Audio in
Send a recording with your question; get an answer about it.
To turn a recording into text and nothing else, Transcription is the guide.
1. Pick a model that listens
Only some text models take sound. Models that take audio in lists them; the sample below uses google/gemini-2.5-flash.
A recording sent to a model that cannot listen is refused before it goes anywhere, with a 400. Nothing is charged.
2. Add the recording as a part of your message
A message's content is a text, or a list of parts. To send a recording, make it a list: one text part for your question, one input_audio part for the recording.
| Part | Holds |
|---|---|
{"type": "text", "text": "..."} | your question |
{"type": "input_audio", "input_audio": {"data": "...", "format": "wav"}} | the recording: its bytes as base64 in data, and its format (wav or mp3) in format |
The recording goes in the request itself: there is no link to give. Read the file, encode it as base64, and put the result in data.
import base64
import json
import os
import urllib.request
API_KEY = os.environ["OPENDUNES_API_KEY"]
ADDRESS = "https://opendunes.com/v1/chat/completions"
MODEL = "google/gemini-2.5-flash"
PROMPT = "List the decisions taken in this meeting."
RECORDING_FILE = "meeting.wav"
with open(RECORDING_FILE, "rb") as file:
recording = base64.b64encode(file.read()).decode()
body = {
"model": MODEL,
"messages": [
{
"role": "user",
"content": [
{"type": "text", "text": PROMPT},
{
"type": "input_audio",
"input_audio": {"data": recording, "format": "wav"},
},
],
}
],
}
request = urllib.request.Request(
ADDRESS,
data=json.dumps(body).encode(),
headers={
"Authorization": "Bearer " + API_KEY,
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(request) as response:
reply = json.load(response)
print(reply["choices"][0]["message"]["content"])
print(reply["usage"]["cost"])Send wav or mp3, and keep the recording short enough that the whole request stays under 32 MB.
Don't put a link in data. It holds the recording's bytes, as base64, never an address.
3. Read the answer
The answer is the text in choices[0].message.content, as on any text request. The recording is charged as audio input, counted in usage.prompt_tokens_details.audio_tokens, and usage.cost is the whole charge.
To send the answer back as a stream, see Streaming. Every field of the request: Chat.
A batch does not take audio. Send the request on its own: see Batches.
Errors
The body says why. What to do on every status: Errors and retries.
| Status | Means |
|---|---|
| 400 | the model takes no audio, or the recording is in a format it does not read |
| 402 | the balanceBalanceWhat you can spend, in DA.More on Balance and charges → cannot cover the request |
| 404 | no such model |
| 413 | the request is larger than the door takes |