Public transcription benchmarks tell you how a model performs on someone else's audio. Your product transcribes your customers' audio: specific accents, product names, and recording conditions. The smallest honest test is to run one of your own files through the endpoint and score it against a transcript you already trust.
One call, two ways.
cURL:
curl https://webwright.ai/v1/audio/transcriptions \
-H "Authorization: Bearer www_v1_YOUR_API_KEY" \
-F "file=@sample.mp3" \
-F "model=whisper-large-v3-turbo" \
-F "response_format=json"Python with the OpenAI SDK (change the base URL and the key, nothing else):
from openai import OpenAI
client = OpenAI(
api_key="www_v1_YOUR_API_KEY",
base_url="https://webwright.ai/v1",
)
with open("sample.mp3", "rb") as audio:
result = client.audio.transcriptions.create(
model="whisper-large-v3-turbo",
file=audio,
response_format="json",
)
print(result.text)Pick the format that fits the app.
json and text are simplest. srt and vtt give you timed captions. verbose_json carries more structure when you need it.
Confirm the format you plan to ship in the same call, so a format
surprise does not surface in production. For the full comparison of
the five formats, see which transcription response format your pipeline should use.
Score it against your own reference.
Keep one file with a transcript you have checked by hand. Compare the two with a simple word-level check: split both on whitespace, then count substitutions, insertions, and deletions against the reference and divide by the reference word count. This is a word-error rate on your audio, not a leaderboard number, and it is the number your team can act on. Run three files from your real workload (one easy, one noisy, one long) before you run a batch.
Then decide.
If the accuracy clears your bar on your own audio, run a larger sample and measure latency and throughput the same way. If you want help turning the result into a migration plan, book a consult.
The WebWright Team