How do I transcribe audio with the OpenAI Whisper API using in-memory bytes?

Below is an example of how you can pass your in-memory file directly to the API without first saving it to disk. The key is to pass both the file-like object and its filename (so that the API knows the file type), like so:

PYTHONCopied
import openai
from fastapi import UploadFile

async def transcribe_audio(upload_file: UploadFile):
    # Note: upload_file.file is a SpooledTemporaryFile instance
    response = openai.Audio.transcriptions.create(
        file=upload_file.file,             # File-like object
        filename=upload_file.filename,     # Required to indicate file type
        model="whisper-1"
    )
    return response["text"]

Explanation

• File-like object: FastAPI’s UploadFile provides a file attribute that is a file-like object (a SpooledTemporaryFile) which you can pass directly.
• Filename: Including the original filename is important because it helps the API understand the file type (e.g., .mp3, .wav).
• OpenAI Python library: This library takes care of packaging the file into the required multipart form data format, so there’s no need to save the file yourself.

This way, you can transcribe audio directly from uploads without dealing with unnecessary temporary disk I/O.