Speech-to-text

POST https://gateway.llm-stats.com/v1/stt/transcribe

Upload an audio file as multipart form data, get a JSON transcript back.

Quickstart

import requests

with open("audio.mp3", "rb") as f:
    response = requests.post(
        "https://gateway.llm-stats.com/v1/stt/transcribe",
        headers={"Authorization": "Bearer YOUR_API_KEY"},
        files={"audio": f},
        data={"model_id": "whisper-1"},
    )

print(response.json()["text"])

const formData = new FormData();
formData.append("audio", audioFile);
formData.append("model_id", "whisper-1");

const response = await fetch(
  "https://gateway.llm-stats.com/v1/stt/transcribe",
  {
    method: "POST",
    headers: { Authorization: "Bearer YOUR_API_KEY" },
    body: formData,
  },
);

const result = await response.json();
console.log(result.text);

curl https://gateway.llm-stats.com/v1/stt/transcribe \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "[email protected]" \
  -F "model_id=whisper-1"

Request

The body is multipart/form-data with the following fields:

file

required

Audio file. Up to 25 MB. Supported formats: wav, mp3, m4a, mp4, webm, ogg, opus, flac.

string

required

STT model ID (e.g. whisper-1, deepgram-nova-3).

string

ISO-639 language hint (e.g. "en", "es"). Skips auto-detection where supported.

string

Force a specific provider for this request. Bypasses the router — use sparingly, only when you need parity with a baseline.

Response

{
  "text": "Hello, this is a test transcription.",
  "duration": 2.45,
  "language": "en",
  "confidence": 0.97,
  "words": [
    { "word": "Hello", "start": 0.00, "end": 0.42, "confidence": 0.98 },
    { "word": "this",  "start": 0.55, "end": 0.74, "confidence": 0.97 }
  ],
  "model": "whisper-1"
}

string

Full transcript.

number

Audio duration in seconds.

string | null

Detected (or supplied) language code.

number | null

Overall confidence between 0 and 1, when the provider exposes it.

array | null

Per-word timestamps and confidences, when the provider supports them.

Streaming

For real-time transcription, open a WebSocket to wss://gateway.llm-stats.com/v1/stt/stream and stream PCM audio frames. The batch HTTP endpoint above is the right choice for files you already have on disk.

Errors

Failures use the shared error envelope. Common ones:

Status	`error.code`	When
`400`	`invalid_input`	Missing fields, unsupported format.
`401`	`unauthenticated`	Missing or invalid API key.
`402`	`insufficient_quota`	Out of credit.
`413`	`invalid_input`	File larger than 25 MB.
`429`	`rate_limited`	Quota exceeded — back off using `Retry-After`.
`502`	`provider_unavailable`	Every STT provider for this model errored.

Overview

Endpoints

Quickstart

Request

Response

Streaming

Errors

​Quickstart

​Request

​Response

​Streaming

​Errors

Quickstart

Request

Response

Streaming

Errors