Hosted Whisper API for Australia

The three options for Whisper in Australia — and the problem with two of them

When you need speech-to-text in an Australian product, Whisper is usually the starting point. The model is open-source, the accuracy is good, and it handles Australian accents reasonably well. But there are three ways to use it, and two of them have serious problems for Australian teams.

Option 1: Run Whisper yourself

Self-hosting Whisper keeps data in Australia and gives you full control. It's also a significant operational burden. You need GPU instances (Whisper is slow on CPU for anything but the smallest model), infrastructure to queue and process jobs, monitoring, scaling, model updates, and someone to be on-call when it breaks. For teams whose core product is not transcription infrastructure, this is usually not the right trade.

Option 2: OpenAI's Whisper API

OpenAI offers a hosted Whisper endpoint at api.openai.com/v1/audio/transcriptions. It's convenient, fast, and cheap. It's also hosted in the United States.

Every audio file you send to OpenAI's API crosses the Australian border. If that audio contains personal information about Australian individuals — which most business recordings do — you have just triggered APP 8 of the Privacy Act 1988 (Cth). APP 8 requires you to take reasonable steps to ensure the overseas recipient won't breach Australian Privacy Principles. OpenAI's terms and data processing agreements are designed for a US-centric market. Most Australian businesses haven't done the due diligence APP 8 requires, which means every API call is a latent privacy breach.

Option 3: A managed Whisper API in Australia

Australian Transcription is built on the same Whisper model family, hosted entirely on AWS ap-southeast-2 (Sydney). Your audio is processed and stored in Australia. It never leaves the country, so APP 8 is never triggered. You get a simple REST API without the operational overhead of running the infrastructure yourself.

Quickstart: transcribe audio in Python

The API follows a submit-then-poll pattern. You post a file, get back a job ID, and poll until the transcription is complete. Here's the minimal working example:

import requests
import time

API_KEY = "your-api-key"
BASE_URL = "https://api.icana.ai/api/v1"
HEADERS = {"X-API-Key": API_KEY}

# Step 1: Submit audio file
with open("meeting.mp3", "rb") as f:
    response = requests.post(
        f"{BASE_URL}/transcribe",
        headers=HEADERS,
        files={"file": f},
        data={
            "num_speakers": 2,       # optional: enables speaker diarization
            "prompt": "APRA, CPS 234, fintech",  # optional: vocabulary hints
        }
    )
response.raise_for_status()
job_id = response.json()["job_id"]
print(f"Job submitted: {job_id}")

# Step 2: Poll for completion (typically 20-60% of audio duration)
for attempt in range(60):
    result = requests.get(
        f"{BASE_URL}/jobs/{job_id}",
        headers=HEADERS
    ).json()

    status = result["status"]
    if status == "complete":
        print("\nTranscription:")
        print(result["transcription"])

        if result.get("diarization"):
            print("\nSpeaker breakdown:")
            for segment in result["diarization"]:
                start = segment["start"]
                speaker = segment["speaker"]
                text = segment["text"]
                print(f"  [{start:.1f}s] {speaker}: {text}")
        break
    elif status == "failed":
        raise RuntimeError(f"Job failed: {result.get('error')}")
    else:
        print(f"  [{attempt+1}/60] Status: {status} — waiting...")
        time.sleep(5)
else:
    raise TimeoutError("Job did not complete within 5 minutes")

from openai import OpenAI

client = OpenAI(api_key="sk-...")

with open("audio.mp3", "rb") as f:
    transcript = client.audio.transcriptions.create(
        model="whisper-1",
        file=f,
        response_format="text"
    )

print(transcript)

# Note: audio crosses the Australian border.
# APP 8 obligations are triggered for
# recordings containing personal information.

import requests, time

HEADERS = {"X-API-Key": "your-key"}
BASE = "https://api.icana.ai/api/v1"

with open("audio.mp3", "rb") as f:
    r = requests.post(
        f"{BASE}/transcribe",
        headers=HEADERS,
        files={"file": f}
    )
job_id = r.json()["job_id"]

for _ in range(60):
    d = requests.get(
        f"{BASE}/jobs/{job_id}",
        headers=HEADERS
    ).json()
    if d["status"] == "complete":
        print(d["transcription"])
        break
    time.sleep(5)

# Audio stays in AWS Sydney.
# APP 8 cross-border obligations are
# never triggered.