Build with accurate medical transcription

Create an API key in the console, export it as OMI_API_KEY, then choose an example. The endpoint is OpenAI-compatible, so existing OpenAI SDK clients only need a different base URL.

curl https://api.omi.health/v1/audio/transcriptions \
  -H "Authorization: Bearer $OMI_API_KEY" \
  -F [email protected] \
  -F model=omi-medical-1 \
  -F language=en \
  -F response_format=verbose_json

python -m pip install openai

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.omi.health/v1",
    api_key=os.environ["OMI_API_KEY"],
)

with open("consultation.wav", "rb") as audio:
    result = client.audio.transcriptions.create(
        model="omi-medical-1",
        file=audio,
        response_format="verbose_json",
        extra_body={"language": "en"},
    )

print(result.text)

npm install openai

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

const client = new OpenAI({
  baseURL: "https://api.omi.health/v1",
  apiKey: process.env.OMI_API_KEY,
});

const result = await client.audio.transcriptions.create({
  model: "omi-medical-1",
  file: fs.createReadStream("consultation.wav"),
  response_format: "verbose_json",
  language: "en",
});

console.log(result.text);

Use verbose_json when you need timestamps, language metadata, model information, or a request ID for support. Use text for plain text and json for the smallest JSON response.

{
  "text": "The patient takes metformin 500 mg twice daily.",
  "language": "en",
  "duration": 4.82,
  "model": "omi-medical-1",
  "request_id": "0b0e…",
  "segments": [
    {
      "id": "seg_0001",
      "speaker": "A",
      "text": "The patient takes metformin 500 mg twice daily.",
      "start": 0.22,
      "end": 4.60,
      "confidence": null
    }
  ],
  "metadata": {
    "api_version": "2026-07-01",
    "runtime": "omi-runtime",
    "processing_ms": {"asr": 640, "diarization": 0}
  }
}

• English is the benchmarked, medically evaluated language.
• Spanish, Portuguese, French, German, Dutch, Arabic, and Hindi are available for testing and evaluation, and are not medically evaluated.
• language=auto
• detects each utterance independently across the eight supported families.
• Encounter vocabulary accepts up to 1,000 terms; focused 5–30 term lists are recommended and use the safest direct mode.
• Speaker labels and word timestamps decorate sealed transcript text; the decoration lane cannot rewrite it.
• Patterns can recover expected dosage and identifier formatting in invited preview requests without acting as a source of transcript truth.