Ormia API

Send audio, get clean, punctuated text back. The API is OpenAI-compatible: tools and SDKs that already call OpenAI’s transcription endpoint switch by changing the base URL and key.

Quickstart

Transcribe a file with cleanup (fillers removed, “no wait” corrections applied, lists formatted):

curl https://ormia.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $ORMIA_API_KEY" \
  -F file=@meeting-note.m4a \
  -F style=formal \
  -F dictionary="Ormia, Qwen3"

from openai import OpenAI

client = OpenAI(base_url="https://ormia.ai/v1", api_key="orm_...")
with open("meeting-note.m4a", "rb") as f:
    result = client.audio.transcriptions.create(model="ormia-1", file=f)
print(result.text)

import fs from "node:fs";
import OpenAI from "openai";

const client = new OpenAI({ baseURL: "https://ormia.ai/v1", apiKey: process.env.ORMIA_API_KEY });
const result = await client.audio.transcriptions.create({
  model: "ormia-1",
  file: fs.createReadStream("meeting-note.m4a"),
});
console.log(result.text);

Response:

{ "text": "The launch is Friday." }

Authentication

Every request needs an API key in the Authorization header:

Authorization: Bearer orm_...

Keys are shown once when created. Ormia stores only a hash of each key. Keep keys on your server; never put them in a web page or a mobile app.

POST/v1/audio/transcriptions

Audio in, clean text out. The request is multipart/form-data.

4.4 MB fits about 4.5 minutes of 16 kHz WAV, or 15 minutes or more of compressed audio. For longer recordings, send compressed audio or split it.

verbose_json adds raw_text (before cleanup), language, duration (billed seconds), cleanup and style. If cleanup fails or takes longer than 4 seconds, you get the raw transcript and cleanup is false.

Every successful response has a Server-Timing header with the milliseconds spent on each step (read, reserve, asr, cleanup, settle, total).

POST/v1/edit

Rewrite text with a spoken or typed instruction. Text up to 20,000 characters, instruction up to 2,000.

curl https://ormia.ai/v1/edit \
  -H "Authorization: Bearer $ORMIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "We should probably meet sometime next week to discuss the roadmap.", "instruction": "make this more direct"}'

{ "text": "Let's meet next week to discuss the roadmap." }

GET/v1/usage

This month’s usage and what is left of your allowance.

curl https://ormia.ai/v1/usage -H "Authorization: Bearer $ORMIA_API_KEY"

{
  "plan": "free",
  "month_audio_seconds": 600, "allowance_audio_seconds": 1800, "remaining_audio_seconds": 1200,
  "month_edits": 3, "allowance_edits": 50, "remaining_edits": 47
}

Languages

Qwen3-ASR recognises 30 languages. Pass one of these codes as language, or leave it out to auto-detect:

zh en yue ar de fr es pt id it ko ru th vi ja tr hi ms nl sv da fi pl cs fil fa el hu mk ro

Allowances and limits

Audio is counted by the duration the speech model reports, rounded up to the second. Ormia reserves an estimate before transcribing and settles to the real duration afterwards, so parallel requests can’t run past an allowance. When an allowance is used up, requests return 429 with code insufficient_quota until the next calendar month (UTC). Each network is also limited to 300 requests a minute.

Errors

Errors use OpenAI’s shape, so SDKs show the message:

{ "error": { "message": "Invalid or revoked API key.", "type": "invalid_request_error", "code": "invalid_api_key" } }

Privacy

Audio and text pass through OpenRouter to the provider that runs the open-weight Qwen models (Qwen3-ASR-0.6B for speech, with Qwen3-ASR-1.7B as a fallback; Qwen3-30B-A3B-Instruct for cleanup), and are not stored by Ormia. Ormia records only the seconds of audio and the number of requests per day, for billing. Read the full privacy policy.