from 'openai'

const client = new OpenAI({
  apiKey: '<YOUR_GATEWAY_API_KEY>',
  baseURL: 'https://gateway-us.pydantic.dev/proxy/openai',
})

const transcription = await client.audio.transcriptions.create({
  file: fs.createReadStream('recording.wav'),
  model: 'gpt-4o-mini-transcribe',
})
console.log(transcription.text)

Usage is recorded on every transcription request like any other gateway call (and estimated cost too, when pricing data is available for the model), so transcriptions show up in your Spending analytics and (when telemetry is enabled) as traces alongside the rest of your LLM traffic, with the transcript as the response text. The raw audio never enters your telemetry — traces carry a summary of the form fields (model, filename, size), not the file bytes — but the transcript text itself is recorded as the response message, just like a chat completion’s output. Keep that in mind if your recordings contain sensitive content.

The prompt form field (context you give the model about the audio) goes through the gateway’s input guardrails like any chat message; the audio content itself is not inspected.

Audio Transcriptions

Which providers can serve transcriptions

Models and response formats

Sending a transcription request

See also