> ## Documentation Index
> Fetch the complete documentation index at: https://docs.geko.sh/llms.txt
> Use this file to discover all available pages before exploring further.

# drop-in replacement for Whisper

> point an existing OpenAI transcription client at geko by changing baseURL. no other code changes.

Seta implements OpenAI's transcription endpoint, so migrating off `whisper-1` is a **base URL change**.

```
POST /v1/audio/transcriptions
```

## the whole migration

<CodeGroup>
  ```ts TypeScript theme={null}
  import OpenAI from "openai";
  import { createReadStream } from "node:fs";

  const openai = new OpenAI({
    apiKey: process.env.GEKO_API_KEY,                                    // sk-tokay-…
    baseURL: "https://geko--seta-serve-transcriber-api.modal.run/v1",    // ← the only change
  });

  const { text } = await openai.audio.transcriptions.create({
    file: createReadStream("call.wav"),
    model: "seta-kk-ru-v2",
  });
  ```

  ```python Python theme={null}
  from openai import OpenAI

  client = OpenAI(
      api_key=os.environ["GEKO_API_KEY"],
      base_url="https://geko--seta-serve-transcriber-api.modal.run/v1",  # ← the only change
  )

  text = client.audio.transcriptions.create(
      file=open("call.wav", "rb"),
      model="seta-kk-ru-v2",
  ).text
  ```

  ```bash curl theme={null}
  curl -X POST "https://geko--seta-serve-transcriber-api.modal.run/v1/audio/transcriptions" \
    -H "Authorization: Bearer $GEKO_API_KEY" \
    -F file=@call.wav \
    -F model=seta-kk-ru-v2
  ```
</CodeGroup>

For Kazakh that swap is worth **8.71% WER instead of 44.95%** — see [accuracy](/stt/accuracy).

## parameter compatibility

| parameter         | behaviour                                                             |
| ----------------- | --------------------------------------------------------------------- |
| `file`            | **required** — the audio                                              |
| `model`           | accepted; the endpoint serves one model, reported by `GET /v1/models` |
| `response_format` | `json` (default), `text`, `verbose_json`                              |
| `language`        | **accepted and ignored** — see below                                  |
| `prompt`          | accepted and ignored                                                  |
| `temperature`     | accepted and ignored                                                  |

### why `language` is ignored

Not laziness — it is the point. Kazakh and Russian share one output vocabulary, so the model is never told which language to expect and a speaker can switch mid-sentence. Forcing one language per utterance is exactly the Whisper limitation this model exists to remove, so honouring `language` would make the model *worse*. It stays accepted so existing clients don't break.

### `srt` and `vtt` are refused

They return **`400`** with a reason, rather than silently handing back JSON:

```json theme={null}
{ "detail": "srt/vtt need word timings, which a CTC model does not produce. Use response_format=json, text or verbose_json." }
```

A character-CTC model has no word-level timestamps to emit. `verbose_json` returns an empty `segments` array for the same reason — it is not an error, there is simply nothing truthful to put there. Word timings are on the [roadmap](/roadmap).

## response shapes

<CodeGroup>
  ```json response_format=json theme={null}
  { "text": "сәлеметсіз бе бүгін ауа райы өте жақсы" }
  ```

  ```text response_format=text theme={null}
  сәлеметсіз бе бүгін ауа райы өте жақсы
  ```

  ```json response_format=verbose_json theme={null}
  {
    "task": "transcribe",
    "language": "kk-ru",
    "duration": 2.84,
    "text": "сәлеметсіз бе бүгін ауа райы өте жақсы",
    "segments": []
  }
  ```
</CodeGroup>

<Note>
  Prefer the native [`/v1/transcribe`](/stt) endpoint for new code — same auth and price, but it also returns `audio_seconds`, `credits_charged`, and real-time factor, which the OpenAI shape has no room for.
</Note>