If your code already talks to the OpenAI API to transcribe or to generate speech, you don't have to rewrite it. We speak the same dialect: same routes, same parameter names, same voice names.
Utterafrom openai import OpenAI
client = OpenAI(
base_url="https://api.uttera.ai/v1", # ← 1
api_key="sk-echo-...", # ← 2
)
That's it. The rest of your code stays untouched.
This isn't a brochure claim: the published examples are run against the live API before
they're published, with the official openai package and no patching. They're in
uttera-examples/openai-sdk,
in Python, Node and curl.
| Works the same | Ours only |
|---|---|
POST /v1/audio/transcriptions |
POST /v1/summarize |
POST /v1/audio/speech |
POST /v1/translate |
GET /v1/models |
Voice analysis with ?extras= |
Voices alloy echo fable nova onyx shimmer |
GET /v1/usage/last |
The extra endpoints don't get in the way: a client that ignores them behaves exactly as it would against OpenAI. You can migrate today and discover them next month.
The timeout. And it is by far the most common mistake when integrating.
We hold the connection open for up to 7200 seconds so a long recording can finish. But the default in many libraries is 30 seconds or a minute, and with that you cut off jobs that were doing fine: the server keeps processing, you've already got an exception, and the wrong conclusion is "this doesn't work".
client = OpenAI(base_url="https://api.uttera.ai/v1",
api_key="sk-echo-...",
timeout=7200)
For reference: a 100-minute recording is transcribed in 10 to 35 seconds. The long wait isn't for the normal case — it's so the rare case isn't lost.
Billing is per second of audio, not per request. Ten six-second requests cost the same as one sixty-second request. That changes how it pays to split the work: there's no penalty for making many small calls.
Every response tells you what it cost. The X-Audio-Duration header carries the
billed seconds, and the credit headers carry what you have left. You don't have to wait
for the end of the month or check somewhere else.
Errors arrive in two shapes. The ones the edge generates carry error and message.
The ones the engine generates — a file that can't be decoded, one over the size limit —
carry detail. When reading an error, look at error and fall back to detail if it
isn't there. It's one line of code and it saves you a useless log the day something
fails.
Your audio is processed in Spain and isn't stored. That's not an API difference — your code won't notice — but it probably is a difference for whoever has to sign the contract. It's explained, article by article, in the privacy documentation.
Anything to add or correct? Write to support@uttera.ai. If you correct us, we edit the post and credit you.